Construction of intelligent visual design system based on convolutional neural network under the background of big data

Construction of intelligent visual design system based on convolutional neural network under the background of big data | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Construction of intelligent visual design system based on convolutional neural network under the background of big data Chunxia Geng This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-2997795/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract In recent years, artificial intelligence has been initially applied to the visual design process, which has had a certain impact on the traditional design methods and working modes. Through the analysis of the working principle of the key technology of artificial intelligence design, the generation antagonism network (GAN), this paper summarizes the advantages and limitations of the existing automatic design methods, and puts forward the transformation path and specific methods of visual design working mode oriented to automation design efficiency optimization and new technology application form innovation under the background of new technology. With the rapid development of cloud computing and artificial intelligence, there are more and more applications in the cloud based on artificial intelligence as a service. Among them, artificial intelligence based on convolutional neural network is widely used. Convolution neural network endogenous control is the first fusion of endogenous security ideas in deep neural network model security control. Based on the core operation structure of convolutional neural network, the key control points of the endogenous control model are proposed; By tightly coupling the endogenous control unit at the control point, a convolutional neural network endogenous control model is designed and proposed. Artificial intelligence Convolution neural network Intelligent visual design big data Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Introduction In recent years, the development and application of artificial intelligence technology in the field of visual design have shown a development trend that has attracted much attention [ 1 ] . It has not only made some breakthroughs and accumulated in the basic theory and technology field, but also appeared relevant industrial application examples in specific application fields [ 2 ] . Although there are still many debates on the relationship between AI and the essence of art, the involvement of AI in the field of design is easier to understand [ 3 ] . This is mainly determined by the production attribute of design and the tool attribute of artificial intelligence technology. Therefore, the in-depth application and development of AI in the design field has become the focus of research in the past 10 years [ 4 ] . A typical representative application example is the Luban intelligent design platform launched by Alibaba Intelligent Design Laboratory, which is the first automatic design platform based on artificial intelligence technology put into practical production [ 5 ] . In April 2019, Slidebean, an AI startup in the United States, also proposed to use the AI technology automatic presentation design engine to replace simple manual design work [ 6 ] . From the perspective of its huge efficiency advantages in the field of actual business design, the emergence of automated design modules based on artificial intelligence will pose a great challenge to traditional art and design methods and even to the industry itself, which has triggered intense discussions in the field of theoretical research [ 7 ] . However, from the perspective of technology itself, there is still a long way to go for the basic technical route and specific implementation methods of current AI to realize the "intelligent ability" stage proposed in the Dartmouth Conference [ 8 ] . First of all, from the current stage of AI technology research, the various functional levels achieved at present are still at the level of "weak AI" [ 9 ] . In terms of application scenarios, it is an automated method with more complex criteria and higher degree of automation than traditional informatics methods [ 10 ] . Secondly, from the perspective of task means, the operation of the existing AI aided design system is based on the four basic modules of framework, element, actor and evaluation network to realize the simulation of general artificial design behavior patterns [ 11 ] . The specific mechanism is to take prefabricated material elements as the center, and make preliminary combination according to different page templates (including fonts, text, material drawings and other elements) and prefabricated frames (such as templates or rules such as layout), And evaluate the effect of the combination results based on big data background by generating a confrontation network (GAN) (Fig. 1 ) (Mainly, according to the big data statistics, according to the performance evaluation in specific application scenarios, the works are graded and calibrated according to their fitness from best to worst, and used as a training set for in-depth learning network training, so as to achieve clustering of design products according to a certain quality evaluation standard, and finally select and retain the best scheme among the results generated under limited constraints) [ 12 ] . The current artificial intelligence design method first constructs a large number of candidate elements ⑤, and conditionally combines the environment as a hidden random variable under certain conditions; In the specific implementation stage, the combination of elements is intervened through the classification of basic elements (such as words, images, graphics, colors, etc.) and the general visual paradigm (such as the layout segmentation framework, etc.), thus forming a limited enumeration of alternatives; Then, based on big data such as user experience evaluation, the adaptive evaluation of the scheme is carried out by means of artificial intelligence - specifically using methods such as confrontation game, that is, the process of identifying potential loss function in the network is defined to achieve the approximation of the optimal scheme [ 13 ] . It can be seen that the basic form of AI design in the contemporary technological environment can be summarized as follows: material selection and evaluation following a specific automated workflow [ 14 ] . Only the selection and pretreatment of original materials, the selection of design patterns and specific materials, and the evaluation of design alternatives, which originally required a lot of experienced manual processing, were handed over to the AI system for combination, evaluation, and screening [ 15 ] . The realization of AI design can not completely replace the work of designers, but to a certain extent, it replaces the material selection and scheme evaluation behavior in the design process [ 16 ] . The essence of this behavior is based on the imitation of design behavior and design strategy [ 17 ] . However, the strategies used in the specific design process are also very different [ 18 ] . For example, in terms of color matching design, layout design and other aspects, the selection strategy, evaluation criteria and implementation methods they follow depend on their own independent systems - chromatography, layout grid, etc [ 19 ] . For the problem of how to effectively implement the design strategy in the automatic design process, researchers have proposed different solutions [ 20 ] . Taking pattern color matching as an example, clustering algorithms such as K-means can be used to create matching rules for existing color systems according to different styles, and with the help of simple secondary development links such as scripts or macro programs, effective color matching behavior simulation can be achieved without using neural networks and other methods that require a large number of training sets [ 21 ] . Materials and Methods 2.1 Overview of visual communication design Visual communication design is an active behavior aimed at spreading specific things through visualization [ 22 ] . Visual presentation is an indispensable part of it. The influence of secondary space such as logo, typesetting, painting, graphic design, illustration, color and electronic equipment is the embodiment of visual presentation [ 23 ] . In short, visual communication design is the process of using media to transmit information to the audience, so that people can receive information through vision. In this process, designers and communication objects run through the whole process [ 24 ] . The former is the sender of information, and the latter is the receiver of information [ 25 ] . First of all, visual communication design is a design that transmits relevant information to the audience through visual media. Visual communication integrates complex and difficult information, and packages it into a beautiful image to visualize the world, so that people can have an intuitive understanding. This reflects the era characteristics and rich connotation of design. Under the background of the rapid development of science and technology, with the emergence of new energy and the development and application of product materials, this field gradually expands its scope, intersects with other fields, and gradually combines with other visual media to form a new design field. Secondly, visual communication design is also called decoration design because most of it uses printed materials as the main medium. From the perspective of development, visual communication design is scientific and rigorous, and it also contains the general trend of future design. As far as the current development of design is concerned, its main content is still "graphic design". Finally, visual communication design came into being in the era of rapid commercial development, and ultimately also serves modern commerce. Visual communication design has a wide range of fields, mainly including signs, advertising, packaging and corporate image, mainly serving for brands. These designs are transmitted to consumers through visual images, so they are called "visual communication design", which is a connecting tool between enterprises, goods and consumers. In the era of new media, designers need to analyze and summarize information, understand the design goals and create. Through emerging technologies in the mobile era, design elements such as text, graphics and color will be formed into visual products and presented in new media. Visual communication design has become a necessary link in the operation of new media platforms. Any design thinking that neglects commercial value will undoubtedly bring huge losses to the design. McKinsey, a well-known global management consulting company, has carried out more than five years of follow-up survey on 300 companies. The measurement of design behavior and the collection of a large number of financial data show that the 12 design behaviors are most relevant to financial returns. In other words, it actually shows the commercial value of design. In fact, a very important feature hidden in today's design behavior - design must have economic value. As the main branch of design, visual communication design plays an important role in promoting the development of the design industry and shaping the core competitiveness of the brand. 2.2 Design and implementation of visual design rules According to the advantages of artificial intelligence and the inefficient problem links caused by the characteristics of low intelligence demand, repetitive operation and time accumulation in the visual design and creation process of designers, we can explore the initial combination points of some human-computer cooperation modes. This usually only solves the basic layout problem between elements, but does not have a more in-depth discussion and research on the layout of elements in the visual aspect. Therefore, the layout rules of templates can be further designed and standardized by using the principle of spatial composition relationship. The core of its work objective is to introduce the visual principles, working methods and general principles of graphic design into the combination process of initial schemes in the form of constraints, so as to effectively reduce the generation of low evaluation results. AI simulates the combination of artificial processes through the graphic and image information of elements, which is determined by its adaptive characteristics facing the complex semantic environment. Compared with the fixed description of target features by traditional information classification methods oriented to general application scenarios, such as license plate recognition, character recognition, voice recognition, etc., and the technical limitations of only being able to single classify objects such as text and voice, the deep learning network can, as required, implement different definitions of the same object in a complex semantic environment by carrying out classification weight statistics on multi category labels associated with the target. It can not only recognize its general concept definition, but also classify its implicit or extended definition according to needs, so as to realize clustering based on complex criteria, which is also the technical basis for current AI methods to effectively realize design behavior. However, the current artificial intelligence design system only relies on neural network for in-depth learning, which is difficult to extract the information features with weak correlation, such as the symbolic meaning of graphic and image materials and the meaning in specific cultural context, and cannot achieve the "creative" function required by functional designers and users from a functional perspective. Although there are corresponding classification methods for the weak annotation environment, due to the great uncertainty of image graphic meaning in the cultural context and other diverse environments, the definition of image meaning dominated by designers in a certain period is of great significance for both traditional design methods and automated artificial intelligence methods. Therefore, the combination and association of implicit semantics and explicit semantics in the deep learning network can be used as the supplement and replacement of basic information labels (explicit semantics). In the context of cross modal information association, the method of constructing implicit semantic association between elements and abstract polygon symbols can be used as a means of realizing the association of design elements such as images in a diversified context. Therefore, from the perspective of design, we can use design theories and methods such as semiotics and morphological principles to mine the semiotic semantics of graphic elements and design related symbol labels while designing specific elements. As the basis for constructing the attribute association network of elements, we can effectively improve the execution efficiency of the generation network. 2.3 Operation characteristics of convolutional neural network in visual design In the first convolution layer of convolutional neural network, multiple convolutions are calculated in parallel to generate a set of linear activation responses. Convolution is a special linear operation, formally a mathematical operation on two real variable functions. Convolution has three characteristics: sparse connectivity, parameter sharing and equivariant representations. Sparse connections are also called sparse interactions or sparse weights. Traditional neural networks use matrix multiplication to establish the connection between input and output, and each output unit interacts with each input unit. Convolutional neural network detects small and meaningful features, such as edge features, by using only a part of the pixel kernel. At this time, the size of the convolution kernel is much smaller than the size of the input, so the convolution neural network needs to store fewer parameters, which improves the statistical efficiency and reduces the computational load of the neural network. Parameter sharing refers to using the same parameters in multiple functions of a convolutional neural network model. In traditional neural networks, each element of the weight matrix is used only once when calculating the output of a layer, that is, the element is no longer used after multiplying the input. This weight can be called tied weights, because the weight used for input will also be bound to other weights. In convolutional neural networks, every element of the convolution kernel acts on every position of the input. The parameter sharing of convolutional neural network ensures that only one parameter set needs to be learned, rather than a separate parameter set for each location, which can significantly reduce the storage requirements of the model. In the full connection model of traditional neural networks, a single black arrow indicates that the parameters are used only once, and parameter sharing is not used; In a convolutional neural network with a convolution kernel width of 3, the black arrow indicates the use of the middle element of the 3-element convolution kernel. This parameter is used at all input locations and parameter sharing is used. Results and Discussion 3.1 Detection layer activation function form of convolutional neural network In the M-P neuron model that has been used so far, neurons receive multiple input signals from other neurons, and the signals propagate through connections with weights. The neurons will compare the total input received with the neuron threshold. The linear activation response of the first layer output will generate new output through a nonlinear activation function, which is sometimes called the detection layer. The activation function of the neural network must be nonlinear. If linear functions are used, it will be meaningless to deepen the number of layers of neural networks, which will not give full play to the advantages of multi-layer neural networks, and even more difficult to fit all possible outputs. The ideal activation function is a step function: $$\operatorname{sgn} (x)=\left\{ {\begin{array}{*{20}{l}} {0,x<0} \\ {1,x \geqslant 0} \end{array}} \right.$$ 1 The step function maps the input to 0 or 1, where 0 corresponds to neuronal inhibition and 1 corresponds to neuronal excitation. However, the step function is discontinuous and unsmooth, so the sigmoid function is often used as the activation function, as shown in Formula (2): $$sigmoid(x)=\frac{1}{{1+{e^{ - x}}}}$$ 2 In the history of neural network development, Sigmaid function has been used for a long time. Sigmaid function or Tanh function of the same period are smooth, derivable in the whole process, and the output is bounded, which is more advantageous in the expression of the output layer. Currently, ReLU (Rectified Linear Unit) is mainly used, as shown in Formula (3): $$ReLU(x)=\left\{ {\begin{array}{*{20}{l}} {0,x \leqslant 0} \\ {x,x>0} \end{array}} \right.$$ 3 The computing power of GPU has been continuously enhanced. Now, the deep neural network in 2012 usually uses a single GPU. The AlexNet deep neural network structure of a single GPU is shown in Fig. 2 . 3.2 Implementation principle of single point endogenous control algorithm In the convolutional neural network endogenous control model, as long as there are control points in the convolutional neural network, the endogenous control model can be used to achieve endogenous control. Then, taking the convolutional neural network with N-layer output as the object, the single point endogenous control algorithm is applied to realize the endogenous control. Arguments of the activation function: $${z_1}=w_{1}^{ - } \cdot x+{b_1}$$ 4 Due to the chain structure, the output of the first layer: $${h_1}={g_1}\left( {w_{1}^{ \top } \cdot x+{b_1}} \right)={g_1}\left( {{z_1}} \right)$$ 5 Then, the output 2h of the second layer is a function of the output 1h of the first layer. If the second layer is a pooling layer, the output of the second layer is: $${h_2}={\text{~pooling}}{{\text{~}}_2}\left( {{h_1}} \right)$$ 6 The output of the third layer is a function of the second layer. If the third layer is a convolution layer, the output of the third layer is: $${h_3}={g_3}\left( {w_{3}^{ \top } \cdot {h_2}+{b_3}} \right)={g_3}\left( {{z_3}} \right)$$ 7 In formula (4), x is the user's input. If the ith hidden layer is convolution layer, ig is the activation function of the layer, iw is the weight of the hidden layer, and ib is the offset of the hidden layer. If the j hidden layer is pooling layer, 1 () j j pooling h − means pooling the output j1h − of j − 1 layer. Similarly, layer n − 1 outputs: $${h_{n - 1}}={g_{n - 1}}\left( {w_{{n - 1}}^{ - } \cdot {h_{n - 2}}+{b_{n - 1}}} \right)={g_{n - 1}}\left( {{z_{n - 1}}} \right)$$ 8 Aiming at the security problems caused by the out of control of the convolutional neural network model, this paper designs a single point endogenous control algorithm of the convolutional neural network. The single point endogenous control algorithm is designed according to the operation characteristics of convolutional neural network, which can be applied to neural networks with convolutional layers and has strong generalization characteristics. The convolutional neural network single point endogenous control algorithm is an endogenous control algorithm tightly coupled in the single operation layer control point of the convolutional neural network. It consists of a single point endogenous control function, a single point endogenous control factor and a single point endogenous authorization operator. After the convolutional neural network is trained, a single point endogenous control unit is coupled to any control point, and the N-layer convolutional neural network structure of the single point endogenous control algorithm is applied. At this time, the control point is the active function control point of the first detection layer. The output 1h of the first layer of the convolutional neural network controlled by the single point endogenous control algorithm can be expressed as: $$h_{1}^{\prime }={g_1}\left[ {\xi \left( {{z_1},k} \right)} \right]$$ 9 Different convolutional neural networks have different structures, so there may not be control points in each layer that can fuse the single point endogenous control algorithm, which leads to different controlled expressions in different operation layers. If the second layer of the convolutional neural network is the pooling layer, and the pooling layer does not contain control points, the output 2h of this layer can be expressed as: $$h_{2}^{\prime }={\text{~pooling}}{{\text{~}}_2}\left( {h_{1}^{\prime }} \right)$$ 10 Then, the output − 1nh of the n − 1 roll can be expressed as: $${h_{n - 1}}={g_{n - 1}}\left( {{z_{n - 1}}} \right)$$ 11 3.3 Implementation principle of mean collapse optimization After the mean collapse optimization scheme is adopted, the weight control factor of the endogenous control algorithm is no longer randomly selected. Let the weight control factor take the expectation of the weight of the layer, then 0 Ω→ ik. Taking N-layer convolutional neural network as the object, the process of applying the multi-point endogenous control algorithm based on mean collapse optimization to realize the endogenous control in each weight layer is analyzed. After mean collapse optimization, the output 1h of the first layer of the convolutional neural network can be expressed as: $$h_{1}^{\prime }={g_1}\left\{ {{{\left[ {\zeta \left( {{w_1},E{{\bar {\omega }}_1}} \right)} \right]}^ - } \cdot x+{b_1}} \right\}$$ 12 Then, the output − 1nh of the n − 1 roll can be expressed as: $$h_{{n - 1}}^{\prime }={g_{n - 1}}\left\{ {{{\left[ {\zeta \left( {{w_{n - 1}},E{{\bar {\omega }}_{n - 1}}} \right)} \right]}^ - } \cdot h_{{n - 2}}^{\prime }+{b_{n - 1}}} \right\}$$ 13 When the endogenous control function is a linear operation 61562ζ (x, k) = x ζ k, the mean collapse optimization scheme first performs a collapse operation with a collapse ratio of ཛྷ ik for the selected convolutional neural network weight control point. At this time, the controlled layer weight \(\omega _{i}^{\prime }={\omega _i} \cdot {k_{{\omega _i}}}={\omega _i} \cdot E{\bar {\omega }_i} \to 0\) of the controlled convolution neural network will output: $$h_{{n - 1}}^{\prime } \approx {g_{n - 1}}\left\{ {0 \cdot h_{{n - 2}}^{\prime }+{b_{n - 1}}} \right\}={g_{n - 1}}\left( {{b_{n - 1}}} \right)$$ 14 The output of the n − 1 layer of the original convolutional neural network is: $${h_{n - 1}}={g_{n - 1}}\left( {w_{{n - 1}}^{ \top } \cdot {h_{n - 2}}+{b_{n - 1}}} \right)$$ 15 By comparing equations ( 13 ) and ( 14 ), it can be found that the output 1n-h and n-1h differ greatly after mean collapse optimization at the weight control point, and the theoretical control effect is good. When using the mean collapse optimization scheme to optimize the endogenous control algorithm, it is necessary to propose a performance index to measure the quality of the mean collapse optimization scheme. In addition to intuitively analyzing the controlled precision of convolutional neural network, the performance indicators also include: algorithm control rate, time cost ratio and algorithm utility ratio. Experimental results and analysis 4.1 Optimization experiment analysis Le Net-5 is selected as the experimental network to study the control effect of the optimization scheme of the endogenous control algorithm based on mean collapse at the weight control points. The weight control points of each convolution layer of LeNet-5 are selected for comparative experiments. One group applies a single point endogenous control algorithm to the single weight control points of C1, C3, and C5 convolutions, and applies a multiple point endogenous control algorithm to multiple weight control points; In the other group, the types of endogenous control algorithms for each weight control point remain unchanged, but they are all optimized using an optimization scheme based on mean collapse, and the experimental results are recorded, as shown in Fig. 3 . For LeNet-5 convolutional neural network, the optimization scheme based on mean value collapse can effectively improve the control accuracy of the endogenous control algorithm and realize the optimization of the endogenous control algorithm. The deeper VGG-16 is selected as the experimental network to study the performance of the mean collapse optimization scheme. For VGG-16, apply single point endogenous control algorithm, multi-point endogenous control algorithm, single point endogenous control algorithm after mean collapse optimization and multi-point endogenous control algorithm after mean collapse optimization respectively, and record the experimental results, as shown in Fig. 4 . For VGG-16 convolutional neural network, the optimization scheme based on mean collapse can optimize the endogenous control algorithm. 4.2 Optimization of endogenous control algorithm based on giant parameters AI simulates the combination of artificial processes through the graphic and image information of elements, which is determined by its adaptive characteristics facing the complex semantic environment. Compared with the fixed description of target features by traditional information classification methods oriented to general application scenarios, such as license plate recognition, character recognition, voice recognition, etc., and the technical limitations of only being able to single classify objects such as text and voice, the deep learning network can, as required, implement different definitions of the same object in a complex semantic environment by carrying out classification weight statistics on multi category labels associated with the target. It can not only recognize its general concept definition, but also classify its implicit or extended definition as needed, so that it can achieve clustering based on complex criteria, which is also the technical basis for current AI methods to effectively implement design behavior. The experimental results are shown in Fig. 5 . According to the analysis of the above two experimental results, for the controlled LeNet-5, the prediction accuracy of unauthorized users is lower when the endogenous control algorithm is applied to the deeper weight control points; For controlled VGG-16, endogenous control algorithm is applied to deeper weight control points after the first layer and C17, and the prediction accuracy of unauthorized users is lower. It can be seen that endogenous control of deeper bias control points has greater impact on the output of convolutional neural network. Compared with other control points, the time cost of applying the endogenous control algorithm to the offset control points in the convolution layer is smaller, because the parameters of the offset control points are often smaller than the weight control points. If the offset matrix of each layer of VGG-16 is different, the maximum parameter quantity of each layer of offset matrix is 4096 × 1. The parameter amount of the weight can reach tens of thousands of times of the offset parameter amount at most. Therefore, the cost of fusing offset control factors and corresponding control algorithms at offset control points is less than that of fusing weight control factors of equal scale at the same level of weight control points. 4.3 Experimental results of visual design under big data In the past, visual design was limited to paper media, and the communication was weak, so the flexibility in content layout was limited by the size of paper printing. Today, new media, as the main carrier of information communication, such as public environment display screen, smart phone, tablet computer, etc., have diversified communication methods. With the development of Internet technology, adaptive page technology is becoming more and more mature. The same content can get the best typesetting effect under different size browsers, and the flexibility is significantly improved, but at the same time, it puts forward higher requirements for the professional quality and professional level of designers. As shown in Fig. 6 . Image is the most intuitive visual trigger, which can convey a lot of information in a short time. The design based on illustration art can be applied to the publicity and promotion of many themes in new media, and has the effect of "1 + 1 ༞ 2". With the cross boundary and integration of various designs, the promotion and application of illustrations today also need to comply with the social progress and market demand. At present, the illustration is mainly used online for enterprise IP, H5, We Media (WeChat official account, Tiktok, Sina Weibo, etc.), large-scale enterprises and product theme promotion, etc. Immersion is to let people focus on the current situation created by the designer and feel happy and satisfied, while forgetting the real world situation. "Tiktok" is typical of immersive experience, as shown in Fig. 7 . With the continuous expansion of mobile Internet communication channels, "two micro and one" has become the main battlefield of media integration and development. As of June 2020, short video has surpassed instant messaging by 110 minutes per person per day. When analyzing the immersive design of "Tiktok" from the perspective of visual design, it has two obvious visual characteristics: one is the de modal design of floating layer operation. During the video viewing process, when making comments, forwarding and other interactive behaviors, the floating layer operation of interactive button floating is a semi transparent temporary floating layer, and without "masking" processing, it will not interrupt the current video playing behavior. Conclusion The application of artificial intelligence has had a profound impact on the development of today's design field. However, from the analysis of the specific implementation methods and effectiveness of the current technology itself, its essence is to replace the relatively basic element combination and example flow in the design process. Moreover, due to the limitations of technology, resources and other aspects, it cannot fundamentally replace the design position of people, so its role and impact are only concentrated on the tool level. However, its self-adaptive association of multiple information provides a good experimental environment and implementation basis for the expansion of design methods. At the same time, the task of design has changed from formal innovation to the understanding and interpretation of form-oriented definition and graphic connotation, as well as the innovation of new technology-oriented application forms. Two kinds of convolutional neural network endogenous control algorithms are designed, and their performance is tested on different convolutional neural networks to study the performance characteristics of different endogenous control algorithms. Aiming at the shortcomings of two endogenous control algorithms, three optimization schemes are proposed and the algorithm performance is studied. The experiment shows that the endogenous control algorithm of the optimization scheme can achieve complete control of the convolutional neural network model. Declarations Data availability The figures used to support the findings of this study are included in the article. Conflicts of interest The authors declare that they have no conflicts of interest. Funding statement This work was not supported by any funds. Acknowledgements The authors would like to show sincere thanks to those techniques who have contributed to this research. References Oh BD, Lee YK, Song HJ (2021) Age group classification to identify the progress of language development based on convolutional neural networks. J Intell Fuzzy Systems: Appl Eng Technol 36(9):2508–2515 Abudureheman A, Nilupaer A, He Y (2020) Performance evaluation of enterprises' innovation capacity based on fuzzy system model and convolutional neural network. J Intell fuzzy systems: Appl Eng Technol 521(7553):436–444 Revi KR, Wilscy M, Antony R (2021) Portrait photography splicing detection using ensemble of convolutional neural networks. J Intell Fuzzy Systems: Appl Eng Technol 5(4):115–133 Mu T, Lin K, Zhang H (2021) Sanitizing hidden activations for improving adversarial robustness of convolutional neural networks. J Intell Fuzzy Systems: Appl Eng Technol 40(6):1229–1251 Chen YM, Chen YJ, Tsai YK (2021) Classification of human electrocardiograms by multi-layer convolutional neural network and hyperparameter optimization. J Intell Fuzzy Systems: Appl Eng Technol 160(1):106–154 Shashank R, Prasad, Srinivasa PP (2021) Intelligent fault diagnosis ofRolling-element bearing using 1-D Deep Convolutional Neural Networks. Int J COMADEM 36(4):193–202 Prosvirin AE, Maliuk AS, Kim JM (2022) Intelligent rubbing fault identification using multivariate signals and a multivariate one-dimensional convolutional neural network. Expert Syst Application 86(11):2278–2324 Sanchez SG, Muns GR, Bocanegra C (2022) AirNN: Neural Networks with Over-the-Air Convolution via Reconfigurable Intelligent Surfaces. 20(3):97–112 Wang Y, Yan J, Jing Q (2021) A novel adversarial transfer learning in deep convolutional neural network for intelligent diagnosis of gas-insulated switchgear insulation defect. IET generation, transmission & distribution, 2(5):359–366 Yang MY, Kumaar S, Ye L (2020) Real-time Semantic Segmentation with Context Aggregation Network. 7(2):279–301 Savary E, Rojas K, Maus M (2021) A search for galaxy-scale strong gravitational lenses in the Ultraviolet Near Infrared Optical Northern Survey (UNIONS). 41(6):93–94 Shafiee MJ, Famouri M, Bathla G, TinyDefectNet (2021) Highly Compact Deep Neural Network Architecture for High-Throughput Manufacturing Visual Quality Inspection. 4(4):79–82 Altan D, Sariel S (2022) CLUE-AI: A Convolutional Three-stream Anomaly Identification Framework for Robot Manipulation. 14(1):1–17 Meletis P (2022) Towards holistic scene understanding: Semantic segmentation and beyond. 20(12):1657–1664 Krishna S, Ss S, Krishnamoorthy S Stain Normalized Breast Histopathology Image Recognition using Convolutional Neural Networks for Cancer Detection. 20229,40(6):141–146 Xompero A, Donaher S, Iashin V (2021) The CORSMAL benchmark for the prediction of the properties of containers. 25(5):33–48 Zhao X, Zhang Y (2022) An intelligent diagnosis method of rolling bearing based on multi-scale residual shrinkage convolutional neural network. Meas Sci Technol 31(10):11–20 Jiangquan ZHANG, Yi (2020) A new bearing fault diagnosis method based on modified convolutional neural networks. Chin J Aeronaut 167(2):54–62 Liu K, Zhao Z, Shi P (2022) Real-time surgical tool detection in computer-aided surgery based on enhanced feature-fusion convolutional neural network. J Comput Des Eng 43(8):173–175 Dehshibi MM, Ashtari-Majlan M, Adhane G (2022) ADVISE: ADaptive Feature Relevance and VISual Explanations for Convolutional Neural Networks. 40(1):17–26 Fang J, Zhou X, Wang P (2021) Densely connected Siamese network visual tracking. Industrial Robot: the international journal of robotics research and application 48(5):680–687 Spagnolo F, Perri S, Corsonello P (2022) Approximate Down-Sampling Strategy for Power-Constrained Intelligent Systems. IEEE Access 38(24):30–36 Pustokhina IV, Pustokhin DA, Rodrigues J (2020) Automatic Vehicle License Plate Recognition Using Optimal K-Means With Convolutional Neural Network for Intelligent Transportation Systems. IEEE Access 8(16):92907–92917 Lin Y, Zhong Q, Sun H (2020) A Pointer Type Instrument Intelligent Reading System Design Based on Convolutional Neural Networks. Front Phys 89(11):22–27 Ma Y, Wen G, Cheng S (2022) Multimodal convolutional neural network model with information fusion for intelligent fault diagnosis in rotating machinery. IOP Publishing Ltd 10(5):705–710 Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-2997795","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":277452037,"identity":"16f884da-9f89-41e4-b2ac-c8db42a9baef","order_by":0,"name":"Chunxia Geng","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAApUlEQVRIiWNgGAWjYBACAwnmBhAtx8befoBYLYxgLcZ8PGcSSNOSOE/CwYA4LebSja0bf7YdTm+TYEhg+FGxjbAWyzkH227znEnLbZNuPMDYc+Y2EQ67kdh2m6HCJrdN5kACM2MbkVpu/jCQSGeTSDAgXssNngqbBNK0gPxi2AYM5INE+iX52E1giMnLt7cffPCjgggtKOAAiepHwSgYBaNgFOACABALPxvkkhiVAAAAAElFTkSuQmCC","orcid":"","institution":"Heze University","correspondingAuthor":true,"prefix":"","firstName":"Chunxia","middleName":"","lastName":"Geng","suffix":""}],"badges":[],"createdAt":"2023-05-30 03:39:23","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-2997795/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-2997795/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":52542509,"identity":"7070b28b-733f-4c57-8b70-c13890d9fb26","added_by":"auto","created_at":"2024-03-12 17:43:16","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":178185,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eBasic Structure of Generation Countermeasures Network (GAN)\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-2997795/v1/f905a54abaaec57dc857e1a5.png"},{"id":52542507,"identity":"d19a350c-0e04-48a5-a3c1-e04d97f99480","added_by":"auto","created_at":"2024-03-12 17:43:16","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":56537,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eAlexNet Structure of Deep Convolution Neural Network for Single GPU\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-2997795/v1/1a08c95d46603c112277080b.png"},{"id":52542506,"identity":"f8437f75-424a-4e07-a621-854f65ef55fd","added_by":"auto","created_at":"2024-03-12 17:43:16","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":31053,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePerformance comparison experiment on endogenous control Le Net-5 of optimization scheme based on mean collapse\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-2997795/v1/c616c20c828220ad0ceec66a.png"},{"id":52542508,"identity":"2872fcf7-95cb-426e-bc5b-04012790b1c5","added_by":"auto","created_at":"2024-03-12 17:43:16","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":45454,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eEndogenous control effect of optimization scheme based on mean collapse in controlled VGG-16\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-2997795/v1/82265605ddd97a3dfa25e472.png"},{"id":52543313,"identity":"8b905cba-6d98-4ce6-af6d-d78f966484f3","added_by":"auto","created_at":"2024-03-12 17:51:16","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":69296,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eApplying endogenous control algorithm to different offset control points of VGG-16\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-2997795/v1/f6c0d7c58976b1ef5115f542.png"},{"id":52542512,"identity":"890d41b5-c32b-4f74-b73f-d216677f74e8","added_by":"auto","created_at":"2024-03-12 17:43:16","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":206160,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eVisual design aided positioning\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-2997795/v1/381d2788b0e068627accb58d.png"},{"id":52542510,"identity":"fd0917ee-9c5f-4223-9b3a-e5d5d2190650","added_by":"auto","created_at":"2024-03-12 17:43:16","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":77524,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eProportion of visual element types\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-2997795/v1/8c09b123659fff5ab6c625b2.png"},{"id":84924594,"identity":"f2d429cf-5bfb-4ebd-82f5-6a66b0466ae0","added_by":"auto","created_at":"2025-06-18 21:24:54","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1451923,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-2997795/v1/ca102c37-e033-45d1-a896-3bcea7927fc6.pdf"}],"financialInterests":"","formattedTitle":"Construction of intelligent visual design system based on convolutional neural network under the background of big data","fulltext":[{"header":"Introduction","content":"\u003cp\u003eIn recent years, the development and application of artificial intelligence technology in the field of visual design have shown a development trend that has attracted much attention \u003csup\u003e[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]\u003c/sup\u003e. It has not only made some breakthroughs and accumulated in the basic theory and technology field, but also appeared relevant industrial application examples in specific application fields \u003csup\u003e[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]\u003c/sup\u003e. Although there are still many debates on the relationship between AI and the essence of art, the involvement of AI in the field of design is easier to understand \u003csup\u003e[\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]\u003c/sup\u003e. This is mainly determined by the production attribute of design and the tool attribute of artificial intelligence technology. Therefore, the in-depth application and development of AI in the design field has become the focus of research in the past 10 years \u003csup\u003e[\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]\u003c/sup\u003e. A typical representative application example is the Luban intelligent design platform launched by Alibaba Intelligent Design Laboratory, which is the first automatic design platform based on artificial intelligence technology put into practical production \u003csup\u003e[\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]\u003c/sup\u003e. In April 2019, Slidebean, an AI startup in the United States, also proposed to use the AI technology automatic presentation design engine to replace simple manual design work \u003csup\u003e[\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eFrom the perspective of its huge efficiency advantages in the field of actual business design, the emergence of automated design modules based on artificial intelligence will pose a great challenge to traditional art and design methods and even to the industry itself, which has triggered intense discussions in the field of theoretical research \u003csup\u003e[\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]\u003c/sup\u003e. However, from the perspective of technology itself, there is still a long way to go for the basic technical route and specific implementation methods of current AI to realize the \"intelligent ability\" stage proposed in the Dartmouth Conference \u003csup\u003e[\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eFirst of all, from the current stage of AI technology research, the various functional levels achieved at present are still at the level of \"weak AI\"\u003csup\u003e[\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]\u003c/sup\u003e. In terms of application scenarios, it is an automated method with more complex criteria and higher degree of automation than traditional informatics methods \u003csup\u003e[\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eSecondly, from the perspective of task means, the operation of the existing AI aided design system is based on the four basic modules of framework, element, actor and evaluation network to realize the simulation of general artificial design behavior patterns \u003csup\u003e[\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]\u003c/sup\u003e. The specific mechanism is to take prefabricated material elements as the center, and make preliminary combination according to different page templates (including fonts, text, material drawings and other elements) and prefabricated frames (such as templates or rules such as layout), And evaluate the effect of the combination results based on big data background by generating a confrontation network (GAN) (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e) (Mainly, according to the big data statistics, according to the performance evaluation in specific application scenarios, the works are graded and calibrated according to their fitness from best to worst, and used as a training set for in-depth learning network training, so as to achieve clustering of design products according to a certain quality evaluation standard, and finally select and retain the best scheme among the results generated under limited constraints)\u003csup\u003e[\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe current artificial intelligence design method first constructs a large number of candidate elements ⑤, and conditionally combines the environment as a hidden random variable under certain conditions; In the specific implementation stage, the combination of elements is intervened through the classification of basic elements (such as words, images, graphics, colors, etc.) and the general visual paradigm (such as the layout segmentation framework, etc.), thus forming a limited enumeration of alternatives; Then, based on big data such as user experience evaluation, the adaptive evaluation of the scheme is carried out by means of artificial intelligence - specifically using methods such as confrontation game, that is, the process of identifying potential loss function in the network is defined to achieve the approximation of the optimal scheme\u003csup\u003e[\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eIt can be seen that the basic form of AI design in the contemporary technological environment can be summarized as follows: material selection and evaluation following a specific automated workflow \u003csup\u003e[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]\u003c/sup\u003e. Only the selection and pretreatment of original materials, the selection of design patterns and specific materials, and the evaluation of design alternatives, which originally required a lot of experienced manual processing, were handed over to the AI system for combination, evaluation, and screening \u003csup\u003e[\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eThe realization of AI design can not completely replace the work of designers, but to a certain extent, it replaces the material selection and scheme evaluation behavior in the design process \u003csup\u003e[\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]\u003c/sup\u003e. The essence of this behavior is based on the imitation of design behavior and design strategy \u003csup\u003e[\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]\u003c/sup\u003e. However, the strategies used in the specific design process are also very different \u003csup\u003e[\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]\u003c/sup\u003e. For example, in terms of color matching design, layout design and other aspects, the selection strategy, evaluation criteria and implementation methods they follow depend on their own independent systems - chromatography, layout grid, etc \u003csup\u003e[\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]\u003c/sup\u003e. For the problem of how to effectively implement the design strategy in the automatic design process, researchers have proposed different solutions \u003csup\u003e[\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]\u003c/sup\u003e. Taking pattern color matching as an example, clustering algorithms such as K-means can be used to create matching rules for existing color systems according to different styles, and with the help of simple secondary development links such as scripts or macro programs, effective color matching behavior simulation can be achieved without using neural networks and other methods that require a large number of training sets \u003csup\u003e[\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]\u003c/sup\u003e.\u003c/p\u003e"},{"header":"Materials and Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Overview of visual communication design\u003c/h2\u003e \u003cp\u003eVisual communication design is an active behavior aimed at spreading specific things through visualization \u003csup\u003e[\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]\u003c/sup\u003e. Visual presentation is an indispensable part of it. The influence of secondary space such as logo, typesetting, painting, graphic design, illustration, color and electronic equipment is the embodiment of visual presentation \u003csup\u003e[\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]\u003c/sup\u003e. In short, visual communication design is the process of using media to transmit information to the audience, so that people can receive information through vision. In this process, designers and communication objects run through the whole process \u003csup\u003e[\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]\u003c/sup\u003e. The former is the sender of information, and the latter is the receiver of information \u003csup\u003e[\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eFirst of all, visual communication design is a design that transmits relevant information to the audience through visual media. Visual communication integrates complex and difficult information, and packages it into a beautiful image to visualize the world, so that people can have an intuitive understanding. This reflects the era characteristics and rich connotation of design. Under the background of the rapid development of science and technology, with the emergence of new energy and the development and application of product materials, this field gradually expands its scope, intersects with other fields, and gradually combines with other visual media to form a new design field.\u003c/p\u003e \u003cp\u003eSecondly, visual communication design is also called decoration design because most of it uses printed materials as the main medium. From the perspective of development, visual communication design is scientific and rigorous, and it also contains the general trend of future design. As far as the current development of design is concerned, its main content is still \"graphic design\".\u003c/p\u003e \u003cp\u003eFinally, visual communication design came into being in the era of rapid commercial development, and ultimately also serves modern commerce. Visual communication design has a wide range of fields, mainly including signs, advertising, packaging and corporate image, mainly serving for brands. These designs are transmitted to consumers through visual images, so they are called \"visual communication design\", which is a connecting tool between enterprises, goods and consumers.\u003c/p\u003e \u003cp\u003eIn the era of new media, designers need to analyze and summarize information, understand the design goals and create. Through emerging technologies in the mobile era, design elements such as text, graphics and color will be formed into visual products and presented in new media. Visual communication design has become a necessary link in the operation of new media platforms. Any design thinking that neglects commercial value will undoubtedly bring huge losses to the design. McKinsey, a well-known global management consulting company, has carried out more than five years of follow-up survey on 300 companies. The measurement of design behavior and the collection of a large number of financial data show that the 12 design behaviors are most relevant to financial returns. In other words, it actually shows the commercial value of design. In fact, a very important feature hidden in today's design behavior - design must have economic value. As the main branch of design, visual communication design plays an important role in promoting the development of the design industry and shaping the core competitiveness of the brand.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Design and implementation of visual design rules\u003c/h2\u003e \u003cp\u003eAccording to the advantages of artificial intelligence and the inefficient problem links caused by the characteristics of low intelligence demand, repetitive operation and time accumulation in the visual design and creation process of designers, we can explore the initial combination points of some human-computer cooperation modes.\u003c/p\u003e \u003cp\u003eThis usually only solves the basic layout problem between elements, but does not have a more in-depth discussion and research on the layout of elements in the visual aspect. Therefore, the layout rules of templates can be further designed and standardized by using the principle of spatial composition relationship. The core of its work objective is to introduce the visual principles, working methods and general principles of graphic design into the combination process of initial schemes in the form of constraints, so as to effectively reduce the generation of low evaluation results.\u003c/p\u003e \u003cp\u003eAI simulates the combination of artificial processes through the graphic and image information of elements, which is determined by its adaptive characteristics facing the complex semantic environment. Compared with the fixed description of target features by traditional information classification methods oriented to general application scenarios, such as license plate recognition, character recognition, voice recognition, etc., and the technical limitations of only being able to single classify objects such as text and voice, the deep learning network can, as required, implement different definitions of the same object in a complex semantic environment by carrying out classification weight statistics on multi category labels associated with the target. It can not only recognize its general concept definition, but also classify its implicit or extended definition according to needs, so as to realize clustering based on complex criteria, which is also the technical basis for current AI methods to effectively realize design behavior. However, the current artificial intelligence design system only relies on neural network for in-depth learning, which is difficult to extract the information features with weak correlation, such as the symbolic meaning of graphic and image materials and the meaning in specific cultural context, and cannot achieve the \"creative\" function required by functional designers and users from a functional perspective. Although there are corresponding classification methods for the weak annotation environment, due to the great uncertainty of image graphic meaning in the cultural context and other diverse environments, the definition of image meaning dominated by designers in a certain period is of great significance for both traditional design methods and automated artificial intelligence methods. Therefore, the combination and association of implicit semantics and explicit semantics in the deep learning network can be used as the supplement and replacement of basic information labels (explicit semantics). In the context of cross modal information association, the method of constructing implicit semantic association between elements and abstract polygon symbols can be used as a means of realizing the association of design elements such as images in a diversified context. Therefore, from the perspective of design, we can use design theories and methods such as semiotics and morphological principles to mine the semiotic semantics of graphic elements and design related symbol labels while designing specific elements. As the basis for constructing the attribute association network of elements, we can effectively improve the execution efficiency of the generation network.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Operation characteristics of convolutional neural network in visual design\u003c/h2\u003e \u003cp\u003eIn the first convolution layer of convolutional neural network, multiple convolutions are calculated in parallel to generate a set of linear activation responses. Convolution is a special linear operation, formally a mathematical operation on two real variable functions. Convolution has three characteristics: sparse connectivity, parameter sharing and equivariant representations.\u003c/p\u003e \u003cp\u003eSparse connections are also called sparse interactions or sparse weights. Traditional neural networks use matrix multiplication to establish the connection between input and output, and each output unit interacts with each input unit. Convolutional neural network detects small and meaningful features, such as edge features, by using only a part of the pixel kernel. At this time, the size of the convolution kernel is much smaller than the size of the input, so the convolution neural network needs to store fewer parameters, which improves the statistical efficiency and reduces the computational load of the neural network.\u003c/p\u003e \u003cp\u003eParameter sharing refers to using the same parameters in multiple functions of a convolutional neural network model. In traditional neural networks, each element of the weight matrix is used only once when calculating the output of a layer, that is, the element is no longer used after multiplying the input. This weight can be called tied weights, because the weight used for input will also be bound to other weights. In convolutional neural networks, every element of the convolution kernel acts on every position of the input. The parameter sharing of convolutional neural network ensures that only one parameter set needs to be learned, rather than a separate parameter set for each location, which can significantly reduce the storage requirements of the model. In the full connection model of traditional neural networks, a single black arrow indicates that the parameters are used only once, and parameter sharing is not used; In a convolutional neural network with a convolution kernel width of 3, the black arrow indicates the use of the middle element of the 3-element convolution kernel. This parameter is used at all input locations and parameter sharing is used.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results and Discussion","content":"\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Detection layer activation function form of convolutional neural network\u003c/h2\u003e \u003cp\u003eIn the M-P neuron model that has been used so far, neurons receive multiple input signals from other neurons, and the signals propagate through connections with weights. The neurons will compare the total input received with the neuron threshold. The linear activation response of the first layer output will generate new output through a nonlinear activation function, which is sometimes called the detection layer.\u003c/p\u003e \u003cp\u003eThe activation function of the neural network must be nonlinear. If linear functions are used, it will be meaningless to deepen the number of layers of neural networks, which will not give full play to the advantages of multi-layer neural networks, and even more difficult to fit all possible outputs. The ideal activation function is a step function:\u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e\n$$\\operatorname{sgn} (x)=\\left\\{ {\\begin{array}{*{20}{l}} {0,x\u0026lt;0} \\\\ {1,x \\geqslant 0} \\end{array}} \\right.$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThe step function maps the input to 0 or 1, where 0 corresponds to neuronal inhibition and 1 corresponds to neuronal excitation. However, the step function is discontinuous and unsmooth, so the sigmoid function is often used as the activation function, as shown in Formula (2):\u003cdiv id=\"Equ2\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ2\" name=\"EquationSource\"\u003e\n$$sigmoid(x)=\\frac{1}{{1+{e^{ - x}}}}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e2\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eIn the history of neural network development, Sigmaid function has been used for a long time. Sigmaid function or Tanh function of the same period are smooth, derivable in the whole process, and the output is bounded, which is more advantageous in the expression of the output layer.\u003c/p\u003e \u003cp\u003eCurrently, ReLU (Rectified Linear Unit) is mainly used, as shown in Formula (3):\u003cdiv id=\"Equ3\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ3\" name=\"EquationSource\"\u003e\n$$ReLU(x)=\\left\\{ {\\begin{array}{*{20}{l}} {0,x \\leqslant 0} \\\\ {x,x\u0026gt;0} \\end{array}} \\right.$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e3\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThe computing power of GPU has been continuously enhanced. Now, the deep neural network in 2012 usually uses a single GPU. The AlexNet deep neural network structure of a single GPU is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Implementation principle of single point endogenous control algorithm\u003c/h2\u003e \u003cp\u003eIn the convolutional neural network endogenous control model, as long as there are control points in the convolutional neural network, the endogenous control model can be used to achieve endogenous control. Then, taking the convolutional neural network with N-layer output as the object, the single point endogenous control algorithm is applied to realize the endogenous control.\u003c/p\u003e \u003cp\u003eArguments of the activation function:\u003cdiv id=\"Equ4\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ4\" name=\"EquationSource\"\u003e\n$${z_1}=w_{1}^{ - } \\cdot x+{b_1}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e4\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eDue to the chain structure, the output of the first layer:\u003cdiv id=\"Equ5\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ5\" name=\"EquationSource\"\u003e\n$${h_1}={g_1}\\left( {w_{1}^{ \\top } \\cdot x+{b_1}} \\right)={g_1}\\left( {{z_1}} \\right)$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e5\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThen, the output 2h of the second layer is a function of the output 1h of the first layer. If the second layer is a pooling layer, the output of the second layer is:\u003cdiv id=\"Equ6\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ6\" name=\"EquationSource\"\u003e\n$${h_2}={\\text{~pooling}}{{\\text{~}}_2}\\left( {{h_1}} \\right)$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e6\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThe output of the third layer is a function of the second layer. If the third layer is a convolution layer, the output of the third layer is:\u003cdiv id=\"Equ7\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ7\" name=\"EquationSource\"\u003e\n$${h_3}={g_3}\\left( {w_{3}^{ \\top } \\cdot {h_2}+{b_3}} \\right)={g_3}\\left( {{z_3}} \\right)$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e7\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eIn formula (4), x is the user's input. If the ith hidden layer is convolution layer, ig is the activation function of the layer, iw is the weight of the hidden layer, and ib is the offset of the hidden layer. If the j hidden layer is pooling layer, 1 () j j pooling h\u0026thinsp;\u0026minus;\u0026thinsp;means pooling the output j1h\u0026thinsp;\u0026minus;\u0026thinsp;of j\u0026thinsp;\u0026minus;\u0026thinsp;1 layer. Similarly, layer n\u0026thinsp;\u0026minus;\u0026thinsp;1 outputs:\u003cdiv id=\"Equ8\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ8\" name=\"EquationSource\"\u003e\n$${h_{n - 1}}={g_{n - 1}}\\left( {w_{{n - 1}}^{ - } \\cdot {h_{n - 2}}+{b_{n - 1}}} \\right)={g_{n - 1}}\\left( {{z_{n - 1}}} \\right)$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e8\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eAiming at the security problems caused by the out of control of the convolutional neural network model, this paper designs a single point endogenous control algorithm of the convolutional neural network. The single point endogenous control algorithm is designed according to the operation characteristics of convolutional neural network, which can be applied to neural networks with convolutional layers and has strong generalization characteristics. The convolutional neural network single point endogenous control algorithm is an endogenous control algorithm tightly coupled in the single operation layer control point of the convolutional neural network. It consists of a single point endogenous control function, a single point endogenous control factor and a single point endogenous authorization operator.\u003c/p\u003e \u003cp\u003eAfter the convolutional neural network is trained, a single point endogenous control unit is coupled to any control point, and the N-layer convolutional neural network structure of the single point endogenous control algorithm is applied. At this time, the control point is the active function control point of the first detection layer.\u003c/p\u003e \u003cp\u003eThe output 1h of the first layer of the convolutional neural network controlled by the single point endogenous control algorithm can be expressed as:\u003cdiv id=\"Equ9\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ9\" name=\"EquationSource\"\u003e\n$$h_{1}^{\\prime }={g_1}\\left[ {\\xi \\left( {{z_1},k} \\right)} \\right]$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e9\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eDifferent convolutional neural networks have different structures, so there may not be control points in each layer that can fuse the single point endogenous control algorithm, which leads to different controlled expressions in different operation layers. If the second layer of the convolutional neural network is the pooling layer, and the pooling layer does not contain control points, the output 2h of this layer can be expressed as:\u003cdiv id=\"Equ10\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ10\" name=\"EquationSource\"\u003e\n$$h_{2}^{\\prime }={\\text{~pooling}}{{\\text{~}}_2}\\left( {h_{1}^{\\prime }} \\right)$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e10\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThen, the output\u0026thinsp;\u0026minus;\u0026thinsp;1nh of the n\u0026thinsp;\u0026minus;\u0026thinsp;1 roll can be expressed as:\u003cdiv id=\"Equ11\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ11\" name=\"EquationSource\"\u003e\n$${h_{n - 1}}={g_{n - 1}}\\left( {{z_{n - 1}}} \\right)$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e11\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003e3.3 Implementation principle of mean collapse optimization\u003c/h3\u003e\n\u003cp\u003eAfter the mean collapse optimization scheme is adopted, the weight control factor of the endogenous control algorithm is no longer randomly selected. Let the weight control factor take the expectation of the weight of the layer, then 0 Ω\u0026rarr; ik. Taking N-layer convolutional neural network as the object, the process of applying the multi-point endogenous control algorithm based on mean collapse optimization to realize the endogenous control in each weight layer is analyzed.\u003c/p\u003e \u003cp\u003eAfter mean collapse optimization, the output 1h of the first layer of the convolutional neural network can be expressed as:\u003cdiv id=\"Equ12\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ12\" name=\"EquationSource\"\u003e\n$$h_{1}^{\\prime }={g_1}\\left\\{ {{{\\left[ {\\zeta \\left( {{w_1},E{{\\bar {\\omega }}_1}} \\right)} \\right]}^ - } \\cdot x+{b_1}} \\right\\}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e12\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThen, the output\u0026thinsp;\u0026minus;\u0026thinsp;1nh of the n\u0026thinsp;\u0026minus;\u0026thinsp;1 roll can be expressed as:\u003cdiv id=\"Equ13\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ13\" name=\"EquationSource\"\u003e\n$$h_{{n - 1}}^{\\prime }={g_{n - 1}}\\left\\{ {{{\\left[ {\\zeta \\left( {{w_{n - 1}},E{{\\bar {\\omega }}_{n - 1}}} \\right)} \\right]}^ - } \\cdot h_{{n - 2}}^{\\prime }+{b_{n - 1}}} \\right\\}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e13\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eWhen the endogenous control function is a linear operation 61562ζ (x, k)\u0026thinsp;=\u0026thinsp;x ζ k, the mean collapse optimization scheme first performs a collapse operation with a collapse ratio of ཛྷ ik for the selected convolutional neural network weight control point. At this time, the controlled layer weight \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\omega _{i}^{\\prime }={\\omega _i} \\cdot {k_{{\\omega _i}}}={\\omega _i} \\cdot E{\\bar {\\omega }_i} \\to 0\\)\u003c/span\u003e\u003c/span\u003e of the controlled convolution neural network will output:\u003cdiv id=\"Equ14\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ14\" name=\"EquationSource\"\u003e\n$$h_{{n - 1}}^{\\prime } \\approx {g_{n - 1}}\\left\\{ {0 \\cdot h_{{n - 2}}^{\\prime }+{b_{n - 1}}} \\right\\}={g_{n - 1}}\\left( {{b_{n - 1}}} \\right)$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e14\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThe output of the n\u0026thinsp;\u0026minus;\u0026thinsp;1 layer of the original convolutional neural network is:\u003cdiv id=\"Equ15\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ15\" name=\"EquationSource\"\u003e\n$${h_{n - 1}}={g_{n - 1}}\\left( {w_{{n - 1}}^{ \\top } \\cdot {h_{n - 2}}+{b_{n - 1}}} \\right)$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e15\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eBy comparing equations (\u003cspan refid=\"Equ13\" class=\"InternalRef\"\u003e13\u003c/span\u003e) and (\u003cspan refid=\"Equ14\" class=\"InternalRef\"\u003e14\u003c/span\u003e), it can be found that the output 1n-h and n-1h differ greatly after mean collapse optimization at the weight control point, and the theoretical control effect is good.\u003c/p\u003e \u003cp\u003eWhen using the mean collapse optimization scheme to optimize the endogenous control algorithm, it is necessary to propose a performance index to measure the quality of the mean collapse optimization scheme. In addition to intuitively analyzing the controlled precision of convolutional neural network, the performance indicators also include: algorithm control rate, time cost ratio and algorithm utility ratio.\u003c/p\u003e"},{"header":"Experimental results and analysis","content":"\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Optimization experiment analysis\u003c/h2\u003e \u003cp\u003eLe Net-5 is selected as the experimental network to study the control effect of the optimization scheme of the endogenous control algorithm based on mean collapse at the weight control points. The weight control points of each convolution layer of LeNet-5 are selected for comparative experiments. One group applies a single point endogenous control algorithm to the single weight control points of C1, C3, and C5 convolutions, and applies a multiple point endogenous control algorithm to multiple weight control points; In the other group, the types of endogenous control algorithms for each weight control point remain unchanged, but they are all optimized using an optimization scheme based on mean collapse, and the experimental results are recorded, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. For LeNet-5 convolutional neural network, the optimization scheme based on mean value collapse can effectively improve the control accuracy of the endogenous control algorithm and realize the optimization of the endogenous control algorithm.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe deeper VGG-16 is selected as the experimental network to study the performance of the mean collapse optimization scheme. For VGG-16, apply single point endogenous control algorithm, multi-point endogenous control algorithm, single point endogenous control algorithm after mean collapse optimization and multi-point endogenous control algorithm after mean collapse optimization respectively, and record the experimental results, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e. For VGG-16 convolutional neural network, the optimization scheme based on mean collapse can optimize the endogenous control algorithm.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Optimization of endogenous control algorithm based on giant parameters\u003c/h2\u003e \u003cp\u003eAI simulates the combination of artificial processes through the graphic and image information of elements, which is determined by its adaptive characteristics facing the complex semantic environment. Compared with the fixed description of target features by traditional information classification methods oriented to general application scenarios, such as license plate recognition, character recognition, voice recognition, etc., and the technical limitations of only being able to single classify objects such as text and voice, the deep learning network can, as required, implement different definitions of the same object in a complex semantic environment by carrying out classification weight statistics on multi category labels associated with the target. It can not only recognize its general concept definition, but also classify its implicit or extended definition as needed, so that it can achieve clustering based on complex criteria, which is also the technical basis for current AI methods to effectively implement design behavior. The experimental results are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAccording to the analysis of the above two experimental results, for the controlled LeNet-5, the prediction accuracy of unauthorized users is lower when the endogenous control algorithm is applied to the deeper weight control points; For controlled VGG-16, endogenous control algorithm is applied to deeper weight control points after the first layer and C17, and the prediction accuracy of unauthorized users is lower. It can be seen that endogenous control of deeper bias control points has greater impact on the output of convolutional neural network. Compared with other control points, the time cost of applying the endogenous control algorithm to the offset control points in the convolution layer is smaller, because the parameters of the offset control points are often smaller than the weight control points. If the offset matrix of each layer of VGG-16 is different, the maximum parameter quantity of each layer of offset matrix is 4096 \u0026times; 1. The parameter amount of the weight can reach tens of thousands of times of the offset parameter amount at most. Therefore, the cost of fusing offset control factors and corresponding control algorithms at offset control points is less than that of fusing weight control factors of equal scale at the same level of weight control points.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Experimental results of visual design under big data\u003c/h2\u003e \u003cp\u003eIn the past, visual design was limited to paper media, and the communication was weak, so the flexibility in content layout was limited by the size of paper printing. Today, new media, as the main carrier of information communication, such as public environment display screen, smart phone, tablet computer, etc., have diversified communication methods. With the development of Internet technology, adaptive page technology is becoming more and more mature. The same content can get the best typesetting effect under different size browsers, and the flexibility is significantly improved, but at the same time, it puts forward higher requirements for the professional quality and professional level of designers. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eImage is the most intuitive visual trigger, which can convey a lot of information in a short time. The design based on illustration art can be applied to the publicity and promotion of many themes in new media, and has the effect of \"1\u0026thinsp;+\u0026thinsp;1 ༞ 2\". With the cross boundary and integration of various designs, the promotion and application of illustrations today also need to comply with the social progress and market demand. At present, the illustration is mainly used online for enterprise IP, H5, We Media (WeChat official account, Tiktok, Sina Weibo, etc.), large-scale enterprises and product theme promotion, etc.\u003c/p\u003e \u003cp\u003eImmersion is to let people focus on the current situation created by the designer and feel happy and satisfied, while forgetting the real world situation. \"Tiktok\" is typical of immersive experience, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e. With the continuous expansion of mobile Internet communication channels, \"two micro and one\" has become the main battlefield of media integration and development. As of June 2020, short video has surpassed instant messaging by 110 minutes per person per day. When analyzing the immersive design of \"Tiktok\" from the perspective of visual design, it has two obvious visual characteristics: one is the de modal design of floating layer operation. During the video viewing process, when making comments, forwarding and other interactive behaviors, the floating layer operation of interactive button floating is a semi transparent temporary floating layer, and without \"masking\" processing, it will not interrupt the current video playing behavior.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThe application of artificial intelligence has had a profound impact on the development of today's design field. However, from the analysis of the specific implementation methods and effectiveness of the current technology itself, its essence is to replace the relatively basic element combination and example flow in the design process. Moreover, due to the limitations of technology, resources and other aspects, it cannot fundamentally replace the design position of people, so its role and impact are only concentrated on the tool level. However, its self-adaptive association of multiple information provides a good experimental environment and implementation basis for the expansion of design methods. At the same time, the task of design has changed from formal innovation to the understanding and interpretation of form-oriented definition and graphic connotation, as well as the innovation of new technology-oriented application forms. Two kinds of convolutional neural network endogenous control algorithms are designed, and their performance is tested on different convolutional neural networks to study the performance characteristics of different endogenous control algorithms. Aiming at the shortcomings of two endogenous control algorithms, three optimization schemes are proposed and the algorithm performance is studied. The experiment shows that the endogenous control algorithm of the optimization scheme can achieve complete control of the convolutional neural network model.\u003c/p\u003e "},{"header":"Declarations","content":"\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eData availability\u003c/h2\u003e \u003cp\u003eThe figures used to support the findings of this study are included in the article.\u003c/p\u003e \u003c/div\u003e\u003cp\u003e \u003ch2\u003eConflicts of interest\u003c/h2\u003e \u003cp\u003eThe authors declare that they have no conflicts of interest.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eFunding statement\u003c/h2\u003e \u003cp\u003eThis work was not supported by any funds.\u003c/p\u003e\u003ch2\u003eAcknowledgements\u003c/h2\u003e \u003cp\u003eThe authors would like to show sincere thanks to those techniques who have contributed to this research.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eOh BD, Lee YK, Song HJ (2021) Age group classification to identify the progress of language development based on convolutional neural networks. J Intell Fuzzy Systems: Appl Eng Technol 36(9):2508\u0026ndash;2515\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAbudureheman A, Nilupaer A, He Y (2020) Performance evaluation of enterprises' innovation capacity based on fuzzy system model and convolutional neural network. J Intell fuzzy systems: Appl Eng Technol 521(7553):436\u0026ndash;444\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRevi KR, Wilscy M, Antony R (2021) Portrait photography splicing detection using ensemble of convolutional neural networks. J Intell Fuzzy Systems: Appl Eng Technol 5(4):115\u0026ndash;133\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMu T, Lin K, Zhang H (2021) Sanitizing hidden activations for improving adversarial robustness of convolutional neural networks. J Intell Fuzzy Systems: Appl Eng Technol 40(6):1229\u0026ndash;1251\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen YM, Chen YJ, Tsai YK (2021) Classification of human electrocardiograms by multi-layer convolutional neural network and hyperparameter optimization. J Intell Fuzzy Systems: Appl Eng Technol 160(1):106\u0026ndash;154\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShashank R, Prasad, Srinivasa PP (2021) Intelligent fault diagnosis ofRolling-element bearing using 1-D Deep Convolutional Neural Networks. Int J COMADEM 36(4):193\u0026ndash;202\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eProsvirin AE, Maliuk AS, Kim JM (2022) Intelligent rubbing fault identification using multivariate signals and a multivariate one-dimensional convolutional neural network. Expert Syst Application 86(11):2278\u0026ndash;2324\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSanchez SG, Muns GR, Bocanegra C (2022) AirNN: Neural Networks with Over-the-Air Convolution via Reconfigurable Intelligent Surfaces. 20(3):97\u0026ndash;112\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang Y, Yan J, Jing Q (2021) A novel adversarial transfer learning in deep convolutional neural network for intelligent diagnosis of gas-insulated switchgear insulation defect. IET generation, transmission \u0026amp; distribution, 2(5):359\u0026ndash;366\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang MY, Kumaar S, Ye L (2020) Real-time Semantic Segmentation with Context Aggregation Network. 7(2):279\u0026ndash;301\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSavary E, Rojas K, Maus M (2021) A search for galaxy-scale strong gravitational lenses in the Ultraviolet Near Infrared Optical Northern Survey (UNIONS). 41(6):93\u0026ndash;94\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShafiee MJ, Famouri M, Bathla G, TinyDefectNet (2021) Highly Compact Deep Neural Network Architecture for High-Throughput Manufacturing Visual Quality Inspection. 4(4):79\u0026ndash;82\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAltan D, Sariel S (2022) CLUE-AI: A Convolutional Three-stream Anomaly Identification Framework for Robot Manipulation. 14(1):1\u0026ndash;17\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMeletis P (2022) Towards holistic scene understanding: Semantic segmentation and beyond. 20(12):1657\u0026ndash;1664\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKrishna S, Ss S, Krishnamoorthy S Stain Normalized Breast Histopathology Image Recognition using Convolutional Neural Networks for Cancer Detection. 20229,40(6):141\u0026ndash;146\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXompero A, Donaher S, Iashin V (2021) The CORSMAL benchmark for the prediction of the properties of containers. 25(5):33\u0026ndash;48\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhao X, Zhang Y (2022) An intelligent diagnosis method of rolling bearing based on multi-scale residual shrinkage convolutional neural network. Meas Sci Technol 31(10):11\u0026ndash;20\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJiangquan ZHANG, Yi (2020) A new bearing fault diagnosis method based on modified convolutional neural networks. Chin J Aeronaut 167(2):54\u0026ndash;62\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu K, Zhao Z, Shi P (2022) Real-time surgical tool detection in computer-aided surgery based on enhanced feature-fusion convolutional neural network. J Comput Des Eng 43(8):173\u0026ndash;175\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDehshibi MM, Ashtari-Majlan M, Adhane G (2022) ADVISE: ADaptive Feature Relevance and VISual Explanations for Convolutional Neural Networks. 40(1):17\u0026ndash;26\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFang J, Zhou X, Wang P (2021) Densely connected Siamese network visual tracking. Industrial Robot: the international journal of robotics research and application 48(5):680\u0026ndash;687\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSpagnolo F, Perri S, Corsonello P (2022) Approximate Down-Sampling Strategy for Power-Constrained Intelligent Systems. IEEE Access 38(24):30\u0026ndash;36\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePustokhina IV, Pustokhin DA, Rodrigues J (2020) Automatic Vehicle License Plate Recognition Using Optimal K-Means With Convolutional Neural Network for Intelligent Transportation Systems. IEEE Access 8(16):92907\u0026ndash;92917\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLin Y, Zhong Q, Sun H (2020) A Pointer Type Instrument Intelligent Reading System Design Based on Convolutional Neural Networks. Front Phys 89(11):22\u0026ndash;27\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMa Y, Wen G, Cheng S (2022) Multimodal convolutional neural network model with information fusion for intelligent fault diagnosis in rotating machinery. IOP Publishing Ltd 10(5):705\u0026ndash;710\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Artificial intelligence, Convolution neural network, Intelligent visual design, big data","lastPublishedDoi":"10.21203/rs.3.rs-2997795/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-2997795/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eIn recent years, artificial intelligence has been initially applied to the visual design process, which has had a certain impact on the traditional design methods and working modes. Through the analysis of the working principle of the key technology of artificial intelligence design, the generation antagonism network (GAN), this paper summarizes the advantages and limitations of the existing automatic design methods, and puts forward the transformation path and specific methods of visual design working mode oriented to automation design efficiency optimization and new technology application form innovation under the background of new technology. With the rapid development of cloud computing and artificial intelligence, there are more and more applications in the cloud based on artificial intelligence as a service. Among them, artificial intelligence based on convolutional neural network is widely used. Convolution neural network endogenous control is the first fusion of endogenous security ideas in deep neural network model security control. Based on the core operation structure of convolutional neural network, the key control points of the endogenous control model are proposed; By tightly coupling the endogenous control unit at the control point, a convolutional neural network endogenous control model is designed and proposed.\u003c/p\u003e","manuscriptTitle":"Construction of intelligent visual design system based on convolutional neural network under the background of big data","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-03-12 17:43:11","doi":"10.21203/rs.3.rs-2997795/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"7526e9c6-fb3c-4519-b52c-5dc3a71ed1d0","owner":[],"postedDate":"March 12th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-06-18T21:16:47+00:00","versionOfRecord":[],"versionCreatedAt":"2024-03-12 17:43:11","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-2997795","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-2997795","identity":"rs-2997795","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}