Realization of normal temperature detection through visible light images by Retinex-CNN

Realization of normal temperature detection through visible light images by Retinex-CNN | 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 Realization of normal temperature detection through visible light images by Retinex-CNN Jiayi Zhu, Zhimin He, Cheng Huang, Jun Zeng, Huichuan Lin, Fuchang Chen, and 5 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-4454734/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 Contactless detection of a target’s temperature of normal range (30°C ~ 150°C), which is based on its visible light images rather than infrared images, is a promising technology. Visible light imaging is primarily based on the visible light reflected from a target, rather than the thermal radiation of its own. The main challenge of contactless normal temperature detection through visible light images exists in the interference introduced by the variation of incident illumination. To solve this, a Retinex convolutional neural network (Retinex-CNN) was proposed in this paper, which was based on the convolutional neural network and Retinex algorithm. This network reduces the interference introduced by illumination variations and effectively improves the accuracy of the temperature detection based on visible light images. The temperature detection results of the Retinex-CNN shows that this network exhibits favorable generalization capability in terms of illumination variations, that is, it is still able to accurately detect temperature as a target is imaged in an illumination condition which is significantly different from that of the images in the training set. In this case, the Retinex-CNN obtains an average absolute error of 2.6°C, a value that is 6.89°C lower than that obtained by the CNN. non-contact normal temperature detection visible light image Retinex algorithm convolutional neural network generalization capability Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 1 Introduction Thermal radiation of objects at normal temperature (30°C ~ 150°C) is infrared electromagnetic waves (6µm ~ 13µm), which are not detectable by the sensors of cameras developed for visible light imaging. Consequently, in non-contact normal temperature detection infrared imaging is the mainstream method, though the infrared cameras are expensive and have low pixels. The principle of thermal modulation of the reflected light indicates that changes in the temperature of an object will cause changes in the surface light reflectance [ 1 – 4 ]. Therefore, in the presence of the incident of visible light, contactless normal temperature detection, in fact, can be achieved by using the surface reflects light of the target, i.e., visible light image of the target. The key technology of realizing this new detection method is how to extract the temperature information exactly from the visible light images of the target. In recent years, with the rise of machine learning techniques, researchers have tried to use the method of machine learning to extract the temperature information from the visible light images. Such as, a prediction model describing the connection between image features and temperature was proposed by M. Wang et al. [ 5 ]; W. Li et al. proposed a full-wavelength multi-incidence angle thermally modulated light reflection temperature measurement mechanism based on visible images [ 6 ]; W. Du reported the visible light thermometry in sunlight conditions using machine learning [ 7 ]; X. Nie et al. proposed a similar machine learning approach in sunlight environment while using videos from surveillance camera as the data for temperature information extraction [ 8 ]. With human assistance to recognize the key features for classification, the research works stated above shows that temperature information can be extracted from the visible light images by using the machine learning. Deep learning is a branch of machine learning, which is a machine learning method based on neural networks [ 9 ]. Without human intervention deep learning can automatically learn features and patterns from raw data and use those features and patterns to classify or predict the data. Moreover, unlike traditional machine learning, deep learning is able to learn features with multiple layers of abstraction, allowing it to work with more complex and high-dimensional data [ 10 – 12 ]. Therefore, exploring and then employing the method of deep learning to extract the temperature information from the visible optical image of the target will make the temperature detection based on visible optical image more convenient and accurate. Convolutional neural network (CNN) is a deep learning model that stands out for its superior ability to extract features in localised regions. In image information extraction, CNN show significant advantages through convolutional operations, parameter sharing and multilayer structure [ 13 ]. In practical applications, CNN are widely used in image classification [ 14 ], object detection [ 15 ], face recognition [ 16 ], medical image analysis [ 17 ], industrial defect detection [ 18 ], etc., showing excellent performance and adaptability. This makes CNN a mainstream deep learning model for processing various vision tasks. Consequently, to realize the temperature extraction from visible light image, CNN would be a very suitable neural network model. One other thing to be noted is that the change of environmental illuminance is able to make a difference of image information of the object. While this disturbance brought about by changes in environmental illuminance will further bring great difficulties to the temperature detection based on the visible light image of the target. According to the principle of thermal modulation of the reflected light, the temperature change of the target changes only the reflectivity of the target surface actually [ 5 ]. If the reflectance distribution map can be separated from the original visible light image, then the temperature detection by using the reflectance distribution map can avoid the influence brought by the change of ambient illumination, so that the temperature detection based on visible light image would more accurate. Therefore, in order to extract the temperature information from visible light images accurately, combining with Retinex algorithm which emphasizes the separation of reflection image and illumination image from original visible light image, a Retinex convolutional neural network (Retinex-CNN) is constructed in this paper. In the subsequent chapters of this paper, we will use Retinex-CNN for temperature recognition of visible image. 2 Experimental data acquisition In order to use the Retinex-CNN to detect the temperature of the target from its visible light image, it is necessary to collect the images corresponding to the target with different temperatures and apply them to train the neural network. Thus, the target which was a copper plate was attached on a temperature-controlled hotplate. Copper is an important metal material that is often used in power grid, so we take copper sheet as the target object for the temperature detection. The hotplate employed is a stainless steel electric hot plate of model type of DB-1, whose controllable temperature range and temperature control precision are from room temperature to 300°C and 0.1°C. For visible light imaging, a Canon EOS 200D II camera was used to capture video of the target, and a still image in 8-bit JPG format was extracted from the video. For the sake of employing Retinex-CNN to realize real-time monitoring of the target temperature, the video images are collected to train the Retinex-CNN. As the experimental device diagram (i.e., Fig. 1 ) shown, camera is placed in front of the copper sheet, sun light is used as the lighting source. When the temperature of the copper sheet is heated up to 130℃, the hotplate is turned off and the video model of the camera is turned on. As the temperature of the copper sheet drops, visible light optical images of the copper sheet at different temperatures will be recorded by the camera. Moreover, a spectral color illuminance meter (SPIC-200) was used to monitor the color temperature and illuminance of the ambient light. As shown in inset of Fig. 1 , the incident light illuminance has been changing throughout the experiment. A total of 46000 target images at different temperatures were collected as the temperature of the target dropped from 130°C to 40°C. The temperature interval was 0.1°C from 40 to 130°C, and the number of target image at each temperature is about 50. Moreover, a series of three experiments for the image capturing were carried out under different illuminance of the ambient light. These visible light images of the target at different temperatures would be used to train and test the Retinex-CNN. 3 Theoretical analysis of the Retinex-CNN In optical imaging, the information actually recorded by the camera is the illuminance incident on the target multiplied by the reflectivity of the target's surface, according to the principle of optical imaging. Thus the target image \(S(x,y)\) can be expressed as $$S(x,y)=L(x,y) \times R(x,y)$$ 1 . Where \(R(x,y)\) is the reflectivity of the target's surface and \(L(x,y)\) is the light illuminance incident on the target. It is obvious that the target image changes with the incident light illuminance, even if the surface reflectivity does not change. In another word, the variation in illuminance will introduce interference to the temperature detection based on visible light imaging. Retinex is an image processing algorithm based on visual perception, and its theory emphasizes the separation of the two key components of the image reflection and illumination, so that the final image highlights the reflection component. Applying this algorithm to the target images for visible light temperature detection, the reflectivity distribution of the target surface, which is independent of the incident illuminance, can be obtained, so that the interference due to the variation of background illuminance can be eliminated. To validate this idea, target images corresponding to illuminances of 692lux, 800lux, and 1039lux were processed through the Retinex algorithm and the effectiveness of the algorithm is shown in Fig. 2 . The frequency distribution of gray values of the images for the target temperature of 80°C before processing are illustrated in Fig. 2 (a), (b) and (c). The curves for all three RGB channels differ greatly from each other, indicating that the variation of illuminance has a significant effect on the image of the target at the same temperature. Figure 2 (c), (d) and (e) are the frequency distribution curves of gray values of the images after processing, in which the curves for all three RGB channels remain stable and do not change with the illuminance. This proves that the interference of the incident light illuminance on the target information in the visible light image can indeed be eliminated using the Retinex algorithm. The capability of the Retinex algorithm to exclude the interference of changes in incident illuminance on the image feature of the mean gray value was also examined. The mean gray values of the original images shot for different target temperatures and in different incident illuminances are plotted in Fig. 3 (a), and the mean gray values of the corresponding images processed by the Retinex algorithm are plotted in Fig. 3 (b). It is obvious that the mean gray value changes significantly with the illuminance, and the magnitude of the change even exceeds the magnitude of the change in the mean gray value due to the change in target temperature. After the images were processed by the Retinex algorithm, the interference of the incident illuminance on the mean gray value of the target image is greatly weakened, as shown in Fig. 3 (b), and the effect of the change in target temperature on the mean gray value of the image is highlighted. Therefore, if the mean gray value is used as the basis for measuring the temperature of the target in an image, the Retinex algorithm is also effective in avoiding the interference caused by changes in incident illuminance. In summary, the introduction of the Retinex algorithm will effectively exclude the errors due to changes in incident illuminance when utilizing the visible light images for temperature detection. The above discussions suggest that the Retinex algorithm is effective in eliminating the interference of the variation in incident illuminance on the temperature features in the visible light images of a target. Therefore, in the approach of temperature detection using visible light optical images and the convolutional neural network, the Retinex algorithm can be incorporated into the network to improve its accuracy in temperature detection. Based on this, after combining Retinex algorithm an upgraded CNN which is named as Retinex-CNN is constructed in this paper, and its network structure is shown in Fig. 2 . Firstly, the target images at different temperature are cropped into 150×150 pixels and then being inputted it into the Retinex-CNN. The auto-MSRCR module of the Retinex algorithm is placed at the front of the Retinex-CNN, so before extracting the temperature information the auto-MSRCR module has separated the target image into reflection image and illumination image. Retaining the reflection image of the target, and then input it into the convolutional layers, which consists of 32, 64, and 128 convolutional kernels. Take advantage of a 3×3 kernel size, the convolutional layers gradually extracts the feature information from the reflection image with a stride of 1. After each convolutional layer, ReLU activation is applied to enhance the network's non-linear fitting capabilities. Moreover, a 2×2 max-pooling layers was added after each convolutional layer to reduce feature map dimensionality and extract significant features. After the features of the reflection image are extracted by the convolutional layer, the features are integrated by a six-layer fully connected layer. Following, the fully connected layer outputs a dedicated neuron for regression tasks, and the network generates continuous values reflecting prediction outcomes. By calculating the mean squared error (MSE) between the predicted results and the labels, the obtaining loss value is sent back to the Adam optimizer, and then updating the model parameters until the desired level of convergence is achieved. In the following chapters of this paper, the collected target images at different temperatures will be used to train the Retinex-CNN, and the trained Retinex-CNN will be used to detect the temperature from the target's visible light image. 4 Temperature detection by Retinex-CNN Three groups of visible light images of the target at different temperatures were collected and a corresponding ambient light illuminance for each group was measured as 692lux, 800lux and 1039lux, respectively. Note that, the image collection process lasted for a certain period of time, and the actual ambient illuminance varied within a certain range due to uncontrollable changes in natural conditions. The above illuminance values of 692lux, 800lux and 1039lux are representatives of the range of ambient illuminance change during three image collection processes, rather than the exact ambient illuminances of each of the images in three groups. In the following discussion, the three illuminance values of 692lux, 800lux and 1039lux would be used as names to distinguish the collected three groups of images. The images in the group of 692lux were chosen to form the dataset for the Retinex-CNN, which were divided into a training set, a validation set, and a testing set in the ratio of 70%, 20%, and 10%. For the purpose of comparison, the same dataset was also used for the training, validation and testing of the CNN. The results of temperature detection by the CNN and the Retinex-CNN using the visible light images in the group of 692lux illuminance are shown in Fig. 5 (a) and (b). The same procedure was also done for the group of 1093lux and the results are illustrated in Fig. 5 (c) and (d). The mean absolute error (MAE) is used as a measure of the accuracy of the temperature detection and it is expressed as $$\text{M}\text{A}\text{E}=\left( {\frac{1}{n}} \right)\sum\nolimits_{i}^{n} {\left| {{y_i} - {x_i}} \right|}$$ 2 where y is the real temperature of the target and x is the target temperature detected by the neural network from the visible light image. From Fig. 5 , it can be seen that the temperature detection of the Retinex-CNN is superior to that of the CNN. For 692lux illuminance, the Retinex-CNN reduces the mean absolute error of temperature detection from 2.6522°C to 2.0479°C. A more significant improvement in the accuracy of temperature detection can be observed for 1039 lux illuminance and the mean absolute error decreases from 2.9541°C to 1.9459°C. These results show that the introduction of the Retinex algorithm, which is incorporated into Retinex-CNN, can significantly improve the performance of temperature detection based on the CNN and the visible light images achieved under natural illumination conditions. This is because the Retinex algorithm can exclude the interference of ambient illuminance variation on the temperature feature in an image. To further validate the illuminance generalization capability of the Retinex-CNN in target temperature detection, all the images in the groups of 692lux and 1039lux were merged together to form the training set and the validation set, and the images in the group of 800lux comprised the testing set used to test the performance of the trained Retinex-CNN model. These data sets were also used for the training, validation and testing of the CNN model for comparison. The results of temperature detection using these two neural network models are shown in Fig. 6 . The images in the group of 800lux were only used for testing and none of them was included in the training set. The huge difference in the background illuminance between the images in the training set and the images in the testing set leads to the deterioration of the accuracy of temperature detection and the mean absolute error reaches as high as 9.5146°C, as shown in Fig. 6 (a). Comparing Fig. 6 (b) with (a), the Retinex-CNN model demonstrates significantly superior performance to the CNN model with an mean absolute error of 2.6238°C, which represents a 6.89°C detection error reduction. This indicates that the Retinex-CNN proposed in this paper has excellent illuminance generalization capability, and its use in temperature detection based on visible light images can effectively reduce the interference caused by changes in background illumination, thus improving the accuracy of temperature detection. This improvement of illuminance generalization ability would be a key technology for the application of the method of normal temperature detection which is based on visible image. 5 Conclusion A Retinex-CNN is proposed in this paper to improve the performance of normal temperature detection based on visible light images. The Retinex algorithm is able to extract an intensity distribution image of the incident light and an intensity distribution image of the reflected light from an image, and the image of reflected light is fed to the trained CNN for target temperature detection. In this way, the interference of ambient illumination variation on the temperature detection is minimized and the performance is therefore improved. The experimental results show that Retinex-CNN has superior detection accuracy compared to the CNN, especially when the image used for temperature detection differs significantly from the images used for neutral network training in terms of background illumination. A temperature detection error of less than 3°C was obtained by the Retinex-CNN, which is about 7°C detection error reduction comparing with the detection result of the CNN. It can be concluded that the Retinex-CNN proposed in this paper has excellent illuminance generalization capability for normal temperature detection based on visible light images, which is ideal for outdoor temperature detection applications where background illumination is difficult to control. Declarations Competing interests: The authors declare no competing interests. Funding: This research was funded by National Natural Science Foundation of China (NSFC), grant numbers 61975072 and 12174173; Natural Science Foundation of Fujian province, grant numbers 2022H0023, 2022J02047 and 2022G02006; Natural Science Foundation of Zhangzhou City, grant number ZZ2023J20. Author Contribution Author contributions: JZ and HL wrote the main manuscript text, ZM is responsible for data collectionall, CH and FC are conducting network construction and development, JZ , CY and YL is responsible for editing and polishing the article, HC, YZ and JP is responsible for conceptualization and methodology, the authors reviewed the manuscript. References Z. Zha, J. Zhu, X. Yang, X. Huang, H. Guo, A. Xie, X. Lu, Z. Fu, in 2022 23rd International Conference on Electronic Packaging Technology (ICEPT), pp.1–5(2022) H. Yang, D. Wang, Q. Yin, Y. Gao, L. Zheng, in 2022 23rd International Conference on Electronic Packaging Technology (ICEPT), pp.1–6(2022) L. Farbaniec, D.E. Eakin, Rev. Sci. Inst. 94 , 034902 (2023) S. Alajlouni, K. Maize, A. Shakouri, in 2022 21st IEEE Intersociety Conference on Thermal and Thermomechanical Phenomena in Electronic Systems (iTherm), pp. 1–10(2022) M. Wang, Research on temperature measurement method based on visible image and machine learning[D], Ph.D. dissertation, Dept. of Optical Engineering, Huazhong University of Science and Technology, Wuhan, China, 2020 W.M. Li, Research on the mechanism and method optimization of artificial intelligence temperature measurement based on visible light images[D] (Huazhong University of Science and Technology, 2021) W. Du, Q. Ye, Z. Yuan, C. Li, High. Volt Appar. 58 , 221–229 (2022). (in Chinese) X. Nie, Q. Ye, Z. Yuan, M. Han, in 16th Annual Conference of China Electrotechnical Society, pp. 616–627(2022) Y. LeCun, Y. Bengio, G. Hinton, nature, 521, 436–444 (2015) J. Yang, P. Qiao, Y. Li, N. Wang, Stat. Decis. Mak. 35 , 36–40 (2019) K. Yu, L. Jia, Y. Chen, W. Xu, Comput. Res. Dev. 50 , 1799–1804 (2013) Y. Zheng, G. Li, Y. Li, Comput. Eng. Appl. 55 , 20–36 (2019) S. Albawi, T.A. Mohammed, S. Al-Zawi, in 2017 International Conference on Engineering and Technology, pp. 1–6(2017) X. Lei, H. Pan, X. Huang, IEEE Access. 7 , 124087–124095 (2019) S. Gidaris, N. Komodakis, in the IEEE international conference on computer vision, pp.1134–1142(2015) S. Lawrence, C.L. Giles, A.C. Tsoi, A.D. Back, IEEE Trans. Neural Netw. 8 , 98–113 (1997) G. Litjens, T. Kooi, B.E. Bejnordi, A.A.A. Setio, F. Ciompi, M. Ghafoorian, J.A.W.M. Laak, B.V. Ginneken, C.I. Sάnches, Med. Image Anal. 42 , 60–88 (2017) X. Tao, D. Zhang, W. Ma, X. Liu, D. Xu, Appl. Sci. 8 , 1575 (2018) Additional Declarations No competing interests reported. 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-4454734","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":305041569,"identity":"76f68cca-19e7-4cb5-890b-110b93548305","order_by":0,"name":"Jiayi Zhu","email":"","orcid":"","institution":"Minnan Normal University","correspondingAuthor":false,"prefix":"","firstName":"Jiayi","middleName":"","lastName":"Zhu","suffix":""},{"id":305041571,"identity":"9564c125-3695-4dec-9c9a-f24bed1d242c","order_by":1,"name":"Zhimin He","email":"","orcid":"","institution":"Minnan Normal University","correspondingAuthor":false,"prefix":"","firstName":"Zhimin","middleName":"","lastName":"He","suffix":""},{"id":305041572,"identity":"93e83acb-142a-437c-86ec-caf68fd0669b","order_by":2,"name":"Cheng Huang","email":"","orcid":"","institution":"Minnan Normal University","correspondingAuthor":false,"prefix":"","firstName":"Cheng","middleName":"","lastName":"Huang","suffix":""},{"id":305041574,"identity":"98e2dee0-1a59-41d7-93d1-aea903f435ae","order_by":3,"name":"Jun Zeng","email":"","orcid":"","institution":"Minnan Normal University","correspondingAuthor":false,"prefix":"","firstName":"Jun","middleName":"","lastName":"Zeng","suffix":""},{"id":305041579,"identity":"97a845d6-293c-4b52-9060-092f4b3d5e4b","order_by":4,"name":"Huichuan Lin","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAxklEQVRIiWNgGAWjYBACPgkGhgMfDCTkoHxmwlrYJBgYD84osDEmSQvzYZ4PaYkNxGuRbn5wcIbB4fS1s3uPSTBUWCc2sJ89gF+LzDEDoF8O5267cy5NguFMemIDT14CAYflMIBsyd12I8dMgrHtcGKDBI8BQS2HeYAOMwNr+Ue8lrQEiJYGorSkGQAdZmMIdJixRcKxdOM2nhz8Wvglkh9/+PBHQh5oi+GNDzXWsv3sZ/BrQQUJIHtJUD8KRsEoGAWjAAcAAIitRJIx52s9AAAAAElFTkSuQmCC","orcid":"","institution":"Minnan Normal University","correspondingAuthor":true,"prefix":"","firstName":"Huichuan","middleName":"","lastName":"Lin","suffix":""},{"id":305041580,"identity":"050bb2fb-1e38-4681-adec-6ec3c3a75239","order_by":5,"name":"Fuchang Chen","email":"","orcid":"","institution":"Minnan Normal University","correspondingAuthor":false,"prefix":"","firstName":"Fuchang","middleName":"","lastName":"Chen","suffix":""},{"id":305041581,"identity":"d40b522b-d48e-499a-8bdf-9cb89038663d","order_by":6,"name":"Chaoqun Yu","email":"","orcid":"","institution":"Minnan Normal University","correspondingAuthor":false,"prefix":"","firstName":"Chaoqun","middleName":"","lastName":"Yu","suffix":""},{"id":305041582,"identity":"7bc427ac-0694-46f1-9310-04ac0ccfc983","order_by":7,"name":"Yan Li","email":"","orcid":"","institution":"Minnan Normal University","correspondingAuthor":false,"prefix":"","firstName":"Yan","middleName":"","lastName":"Li","suffix":""},{"id":305041586,"identity":"5a50ef6f-e185-4d61-92d5-ed9427d1e528","order_by":8,"name":"Yongtao Zhang","email":"","orcid":"","institution":"Minnan Normal University","correspondingAuthor":false,"prefix":"","firstName":"Yongtao","middleName":"","lastName":"Zhang","suffix":""},{"id":305041588,"identity":"452c3bec-d4fa-446d-90a8-47b45fa7626b","order_by":9,"name":"Huanting Chen","email":"","orcid":"","institution":"Minnan Normal University","correspondingAuthor":false,"prefix":"","firstName":"Huanting","middleName":"","lastName":"Chen","suffix":""},{"id":305041589,"identity":"62e5e12f-241b-4f75-a47b-b9734206bbc2","order_by":10,"name":"Jixiong Pu","email":"","orcid":"","institution":"Huaqiao University","correspondingAuthor":false,"prefix":"","firstName":"Jixiong","middleName":"","lastName":"Pu","suffix":""}],"badges":[],"createdAt":"2024-05-21 12:14:17","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-4454734/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-4454734/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":57872912,"identity":"15399fce-56fa-41b6-9bd9-1afee4460a84","added_by":"auto","created_at":"2024-06-06 18:34:59","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":4258432,"visible":true,"origin":"","legend":"\u003cp\u003eSchematic of imaging of a copper sheet at different temperatures.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-4454734/v1/44d5836c6bfda1f93843ddbd.png"},{"id":57872908,"identity":"fb07ef50-3b93-436c-91b5-09a5462cf6ed","added_by":"auto","created_at":"2024-06-06 18:34:59","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":451873,"visible":true,"origin":"","legend":"\u003cp\u003eThe frequency distribution of gray values of the images for the target at the same temperature under different illuminations. (a), (b), and (c) are the frequency distribution of gray values before being processed for all three RGB channels; (c), (d), and (f) are the frequency distribution of gray values after being processed for all three RGB channels.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-4454734/v1/99701b9df65cd76dc8fa3339.png"},{"id":57872911,"identity":"af727d4a-4238-4a5f-a436-6893dc30907a","added_by":"auto","created_at":"2024-06-06 18:34:59","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":233689,"visible":true,"origin":"","legend":"\u003cp\u003ethe mean gray values of the images at different temperature under different illuminations. (a) the mean gray values of the original images before processing by the Retinex algorithm, (b) the mean gray values of the images after processing by the Retinex algorithm.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-4454734/v1/2920a1080e51ba1861841a85.png"},{"id":57873996,"identity":"30e0acea-e935-4ef9-b116-b5c70456496a","added_by":"auto","created_at":"2024-06-06 18:42:59","extension":"jpeg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":125667,"visible":true,"origin":"","legend":"\u003cp\u003eThe network structure of Retinex-CNN.\u003c/p\u003e","description":"","filename":"floatimage4.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-4454734/v1/7f1cfb76862b4dc2fbe68a1c.jpeg"},{"id":57872913,"identity":"68df35f1-728b-4d4b-8b2c-7d11a8b32e22","added_by":"auto","created_at":"2024-06-06 18:34:59","extension":"jpeg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":441787,"visible":true,"origin":"","legend":"\u003cp\u003eTemperature prediction results by CNN and Retinex-CNN under different illuminations. (a) and (b) correspond to data sets of 692lux, (c) and (d) correspond to 1039lux.\u003c/p\u003e","description":"","filename":"floatimage5.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-4454734/v1/33130beb18b9a4adb4e51685.jpeg"},{"id":57872914,"identity":"c6cdbd62-516b-476f-b92d-89df33e09229","added_by":"auto","created_at":"2024-06-06 18:34:59","extension":"jpeg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":268571,"visible":true,"origin":"","legend":"\u003cp\u003eTemperature detection for untrained visible light image, (a) by CNN and (b) by Retinex-CNN.\u003c/p\u003e","description":"","filename":"floatimage6.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-4454734/v1/006f028cea58c75f0fd98eb6.jpeg"},{"id":57875265,"identity":"e347c4e6-64e4-440a-a164-544dc421434f","added_by":"auto","created_at":"2024-06-06 18:59:07","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":6827905,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-4454734/v1/4e924afd-5d7c-4338-86f6-8c56beca5567.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Realization of normal temperature detection through visible light images by Retinex-CNN","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eThermal radiation of objects at normal temperature (30\u0026deg;C\u0026thinsp;~\u0026thinsp;150\u0026deg;C) is infrared electromagnetic waves (6\u0026micro;m\u0026thinsp;~\u0026thinsp;13\u0026micro;m), which are not detectable by the sensors of cameras developed for visible light imaging. Consequently, in non-contact normal temperature detection infrared imaging is the mainstream method, though the infrared cameras are expensive and have low pixels. The principle of thermal modulation of the reflected light indicates that changes in the temperature of an object will cause changes in the surface light reflectance [\u003cspan additionalcitationids=\"CR2 CR3\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. Therefore, in the presence of the incident of visible light, contactless normal temperature detection, in fact, can be achieved by using the surface reflects light of the target, i.e., visible light image of the target. The key technology of realizing this new detection method is how to extract the temperature information exactly from the visible light images of the target. In recent years, with the rise of machine learning techniques, researchers have tried to use the method of machine learning to extract the temperature information from the visible light images. Such as, a prediction model describing the connection between image features and temperature was proposed by M. Wang et al. [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]; W. Li et al. proposed a full-wavelength multi-incidence angle thermally modulated light reflection temperature measurement mechanism based on visible images [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]; W. Du reported the visible light thermometry in sunlight conditions using machine learning [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]; X. Nie et al. proposed a similar machine learning approach in sunlight environment while using videos from surveillance camera as the data for temperature information extraction [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. With human assistance to recognize the key features for classification, the research works stated above shows that temperature information can be extracted from the visible light images by using the machine learning. Deep learning is a branch of machine learning, which is a machine learning method based on neural networks [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Without human intervention deep learning can automatically learn features and patterns from raw data and use those features and patterns to classify or predict the data. Moreover, unlike traditional machine learning, deep learning is able to learn features with multiple layers of abstraction, allowing it to work with more complex and high-dimensional data [\u003cspan additionalcitationids=\"CR11\" citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. Therefore, exploring and then employing the method of deep learning to extract the temperature information from the visible optical image of the target will make the temperature detection based on visible optical image more convenient and accurate.\u003c/p\u003e \u003cp\u003eConvolutional neural network (CNN) is a deep learning model that stands out for its superior ability to extract features in localised regions. In image information extraction, CNN show significant advantages through convolutional operations, parameter sharing and multilayer structure [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. In practical applications, CNN are widely used in image classification [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e], object detection [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e], face recognition [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e], medical image analysis [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e], industrial defect detection [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e], etc., showing excellent performance and adaptability. This makes CNN a mainstream deep learning model for processing various vision tasks. Consequently, to realize the temperature extraction from visible light image, CNN would be a very suitable neural network model. One other thing to be noted is that the change of environmental illuminance is able to make a difference of image information of the object. While this disturbance brought about by changes in environmental illuminance will further bring great difficulties to the temperature detection based on the visible light image of the target. According to the principle of thermal modulation of the reflected light, the temperature change of the target changes only the reflectivity of the target surface actually [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. If the reflectance distribution map can be separated from the original visible light image, then the temperature detection by using the reflectance distribution map can avoid the influence brought by the change of ambient illumination, so that the temperature detection based on visible light image would more accurate. Therefore, in order to extract the temperature information from visible light images accurately, combining with Retinex algorithm which emphasizes the separation of reflection image and illumination image from original visible light image, a Retinex convolutional neural network (Retinex-CNN) is constructed in this paper. In the subsequent chapters of this paper, we will use Retinex-CNN for temperature recognition of visible image.\u003c/p\u003e"},{"header":"2 Experimental data acquisition","content":"\u003cp\u003eIn order to use the Retinex-CNN to detect the temperature of the target from its visible light image, it is necessary to collect the images corresponding to the target with different temperatures and apply them to train the neural network. Thus, the target which was a copper plate was attached on a temperature-controlled hotplate. Copper is an important metal material that is often used in power grid, so we take copper sheet as the target object for the temperature detection. The hotplate employed is a stainless steel electric hot plate of model type of DB-1, whose controllable temperature range and temperature control precision are from room temperature to 300\u0026deg;C and 0.1\u0026deg;C. For visible light imaging, a Canon EOS 200D II camera was used to capture video of the target, and a still image in 8-bit JPG format was extracted from the video. For the sake of employing Retinex-CNN to realize real-time monitoring of the target temperature, the video images are collected to train the Retinex-CNN. As the experimental device diagram (i.e., Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e) shown, camera is placed in front of the copper sheet, sun light is used as the lighting source. When the temperature of the copper sheet is heated up to 130℃, the hotplate is turned off and the video model of the camera is turned on. As the temperature of the copper sheet drops, visible light optical images of the copper sheet at different temperatures will be recorded by the camera. Moreover, a spectral color illuminance meter (SPIC-200) was used to monitor the color temperature and illuminance of the ambient light. As shown in inset of Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, the incident light illuminance has been changing throughout the experiment. A total of 46000 target images at different temperatures were collected as the temperature of the target dropped from 130\u0026deg;C to 40\u0026deg;C. The temperature interval was 0.1\u0026deg;C from 40 to 130\u0026deg;C, and the number of target image at each temperature is about 50. Moreover, a series of three experiments for the image capturing were carried out under different illuminance of the ambient light. These visible light images of the target at different temperatures would be used to train and test the Retinex-CNN.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"3 Theoretical analysis of the Retinex-CNN","content":"\u003cp\u003eIn optical imaging, the information actually recorded by the camera is the illuminance incident on the target multiplied by the reflectivity of the target's surface, according to the principle of optical imaging. Thus the target image \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(S(x,y)\\)\u003c/span\u003e\u003c/span\u003e can be expressed as\u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e\n$$S(x,y)=L(x,y) \\times R(x,y)$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\u003c/div\u003e.\u003c/p\u003e \u003cp\u003eWhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(R(x,y)\\)\u003c/span\u003e\u003c/span\u003e is the reflectivity of the target's surface and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(L(x,y)\\)\u003c/span\u003e\u003c/span\u003e is the light illuminance incident on the target. It is obvious that the target image changes with the incident light illuminance, even if the surface reflectivity does not change. In another word, the variation in illuminance will introduce interference to the temperature detection based on visible light imaging. Retinex is an image processing algorithm based on visual perception, and its theory emphasizes the separation of the two key components of the image reflection and illumination, so that the final image highlights the reflection component. Applying this algorithm to the target images for visible light temperature detection, the reflectivity distribution of the target surface, which is independent of the incident illuminance, can be obtained, so that the interference due to the variation of background illuminance can be eliminated. To validate this idea, target images corresponding to illuminances of 692lux, 800lux, and 1039lux were processed through the Retinex algorithm and the effectiveness of the algorithm is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. The frequency distribution of gray values of the images for the target temperature of 80\u0026deg;C before processing are illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e(a), (b) and (c). The curves for all three RGB channels differ greatly from each other, indicating that the variation of illuminance has a significant effect on the image of the target at the same temperature. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e(c), (d) and (e) are the frequency distribution curves of gray values of the images after processing, in which the curves for all three RGB channels remain stable and do not change with the illuminance. This proves that the interference of the incident light illuminance on the target information in the visible light image can indeed be eliminated using the Retinex algorithm.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe capability of the Retinex algorithm to exclude the interference of changes in incident illuminance on the image feature of the mean gray value was also examined. The mean gray values of the original images shot for different target temperatures and in different incident illuminances are plotted in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e(a), and the mean gray values of the corresponding images processed by the Retinex algorithm are plotted in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e(b). It is obvious that the mean gray value changes significantly with the illuminance, and the magnitude of the change even exceeds the magnitude of the change in the mean gray value due to the change in target temperature. After the images were processed by the Retinex algorithm, the interference of the incident illuminance on the mean gray value of the target image is greatly weakened, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e(b), and the effect of the change in target temperature on the mean gray value of the image is highlighted. Therefore, if the mean gray value is used as the basis for measuring the temperature of the target in an image, the Retinex algorithm is also effective in avoiding the interference caused by changes in incident illuminance. In summary, the introduction of the Retinex algorithm will effectively exclude the errors due to changes in incident illuminance when utilizing the visible light images for temperature detection.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe above discussions suggest that the Retinex algorithm is effective in eliminating the interference of the variation in incident illuminance on the temperature features in the visible light images of a target. Therefore, in the approach of temperature detection using visible light optical images and the convolutional neural network, the Retinex algorithm can be incorporated into the network to improve its accuracy in temperature detection.\u003c/p\u003e \u003cp\u003eBased on this, after combining Retinex algorithm an upgraded CNN which is named as Retinex-CNN is constructed in this paper, and its network structure is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. Firstly, the target images at different temperature are cropped into 150\u0026times;150 pixels and then being inputted it into the Retinex-CNN. The auto-MSRCR module of the Retinex algorithm is placed at the front of the Retinex-CNN, so before extracting the temperature information the auto-MSRCR module has separated the target image into reflection image and illumination image. Retaining the reflection image of the target, and then input it into the convolutional layers, which consists of 32, 64, and 128 convolutional kernels. Take advantage of a 3\u0026times;3 kernel size, the convolutional layers gradually extracts the feature information from the reflection image with a stride of 1. After each convolutional layer, ReLU activation is applied to enhance the network's non-linear fitting capabilities. Moreover, a 2\u0026times;2 max-pooling layers was added after each convolutional layer to reduce feature map dimensionality and extract significant features. After the features of the reflection image are extracted by the convolutional layer, the features are integrated by a six-layer fully connected layer. Following, the fully connected layer outputs a dedicated neuron for regression tasks, and the network generates continuous values reflecting prediction outcomes. By calculating the mean squared error (MSE) between the predicted results and the labels, the obtaining loss value is sent back to the Adam optimizer, and then updating the model parameters until the desired level of convergence is achieved. In the following chapters of this paper, the collected target images at different temperatures will be used to train the Retinex-CNN, and the trained Retinex-CNN will be used to detect the temperature from the target's visible light image.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"4 Temperature detection by Retinex-CNN","content":"\u003cp\u003eThree groups of visible light images of the target at different temperatures were collected and a corresponding ambient light illuminance for each group was measured as 692lux, 800lux and 1039lux, respectively. Note that, the image collection process lasted for a certain period of time, and the actual ambient illuminance varied within a certain range due to uncontrollable changes in natural conditions. The above illuminance values of 692lux, 800lux and 1039lux are representatives of the range of ambient illuminance change during three image collection processes, rather than the exact ambient illuminances of each of the images in three groups. In the following discussion, the three illuminance values of 692lux, 800lux and 1039lux would be used as names to distinguish the collected three groups of images. The images in the group of 692lux were chosen to form the dataset for the Retinex-CNN, which were divided into a training set, a validation set, and a testing set in the ratio of 70%, 20%, and 10%. For the purpose of comparison, the same dataset was also used for the training, validation and testing of the CNN. The results of temperature detection by the CNN and the Retinex-CNN using the visible light images in the group of 692lux illuminance are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e(a) and (b). The same procedure was also done for the group of 1093lux and the results are illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e(c) and (d). The mean absolute error (MAE) is used as a measure of the accuracy of the temperature detection and it is expressed as\u003cdiv id=\"Equ2\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ2\" name=\"EquationSource\"\u003e\n$$\\text{M}\\text{A}\\text{E}=\\left( {\\frac{1}{n}} \\right)\\sum\\nolimits_{i}^{n} {\\left| {{y_i} - {x_i}} \\right|}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e2\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003ewhere y is the real temperature of the target and x is the target temperature detected by the neural network from the visible light image. From Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e, it can be seen that the temperature detection of the Retinex-CNN is superior to that of the CNN. For 692lux illuminance, the Retinex-CNN reduces the mean absolute error of temperature detection from 2.6522\u0026deg;C to 2.0479\u0026deg;C. A more significant improvement in the accuracy of temperature detection can be observed for 1039 lux illuminance and the mean absolute error decreases from 2.9541\u0026deg;C to 1.9459\u0026deg;C. These results show that the introduction of the Retinex algorithm, which is incorporated into Retinex-CNN, can significantly improve the performance of temperature detection based on the CNN and the visible light images achieved under natural illumination conditions. This is because the Retinex algorithm can exclude the interference of ambient illuminance variation on the temperature feature in an image.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eTo further validate the illuminance generalization capability of the Retinex-CNN in target temperature detection, all the images in the groups of 692lux and 1039lux were merged together to form the training set and the validation set, and the images in the group of 800lux comprised the testing set used to test the performance of the trained Retinex-CNN model. These data sets were also used for the training, validation and testing of the CNN model for comparison. The results of temperature detection using these two neural network models are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e. The images in the group of 800lux were only used for testing and none of them was included in the training set. The huge difference in the background illuminance between the images in the training set and the images in the testing set leads to the deterioration of the accuracy of temperature detection and the mean absolute error reaches as high as 9.5146\u0026deg;C, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e(a). Comparing Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e(b) with (a), the Retinex-CNN model demonstrates significantly superior performance to the CNN model with an mean absolute error of 2.6238\u0026deg;C, which represents a 6.89\u0026deg;C detection error reduction. This indicates that the Retinex-CNN proposed in this paper has excellent illuminance generalization capability, and its use in temperature detection based on visible light images can effectively reduce the interference caused by changes in background illumination, thus improving the accuracy of temperature detection. This improvement of illuminance generalization ability would be a key technology for the application of the method of normal temperature detection which is based on visible image.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"5 Conclusion","content":"\u003cp\u003eA Retinex-CNN is proposed in this paper to improve the performance of normal temperature detection based on visible light images. The Retinex algorithm is able to extract an intensity distribution image of the incident light and an intensity distribution image of the reflected light from an image, and the image of reflected light is fed to the trained CNN for target temperature detection. In this way, the interference of ambient illumination variation on the temperature detection is minimized and the performance is therefore improved. The experimental results show that Retinex-CNN has superior detection accuracy compared to the CNN, especially when the image used for temperature detection differs significantly from the images used for neutral network training in terms of background illumination. A temperature detection error of less than 3\u0026deg;C was obtained by the Retinex-CNN, which is about 7\u0026deg;C detection error reduction comparing with the detection result of the CNN. It can be concluded that the Retinex-CNN proposed in this paper has excellent illuminance generalization capability for normal temperature detection based on visible light images, which is ideal for outdoor temperature detection applications where background illumination is difficult to control.\u003c/p\u003e"},{"header":"Declarations","content":" \u003ch2\u003eCompeting interests:\u003c/h2\u003e \u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eFunding:\u003c/h2\u003e \u003cp\u003eThis research was funded by National Natural Science Foundation of China (NSFC), grant numbers 61975072 and 12174173; Natural Science Foundation of Fujian province, grant numbers 2022H0023, 2022J02047 and 2022G02006; Natural Science Foundation of Zhangzhou City, grant number ZZ2023J20.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eAuthor contributions: JZ and HL wrote the main manuscript text, ZM is responsible for data collectionall, CH and FC are conducting network construction and development, JZ , CY and YL is responsible for editing and polishing the article, HC, YZ and JP is responsible for conceptualization and methodology, the authors reviewed the manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eZ. Zha, J. Zhu, X. Yang, X. Huang, H. Guo, A. Xie, X. Lu, Z. Fu, in 2022 23rd International Conference on Electronic Packaging Technology (ICEPT), pp.1\u0026ndash;5(2022)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eH. Yang, D. Wang, Q. Yin, Y. Gao, L. Zheng, in 2022 23rd International Conference on Electronic Packaging Technology (ICEPT), pp.1\u0026ndash;6(2022)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eL. Farbaniec, D.E. Eakin, Rev. Sci. Inst. \u003cb\u003e94\u003c/b\u003e, 034902 (2023)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eS. Alajlouni, K. Maize, A. Shakouri, in 2022 21st IEEE Intersociety Conference on Thermal and Thermomechanical Phenomena in Electronic Systems (iTherm), pp. 1\u0026ndash;10(2022)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eM. Wang, Research on temperature measurement method based on visible image and machine learning[D], Ph.D. dissertation, Dept. of Optical Engineering, Huazhong University of Science and Technology, Wuhan, China, 2020\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eW.M. Li, \u003cem\u003eResearch on the mechanism and method optimization of artificial intelligence temperature measurement based on visible light images[D]\u003c/em\u003e (Huazhong University of Science and Technology, 2021)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eW. Du, Q. Ye, Z. Yuan, C. Li, High. Volt Appar. \u003cb\u003e58\u003c/b\u003e, 221\u0026ndash;229 (2022). (in Chinese)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eX. Nie, Q. Ye, Z. Yuan, M. Han, in 16th Annual Conference of China Electrotechnical Society, pp. 616\u0026ndash;627(2022)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eY. LeCun, Y. Bengio, G. Hinton, nature, 521, 436\u0026ndash;444 (2015)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJ. Yang, P. Qiao, Y. Li, N. Wang, Stat. Decis. Mak. \u003cb\u003e35\u003c/b\u003e, 36\u0026ndash;40 (2019)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eK. Yu, L. Jia, Y. Chen, W. Xu, Comput. Res. Dev. \u003cb\u003e50\u003c/b\u003e, 1799\u0026ndash;1804 (2013)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eY. Zheng, G. Li, Y. Li, Comput. Eng. Appl. \u003cb\u003e55\u003c/b\u003e, 20\u0026ndash;36 (2019)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eS. Albawi, T.A. Mohammed, S. Al-Zawi, in 2017 International Conference on Engineering and Technology, pp. 1\u0026ndash;6(2017)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eX. Lei, H. Pan, X. Huang, IEEE Access. \u003cb\u003e7\u003c/b\u003e, 124087\u0026ndash;124095 (2019)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eS. Gidaris, N. Komodakis, in the IEEE international conference on computer vision, pp.1134\u0026ndash;1142(2015)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eS. Lawrence, C.L. Giles, A.C. Tsoi, A.D. Back, IEEE Trans. Neural Netw. \u003cb\u003e8\u003c/b\u003e, 98\u0026ndash;113 (1997)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eG. Litjens, T. Kooi, B.E. Bejnordi, A.A.A. Setio, F. Ciompi, M. Ghafoorian, J.A.W.M. Laak, B.V. Ginneken, C.I. Sάnches, Med. Image Anal. \u003cb\u003e42\u003c/b\u003e, 60\u0026ndash;88 (2017)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eX. Tao, D. Zhang, W. Ma, X. Liu, D. Xu, Appl. Sci. \u003cb\u003e8\u003c/b\u003e, 1575 (2018)\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":"non-contact normal temperature detection, visible light image, Retinex algorithm, convolutional neural network, generalization capability","lastPublishedDoi":"10.21203/rs.3.rs-4454734/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-4454734/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eContactless detection of a target\u0026rsquo;s temperature of normal range (30\u0026deg;C\u0026thinsp;~\u0026thinsp;150\u0026deg;C), which is based on its visible light images rather than infrared images, is a promising technology. Visible light imaging is primarily based on the visible light reflected from a target, rather than the thermal radiation of its own. The main challenge of contactless normal temperature detection through visible light images exists in the interference introduced by the variation of incident illumination. To solve this, a Retinex convolutional neural network (Retinex-CNN) was proposed in this paper, which was based on the convolutional neural network and Retinex algorithm. This network reduces the interference introduced by illumination variations and effectively improves the accuracy of the temperature detection based on visible light images. The temperature detection results of the Retinex-CNN shows that this network exhibits favorable generalization capability in terms of illumination variations, that is, it is still able to accurately detect temperature as a target is imaged in an illumination condition which is significantly different from that of the images in the training set. In this case, the Retinex-CNN obtains an average absolute error of 2.6\u0026deg;C, a value that is 6.89\u0026deg;C lower than that obtained by the CNN.\u003c/p\u003e","manuscriptTitle":"Realization of normal temperature detection through visible light images by Retinex-CNN","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-06-06 18:34:54","doi":"10.21203/rs.3.rs-4454734/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":"5cd47b63-173a-4c33-9ba5-e28aafef6403","owner":[],"postedDate":"June 6th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2024-06-06T18:34:57+00:00","versionOfRecord":[],"versionCreatedAt":"2024-06-06 18:34:54","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-4454734","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-4454734","identity":"rs-4454734","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}