Kyouiku Kanji Grade 1 Recognition Using MobileNet V2 Based on Android

Kyouiku Kanji Grade 1 Recognition Using MobileNet V2 Based on Android | 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 Kyouiku Kanji Grade 1 Recognition Using MobileNet V2 Based on Android Fathir Fathan, Alvi Syahrini Utami, Danny Matthew Saputra This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7473659/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 Character recognition has become a popular research topic in the field of pattern recognition and machine learning, including handwriting recognition, specifically kanji handwriting. This study performs handwriting recognition of kyouiku kanji grade 1, which is the kanji required to be learnt by grade 1 elementary school students in Japan. This research uses ETL-9B dataset from Electrotechnical Laboratory (now AIST), uses CNN MobileNet V2 deep learning method that has been customized for mobile devices, and uses Android application as the user interface implementation. Based on the study results, the highest accuracy model was obtained with an accuracy of 96,6875% and a size of 27.4MB for the alpha 1.0 hyperparameter. It can be concluded that the CNN MobileNet V2 deep learning method has performed quite well in the process of recognizing handwritten kyouiku kanji grade 1. Theoretical Computer Science MobileNet V2 ETL-9B Handwritten Kanji Recognition Kyouiku Kanji Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 1. Introduction Character recognition has become a popular research topic in the fields of pattern recognition and machine learning. With the development of digital technology, character recognition has become an important tool in many fields, including handwritten character recognition [ 1 ]. Character recognition involves complex phases such as preprocessing, segmentation, normalization, feature extraction, classification, and postprocessing. The task of character recognition also has its own difficulties, such as background complexity, uneven lighting, rotation, blur, degradation, aspect ratio, font, language, and others [ 1 ]. Recent advances in machine learning, especially deep learning, have shown promising results in the field of handwritten character recognition. Convolutional neural networks (CNN) are very effective in understanding the structure of handwritten characters by aiding in the automatic extraction of different features, making CNN a suitable approach for solving handwriting recognition problems [ 2 ]. Handwritten kanji recognition technology is widely used in pen-input interfaces, such as PDAs or mobile phone devices, and is expected to grow in popularity as its application scope expands in the future. However, its accuracy is still far from human capabilities [ 3 ]. Previous research on handwritten kanji usually focuses on the Kuzushiji-Kanji (ancient cursive kanji) dataset from Research Organization of Information and Systems - Center for Open Data in the Humanities (ROIS-CODH) due to the easy accessibility and format of the dataset, as well as its usefulness for recognizing classical documents. Regarding this dataset, there was a recognition study of 63 kanji for reading the Kiritsubo chapter of Genji Monogatari (a classical Japanese document) [ 4 ]. There is also research on recognizing 150 kanji that have the most samples in the dataset with the CNN method [ 5 ]. For the purposes of modern handwritten kanji recognition, the Electrotechnical Laboratory (ETL) dataset is used in this research. This dataset has a relatively large number of samples, but it is relatively more difficult to use because of its binary format. Regarding this dataset, there are several recognition studies for 878 kanji using CNN method, including [ 6 ] and [ 7 ]. However, it is difficult to find research specific to kyouiku kanji grade 1, 80 kanji that all 1st grade elementary school children in Japan must learn, which is also a subset of jouyou kanji (2136 kanji designated by the Japanese government for daily use). In this research, grade 1 kyouiku kanji recognition is carried out with an Android application as its user interface (UI) implementation, and using MobileNet V2 as CNN implementation that has been customized for mobile device usage [ 8 ]. MobileNetV2 is used in this research because, out of all the models available on Keras, this model has the smallest size and number of parameters, but still has a high accuracy value making it suitable for use on Android [ 9 ]. 2. Literature Study / Hypotheses Development a. Kanji Kanji are Japanese morphosyllabic characters derived from Chinese characters and are used in Japanese writing [ 10 ]. Japanese scholars attach Japanese meanings to Chinese characters, so that each Chinese character is considered to have, in addition to the Chinese sound, an additional Japanese sound that corresponds to its meaning in Japanese. These characters are now known as Japanese characters [ 11 ]. The Japanese government itself has published character lists periodically to help guide the education of its citizens through the myriad of characters that exist. Among them are 863 kanji that can be used in Japanese names [ 12 ], as well as 2136 kanji that can be used in general communication media [ 13 ]. b. Jouyou Kanji Jouyou kanji are kanji used to write modern Japanese in general daily social life, such as in laws and regulations, official documents, newspapers, magazines, and broadcasts. The currently used jouyou kanji was issued in 2010 and contains 2136 characters [ 14 ]. Students by the end of the sixth year in elementary school have learned 1006 of the 2136 kanji , and it is estimated that these 1006 kanji alone account for 95% of kanji usage in print media [ 15 ]. c. Kyouiku Kanji Grade 1 Kyouiku kanji is the common name for Gakunen Betsu Kanji Haitouhyou in Gakushuu Shidou Youryou (general guide to teaching and learning) published by the Japanese Ministry of Education, Culture, Sports, Science, and Technology, which is a subset of jouyou kanji intended to be learned by students in grades 1 to 6 during compulsory education in Japan [ 16 ]. Kyouiku kanji itself is divided into six subsets, from grade 1 (Dai-1 Gakunen) to grade 6 (Dai-6 Gakunen). Here is a list of 80 kanji included in kyouiku kanji grade 1 [ 17 ]: 一 右 雨 円 王 音 下 火花 貝 学 気 九 休 玉 金 空 月 犬 見 五 口 校 左 三 山 子 四 糸 字 耳 七 車 手 十 出 女 小 上森 人 水 正 生 青 夕 石 赤 千 川 先 早 草 足 村 大 男 竹 中 虫 町 天 田 土 二 日 入 年 白 八百 文 木 本 名 目 立 力 林 六 d. MobileNet V1 The idea behind the development of MobileNet V1, an open-source mobile-first Convolutional Neural Network (CNN) architecture developed by Google, is that the convolution layer, which is essential for computer vision but quite expensive to compute, can be replaced with a depthwise separable convolution layer [18]. Depthwise separable convolutions factorize the standard convolution into depthwise convolution and pointwise convolution, as illustrated by Fig. 1 . Standard convolution filters and combines inputs into one new output in one step. In MobileNet V1, the step is factorized into two steps. The depthwise convolution layer is used to apply one filter per input channel, and the pointwise convolution layer, a simple 1×1 convolution, is used to perform a linear combination of the depthwise convolution layer outputs. MobileNet V1 uses batch normalization and Rectified Linear Unit (ReLU) for both layers, as visualized in Fig. 2 . This factorization has the effect of drastically reducing computation and model size. MobileNet V1 requires between 8 and 9 times less computation than standard convolutions, with only a slight reduction in accuracy [18]. In MobileNet V1 and V2, to build a smaller and computationally cheaper model, a parameter α called the width multiplier can be used. The width multiplier α is used to adjust the number of channels in each layer. Given the number of input channels M and output channels N, the width multiplier α will set the number of channels to αM and αN. MobileNet V1 and V2 use α = 1 as the baseline and α < 1 as the reduced MobileNet. The width multiplier can reduce the computational cost and number of parameters quadratically around α 2 [18]. e. MobileNet V2 MobileNet V2 modifies the depthwise separable convolution into a bottleneck residual block by adding the utilization of linear bottleneck and inverted residual, as well as changing the usage of ReLU to ReLU6 [ 8 ]. In Fig. 3 , the first block from the left, which is low-dimensional (channel), is expanded into a high-dimensional block by pointwise convolution. Then the block will be filtered by depthwise convolution and projected back to the low-dimensional block by pointwise convolution. If the first and last blocks are the same, they are added together. The last block in Fig. 3 is called the linear bottleneck because it uses linear activation function in place of ReLu6. The flow of data from the low dimension being expanded to the high dimension, filtered in the high dimension, and projected back to the low dimension is called inverted residual. Figure 4 illustrates the comparison between residual block and inverted residual block. Table 1 lists the performance of MobileNet V1 and V2 on ImageNet by comparing the accuracy, number of parameters, and number of Multiply-Adds (MAdds) operations. It is evident that MobileNet V2 achieves higher results in those three metrics [ 8 ]. Table 1 MobileNet V1 and V2 Comparation Accuracy Parameter Multipy-Adds MobileNet V1 70.6 4.2 Million 575 Million MobileNet V2 72.0 3.4 Million 300 Million 3. Methodology a. Data Collection The data used in this research is secondary data, namely the ETL Character Database, which is a collection of images of approximately 1.2 million handwritten and machine-printed numbers, symbols, Latin letters, and Japanese characters, compiled into ETL-1 to ETL-9 datasets. In this research, the ETL-9B dataset was used. This database was collected by the Electrotechnical Laboratory for character recognition research from 1973 to 1984. Since January 2014, this database can be downloaded at etlcdb.db.aist.go.jp. b. Framework Figure 5 highlights the framework of this research. The research starts with the extraction of handwritten kanji data from the raw binary ETL-9B dataset into PNG format. The kanji used will also be limited to kyouiku kanji grade 1. Then the dataset will be divided into three different datasets, namely training, validation, and test datasets. The training and validation datasets will then be used to train the MobileNetV2 model implementation of the Keras API. For each epoch that has been trained, the model data and metrics results will be stored as a checkpoint, so that when the training process is complete, a list of models with their metrics will be obtained for each epoch. This research uses 200 epochs. From the list, one epoch will be selected whose accuracy is the highest. The model at that epoch will then be used as the output model. This output model can then be used to recognize the test dataset, and the test metrics results will be analyzed further. The model output will also be converted into the TensorFlow Lite model format so that it can be implemented into the Android application. After that, the Android application as a user interface will go through a testing and bug hunting process, and when it is considered that there are no more problems, then the final Android application can be deployed. 4. Result and Discussion Table 2 shows the results of validation accuracy, test accuracy, and model size for each different width multiplier alpha value. The Android application implementation result in Fig. 6 can be used as the UI for model inference. The research was conducted using a training dataset of 12800 images, a validation dataset of 1600 images, and a test dataset of 1600 images. The test dataset is used on the model at the epoch that has the highest validation accuracy value to get the test accuracy value. Each experiment uses the same settings and only differs in alpha value, with the aim of observing the relationship between accuracy and alpha value. Table 2 Model Accuracy and Sizes Alpha 0.35 0.5 0.75 1.0 1.3 1.4 Size 6.4 MB 9.7 MB 17.4 MB 27.4 MB 44.9 MB 51.8 MB Validation 92.0625% 95.3125% 93.1875% 97.1250% 70.6250% 47.5625% Test 90.5000% 94.3750% 90.8125% 96.6875% 72.6875% 42.6250% Difference 1.5625% 0.9375% 2.375% 0.4375% -2.0625% 4.9375% It can be observed in Table 2 that the higher the alpha value, the larger the model size. After alpha = 1.0, the resulting accuracy value drops significantly. Furthermore, it can be observed that the accuracy relationship between alpha values for validation and test results is the same, with accuracy at alpha 0.35 0.75 1.3 > 1.4. Finally, the model with the highest accuracy value is obtained at alpha = 1.0, with a test value of 96.6875%. 5. Conclusion Based on the explanation in the previous chapters, it can be concluded that: Handwritten kyouiku kanji grade 1 recognition using MobileNet V2 based on Android achieves the highest test accuracy value of 96.6875% in this research by using width multiplier alpha 1.0. Handwritten kyouiku kanji grade 1 recognition using MobileNet V2 based on Android has been implemented. References Hamad K, Kaya M (2016) A Detailed Analysis of Optical Character Recognition Technology, International Journal of Applied Mathematics Electronics and Computers, no. Special Issue-1, pp. 244– 249, December 10.18100/ijamec.270374 Ahlawat S, Choudhary A, Nayyar A, Singh S, Yoon B (2020) Improved Handwritten Digit Recognition Using Convolutional Neural Networks (CNN), Sensors, vol. 20, no. 12, p. 3344. 10.3390/s20123344 Ota I, Yamamoto R, Sako S, Sagayama S (2007) Online Handwritten Kanji Recognition Based on Inter- stroke Grammar, Ninth International Conference on Document Analysis and Recognition (ICDAR 2007), Curitiba, Brazil, pp. 1188–1192. 10.1109/ICDAR.2007.4377103 Hu X, Inamoto M, Konagaya A (2019) Recognition of Kuzushi-ji with Deep Learning Method, in Proc. Annu. Conf. JSAI, vol. JSAI2019, 33rd, Session ID 4H2-E-5-01, p. 4H2E501, Jun. 10.11517/pjsai.JSAI2019.0_4H2E501 Solis AI, Zarkovacki J, Ly J, Atyabi A (2023) Recognition of Handwritten Japanese Characters Using Ensemble of Convolutional Neural Networks. arXiv e-prints. 10.48550/arXiv.2306.03954 Tsai C (2016) Recognizing Handwritten Japanese Characters Using Deep Convolutional Neural Networks, University Stanford, Stanford, CA, USA, pp. 405–410 Pérez JVT (2020) Recognition of Japanese handwritten characters with machine learning techniques, B.S. thesis, Multimedia Engineering, Universidad de Alicante Sandler M, Howard AG, Zhu M, Zhmoginov A, Chen L-C, MobileNetV2: Inverted Residuals and Linear Bottlenecks, in 2018 IEEE/CVF Conference on Computer Vision and Pattern, Recognition (2018) pp. 4510–4520 Keras Applications Keras (2023) https://keras.io/api/applications/ Suski PM (1931) The Phonetics of Japanese Language. Routledge Library Edition, vol 59. Routledge, London Jinmeiyou, Kanjihyou Ministry of Justice, Japan, 2017. Accessed: Nov. 22, 2023. [Online]. Available: https://www.moj.go.jp/content/001131003.pdf Jouyou, Kanjihyou (2010) Agency for Cultural Affairs, Japan, Accessed: Nov. 22, 2023. [Online]. Available: https://www.bunka.go.jp/kokugo_nihongo/sisaku/joho/joho/kijun/naikaku/pdf/joyokanjihyo_2 0101130.pdf Jouyou Kanjihyou Japan Electronic Publishing Association (2023) https://www.jepa.or.jp/ebookpedia/201701_3385/ Paxton S, Svetanant C (2014) Tackling the Kanji hurdle: investigation of Kanji learning in non-Kanji background learners. Int J Res Stud Lang Learn 3(3):89–104 Kyouiku Kanji, NTT DOCOMO (2024) https://dictionary.goo.ne.jp/word/教育漢字/ (accessed Jun. 4 Shougakkou Gakushuu Shidou Youryou, Ministry of Education Culture Sports Science and Technology, Japan (2017) Accessed: Nov. 22, 2023. [Online]. Available: https://www.bunka.go.jp/kokugo_nihongo/sisaku/joho/joho/kijun/naikaku/pdf/joyokanjihyo_20101130.p df Howard AG et al (2017) ‘MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications’, arXiv [cs.CV] Additional Declarations The authors declare no competing interests. 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-7473659","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":506493637,"identity":"73d7257c-e00a-4d89-8003-84e51e3841e8","order_by":0,"name":"Fathir Fathan","email":"","orcid":"","institution":"Sriwijaya University","correspondingAuthor":false,"prefix":"","firstName":"Fathir","middleName":"","lastName":"Fathan","suffix":""},{"id":506493638,"identity":"9a9af13d-a696-4737-82ed-1b662a3f4cc3","order_by":1,"name":"Alvi Syahrini Utami","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Alvi","middleName":"Syahrini","lastName":"Utami","suffix":""},{"id":506493639,"identity":"ab0d4f35-fe46-4588-b619-3495cf4bfecb","order_by":2,"name":"Danny Matthew Saputra","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA9klEQVRIiWNgGAWjYDCCAyi8CiBmZm4gVgszEJ8B0YykaGFsAzEIaOE7fvbYYx6GOnlzifxjH37Oq43mbwdq+VGxDacWyTN56cY8DIcNd85IZp7Zu+147ozDjA2MPWdu49RicCDHTJqH4QDjhhvJzAy8247lNgC1MDO24dFy/g1IS509SAvj3znHcucT1HIDbAtzIkgLM29DTe4GQlokb7wxk5xjcDh5Z89jY2aZYwdyNwK1HMTnF77zOWYSbyrqbLezJz5mfFNTlzvv/OGDD35U4NYCAkw8BkAXQtiHweQBvOqBgPEHA1xLHSHFo2AUjIJRMAIBAELPWi2nBW0CAAAAAElFTkSuQmCC","orcid":"https://orcid.org/0009-0000-1144-7092","institution":"Sriwijaya University","correspondingAuthor":true,"prefix":"","firstName":"Danny","middleName":"Matthew","lastName":"Saputra","suffix":""}],"badges":[],"createdAt":"2025-08-27 16:47:54","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-7473659/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7473659/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":90102480,"identity":"b495bc8b-056b-4c1a-9ed6-e5bbe3cf5321","added_by":"auto","created_at":"2025-08-28 13:36:13","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":15301,"visible":true,"origin":"","legend":"\u003cp\u003eStandard and depthwise separable convolution\u003c/p\u003e","description":"","filename":"1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7473659/v1/a0ccd3f54c1c01a02cd9f997.jpg"},{"id":90102481,"identity":"23826f76-80ad-4d63-a5ce-6f3b5dbfca1f","added_by":"auto","created_at":"2025-08-28 13:36:13","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":11435,"visible":true,"origin":"","legend":"\u003cp\u003eBatch Normalization and ReLU after each Convolution Layer\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-7473659/v1/fae7c886080b282d7d5bde58.png"},{"id":90102482,"identity":"3a14af95-6364-46f7-80ca-e77f43d7ef9d","added_by":"auto","created_at":"2025-08-28 13:36:13","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":18282,"visible":true,"origin":"","legend":"\u003cp\u003eBottleneck Residual Block\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-7473659/v1/a0edc643159f36473b44c6f0.png"},{"id":90102483,"identity":"f6dc05d1-abbb-4e2c-b2cc-535ec08aae4b","added_by":"auto","created_at":"2025-08-28 13:36:13","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":13552,"visible":true,"origin":"","legend":"\u003cp\u003eResidual (left) and Inverted Residual (right) Block\u003c/p\u003e","description":"","filename":"4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7473659/v1/43cc54e85ed6c8e44377a8f4.jpg"},{"id":90102496,"identity":"d04b3050-872c-45ad-825b-5cfe87d25590","added_by":"auto","created_at":"2025-08-28 13:36:14","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":22287,"visible":true,"origin":"","legend":"\u003cp\u003eResearch Framework\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-7473659/v1/1739e4edc0f9e0c726a299e1.png"},{"id":90104226,"identity":"c6c47067-7f95-45e9-a97c-e307747c96c0","added_by":"auto","created_at":"2025-08-28 13:52:13","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":11826,"visible":true,"origin":"","legend":"\u003cp\u003eAndroid App UI for Model Inference\u003c/p\u003e","description":"","filename":"6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7473659/v1/329f4069ec64b7d1219d5eb4.jpg"},{"id":90105262,"identity":"620e6be4-4a28-4b76-a7fc-35f233fc37dd","added_by":"auto","created_at":"2025-08-28 14:08:14","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":504270,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7473659/v1/d06b06ba-bf91-4c9a-9973-9aa8d93d3937.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003eKyouiku Kanji Grade 1 Recognition Using MobileNet V2 Based on Android\u003c/p\u003e","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eCharacter recognition has become a popular research topic in the fields of pattern recognition and machine learning. With the development of digital technology, character recognition has become an important tool in many fields, including handwritten character recognition [\u003cspan class=\"CitationRef\"\u003e1\u003c/span\u003e].\u003c/p\u003e\n\u003cp\u003eCharacter recognition involves complex phases such as preprocessing, segmentation, normalization, feature extraction, classification, and postprocessing. The task of character recognition also has its own difficulties, such as background complexity, uneven lighting, rotation, blur, degradation, aspect ratio, font, language, and others [\u003cspan class=\"CitationRef\"\u003e1\u003c/span\u003e].\u003c/p\u003e\n\u003cp\u003eRecent advances in machine learning, especially deep learning, have shown promising results in the field of handwritten character recognition. Convolutional neural networks (CNN) are very effective in understanding the structure of handwritten characters by aiding in the automatic extraction of different features, making CNN a suitable approach for solving handwriting recognition problems [\u003cspan class=\"CitationRef\"\u003e2\u003c/span\u003e].\u003c/p\u003e\n\u003cp\u003eHandwritten kanji recognition technology is widely used in pen-input interfaces, such as PDAs or mobile phone devices, and is expected to grow in popularity as its application scope expands in the future. However, its accuracy is still far from human capabilities [\u003cspan class=\"CitationRef\"\u003e3\u003c/span\u003e].\u003c/p\u003e\n\u003cp\u003ePrevious research on handwritten \u003cem\u003ekanji\u003c/em\u003e usually focuses on the Kuzushiji-Kanji (ancient cursive kanji) dataset from Research Organization of Information and Systems - Center for Open Data in the Humanities (ROIS-CODH) due to the easy accessibility and format of the dataset, as well as its usefulness for recognizing classical documents. Regarding this dataset, there was a recognition study of 63 \u003cem\u003ekanji\u003c/em\u003e for reading the Kiritsubo chapter of Genji Monogatari (a classical Japanese document) [\u003cspan class=\"CitationRef\"\u003e4\u003c/span\u003e]. There is also research on recognizing 150 \u003cem\u003ekanji\u003c/em\u003e that have the most samples in the dataset with the CNN method [\u003cspan class=\"CitationRef\"\u003e5\u003c/span\u003e].\u003c/p\u003e\n\u003cp\u003eFor the purposes of modern handwritten \u003cem\u003ekanji\u003c/em\u003e recognition, the Electrotechnical Laboratory (ETL) dataset is used in this research. This dataset has a relatively large number of samples, but it is relatively more difficult to use because of its binary format. Regarding this dataset, there are several recognition studies for 878 kanji using CNN method, including [\u003cspan class=\"CitationRef\"\u003e6\u003c/span\u003e] and [\u003cspan class=\"CitationRef\"\u003e7\u003c/span\u003e]. However, it is difficult to find research specific to \u003cem\u003ekyouiku kanji\u003c/em\u003e grade 1, 80 \u003cem\u003ekanji\u003c/em\u003e that all 1st grade elementary school children in Japan must learn, which is also a subset of \u003cem\u003ejouyou kanji\u003c/em\u003e (2136 \u003cem\u003ekanji\u003c/em\u003e designated by the Japanese government for daily use).\u003c/p\u003e\n\u003cp\u003eIn this research, grade 1 \u003cem\u003ekyouiku kanji\u003c/em\u003e recognition is carried out with an Android application as its user interface (UI) implementation, and using MobileNet V2 as CNN implementation that has been customized for mobile device usage [\u003cspan class=\"CitationRef\"\u003e8\u003c/span\u003e]. MobileNetV2 is used in this research because, out of all the models available on Keras, this model has the smallest size and number of parameters, but still has a high accuracy value making it suitable for use on Android [\u003cspan class=\"CitationRef\"\u003e9\u003c/span\u003e].\u003c/p\u003e"},{"header":"2.\tLiterature Study / Hypotheses Development","content":"\u003cp style=\"display: inline !important;\"\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ea. Kanji\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eKanji\u003c/em\u003e are Japanese morphosyllabic characters derived from Chinese characters and are used in Japanese writing [\u003cspan class=\"CitationRef\"\u003e10\u003c/span\u003e]. Japanese scholars attach Japanese meanings to Chinese characters, so that each Chinese character is considered to have, in addition to the Chinese sound, an additional Japanese sound that corresponds to its meaning in Japanese. These characters are now known as Japanese characters [\u003cspan class=\"CitationRef\"\u003e11\u003c/span\u003e].\u003c/p\u003e\n\u003cp\u003eThe Japanese government itself has published character lists periodically to help guide the education of its citizens through the myriad of characters that exist. Among them are 863 \u003cem\u003ekanji\u003c/em\u003e that can be used in Japanese names [\u003cspan class=\"CitationRef\"\u003e12\u003c/span\u003e], as well as 2136 \u003cem\u003ekanji\u003c/em\u003e that can be used in general communication media [\u003cspan class=\"CitationRef\"\u003e13\u003c/span\u003e].\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eb. Jouyou Kanji\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eJouyou kanji are \u003cem\u003ekanji\u003c/em\u003e used to write modern Japanese in general daily social life, such as in laws and regulations, official documents, newspapers, magazines, and broadcasts. The currently used \u003cem\u003ejouyou kanji\u003c/em\u003e was issued in 2010 and contains 2136 characters [\u003cspan class=\"CitationRef\"\u003e14\u003c/span\u003e]. Students by the end of the sixth year in elementary school have learned 1006 of the 2136 \u003cem\u003ekanji\u003c/em\u003e, and it is estimated that these 1006 \u003cem\u003ekanji\u003c/em\u003e alone account for 95% of \u003cem\u003ekanji\u003c/em\u003e usage in print media [\u003cspan class=\"CitationRef\"\u003e15\u003c/span\u003e].\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ec. Kyouiku Kanji\u003c/strong\u003e \u003cstrong\u003eGrade 1\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eKyouiku kanji\u003c/em\u003e is the common name for Gakunen Betsu Kanji Haitouhyou in Gakushuu Shidou Youryou (general guide to teaching and learning) published by the Japanese Ministry of Education, Culture, Sports, Science, and Technology, which is a subset of \u003cem\u003ejouyou kanji\u003c/em\u003e intended to be learned by students in grades 1 to 6 during compulsory education in Japan [\u003cspan class=\"CitationRef\"\u003e16\u003c/span\u003e].\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eKyouiku kanji\u003c/em\u003e itself is divided into six subsets, from grade 1 (Dai-1 Gakunen) to grade 6 (Dai-6 Gakunen). Here is a list of 80 \u003cem\u003ekanji\u003c/em\u003e included in \u003cem\u003ekyouiku kanji\u003c/em\u003e grade 1 [\u003cspan class=\"CitationRef\"\u003e17\u003c/span\u003e]: 一 右 雨 円 王 音 下 火花 貝 学 気 九 休 玉 金 空 月 犬 見 五 口 校 左 三 山 子 四 糸 字 耳 七 車 手 十 出 女 小 上森 人 水 正 生 青 夕 石 赤 千 川 先 早 草 足 村 大 男 竹 中 虫 町 天 田 土 二 日 入 年 白 八百 文 木 本 名 目 立 力 林 六\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ed. MobileNet V1\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe idea behind the development of MobileNet V1, an open-source mobile-first Convolutional Neural Network (CNN) architecture developed by Google, is that the convolution layer, which is essential for computer vision but quite expensive to compute, can be replaced with a depthwise separable convolution layer [18]. Depthwise separable convolutions factorize the standard convolution into depthwise convolution and pointwise convolution, as illustrated by Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e\n\u003cp\u003eStandard convolution filters and combines inputs into one new output in one step. In MobileNet V1, the step is factorized into two steps. The depthwise convolution layer is used to apply one filter per input channel, and the pointwise convolution layer, a simple 1\u0026times;1 convolution, is used to perform a linear combination of the depthwise convolution layer outputs. MobileNet V1 uses batch normalization and Rectified Linear Unit (ReLU) for both layers, as visualized in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e. This factorization has the effect of drastically reducing computation and model size. MobileNet V1 requires between 8 and 9 times less computation than standard convolutions, with only a slight reduction in accuracy [18].\u003c/p\u003e\n\u003cp\u003eIn MobileNet V1 and V2, to build a smaller and computationally cheaper model, a parameter \u0026alpha; called the width multiplier can be used. The width multiplier \u0026alpha; is used to adjust the number of channels in each layer. Given the number of input channels M and output channels N, the width multiplier \u0026alpha; will set the number of channels to \u0026alpha;M and \u0026alpha;N. MobileNet V1 and V2 use \u0026alpha;\u0026thinsp;=\u0026thinsp;1 as the baseline and \u0026alpha;\u0026thinsp;\u0026lt;\u0026thinsp;1 as the reduced MobileNet. The width multiplier can reduce the computational cost and number of parameters quadratically around \u0026alpha;\u003csup\u003e2\u003c/sup\u003e [18].\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ee. MobileNet V2\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eMobileNet V2 modifies the depthwise separable convolution into a bottleneck residual block by adding the utilization of linear bottleneck and inverted residual, as well as changing the usage of ReLU to ReLU6 [\u003cspan class=\"CitationRef\"\u003e8\u003c/span\u003e]. In Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e, the first block from the left, which is low-dimensional (channel), is expanded into a high-dimensional block by pointwise convolution. Then the block will be filtered by depthwise convolution and projected back to the low-dimensional block by pointwise convolution. If the first and last blocks are the same, they are added together.\u003c/p\u003e\n\u003cp\u003eThe last block in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e is called the linear bottleneck because it uses linear activation function in place of ReLu6. The flow of data from the low dimension being expanded to the high dimension, filtered in the high dimension, and projected back to the low dimension is called inverted residual. Figure\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e illustrates the comparison between residual block and inverted residual block.\u003c/p\u003e\n\u003cp\u003eTable\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e lists the performance of MobileNet V1 and V2 on ImageNet by comparing the accuracy, number of parameters, and number of Multiply-Adds (MAdds) operations. It is evident that MobileNet V2 achieves higher results in those three metrics [\u003cspan class=\"CitationRef\"\u003e8\u003c/span\u003e].\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\n\u003ctable id=\"Tab1\" style=\"width: 297px;\" border=\"1\"\u003e\u003ccaption\u003e\n\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\n\u003cdiv class=\"CaptionContent\"\u003e\n\u003cp\u003eMobileNet V1 and V2 Comparation\u003c/p\u003e\n\u003c/div\u003e\n\u003c/caption\u003e\n\u003cthead\u003e\n\u003ctr style=\"height: 35px;\"\u003e\n\u003cth style=\"width: 72px; height: 35px;\" align=\"left\"\u003e\u0026nbsp;\u003c/th\u003e\n\u003cth style=\"width: 55.1111px; height: 35px;\" align=\"left\"\u003e\n\u003cp\u003eAccuracy\u003c/p\u003e\n\u003c/th\u003e\n\u003cth style=\"width: 63.8889px; height: 35px;\" align=\"left\"\u003e\n\u003cp\u003eParameter\u003c/p\u003e\n\u003c/th\u003e\n\u003cth style=\"width: 81px; height: 35px;\" align=\"left\"\u003e\n\u003cp\u003eMultipy-Adds\u003c/p\u003e\n\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr style=\"height: 48px;\"\u003e\n\u003ctd style=\"width: 72px; height: 48px;\" align=\"left\"\u003e\n\u003cp\u003eMobileNet V1\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd style=\"width: 55.1111px; height: 48px;\" align=\"char\" char=\".\"\u003e\n\u003cp\u003e70.6\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd style=\"width: 63.8889px; height: 48px;\" align=\"left\"\u003e\n\u003cp\u003e4.2 Million\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd style=\"width: 81px; height: 48px;\" align=\"left\"\u003e\n\u003cp\u003e575 Million\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr style=\"height: 48px;\"\u003e\n\u003ctd style=\"width: 72px; height: 48px;\" align=\"left\"\u003e\n\u003cp\u003eMobileNet V2\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd style=\"width: 55.1111px; height: 48px;\" align=\"char\" char=\".\"\u003e\n\u003cp\u003e72.0\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd style=\"width: 63.8889px; height: 48px;\" align=\"left\"\u003e\n\u003cp\u003e3.4 Million\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd style=\"width: 81px; height: 48px;\" align=\"left\"\u003e\n\u003cp\u003e300 Million\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e"},{"header":"3.\tMethodology","content":"\u003cp\u003ea. \u003cstrong\u003eData Collection\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe data used in this research is secondary data, namely the ETL Character Database, which is a collection of images of approximately 1.2\u0026nbsp;million handwritten and machine-printed numbers, symbols, Latin letters, and Japanese characters, compiled into ETL-1 to ETL-9 datasets. In this research, the ETL-9B dataset was used. This database was collected by the Electrotechnical Laboratory for character recognition research from 1973 to 1984. Since January 2014, this database can be downloaded at etlcdb.db.aist.go.jp.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eb. Framework\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFigure\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e highlights the framework of this research. The research starts with the extraction of handwritten \u003cem\u003ekanji\u003c/em\u003e data from the raw binary ETL-9B dataset into PNG format. The \u003cem\u003ekanji\u003c/em\u003e used will also be limited to \u003cem\u003ekyouiku kanji\u003c/em\u003e grade 1. Then the dataset will be divided into three different datasets, namely training, validation, and test datasets. The training and validation datasets will then be used to train the MobileNetV2 model implementation of the Keras API.\u003c/p\u003e\n\u003cp\u003eFor each epoch that has been trained, the model data and metrics results will be stored as a checkpoint, so that when the training process is complete, a list of models with their metrics will be obtained for each epoch. This research uses 200 epochs. From the list, one epoch will be selected whose accuracy is the highest. The model at that epoch will then be used as the output model. This output model can then be used to recognize the test dataset, and the test metrics results will be analyzed further.\u003c/p\u003e\n\u003cp\u003eThe model output will also be converted into the TensorFlow Lite model format so that it can be implemented into the Android application. After that, the Android application as a user interface will go through a testing and bug hunting process, and when it is considered that there are no more problems, then the final Android application can be deployed.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e"},{"header":"4. Result and Discussion","content":"\u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e shows the results of validation accuracy, test accuracy, and model size for each different width multiplier alpha value. The Android application implementation result in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e can be used as the UI for model inference. The research was conducted using a training dataset of 12800 images, a validation dataset of 1600 images, and a test dataset of 1600 images. The test dataset is used on the model at the epoch that has the highest validation accuracy value to get the test accuracy value. Each experiment uses the same settings and only differs in alpha value, with the aim of observing the relationship between accuracy and alpha value.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eModel Accuracy and Sizes\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"7\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colspan=\"6\" nameend=\"c7\" namest=\"c2\"\u003e\u003cp\u003eAlpha\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.35\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.5\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.75\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003e1.0\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003e1.3\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003e1.4\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSize\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e6.4 MB\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e9.7 MB\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e17.4 MB\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e27.4 MB\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e44.9 MB\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e51.8 MB\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eValidation\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e92.0625%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e95.3125%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e93.1875%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e97.1250%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e70.6250%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e47.5625%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTest\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e90.5000%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e94.3750%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e90.8125%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e96.6875%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e72.6875%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e42.6250%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eDifference\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e1.5625%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.9375%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e2.375%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.4375%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e-2.0625%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e4.9375%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eIt can be observed in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e that the higher the alpha value, the larger the model size. After alpha\u0026thinsp;=\u0026thinsp;1.0, the resulting accuracy value drops significantly. Furthermore, it can be observed that the accuracy relationship between alpha values for validation and test results is the same, with accuracy at alpha 0.35\u0026thinsp;\u0026lt;\u0026thinsp;0.5\u0026thinsp;\u0026gt;\u0026thinsp;0.75\u0026thinsp;\u0026lt;\u0026thinsp;1.0\u0026thinsp;\u0026gt;\u0026thinsp;1.3\u0026thinsp;\u0026gt;\u0026thinsp;1.4. Finally, the model with the highest accuracy value is obtained at alpha\u0026thinsp;=\u0026thinsp;1.0, with a test value of 96.6875%.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e"},{"header":"5. Conclusion","content":"\u003cp\u003eBased on the explanation in the previous chapters, it can be concluded that:\u003c/p\u003e\u003cp\u003e\u003col\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003eHandwritten \u003cem\u003ekyouiku kanji\u003c/em\u003e grade 1 recognition using MobileNet V2 based on Android achieves the highest test accuracy value of 96.6875% in this research by using width multiplier alpha 1.0.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003eHandwritten \u003cem\u003ekyouiku kanji\u003c/em\u003e grade 1 recognition using MobileNet V2 based on Android has been implemented.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003c/ol\u003e\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eHamad K, Kaya M (2016) A Detailed Analysis of Optical Character Recognition Technology, International Journal of Applied Mathematics Electronics and Computers, no. Special Issue-1, pp. 244\u0026ndash; 249, December \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.18100/ijamec.270374\u003c/span\u003e\u003cspan address=\"10.18100/ijamec.270374\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAhlawat S, Choudhary A, Nayyar A, Singh S, Yoon B (2020) Improved Handwritten Digit Recognition Using Convolutional Neural Networks (CNN), Sensors, vol. 20, no. 12, p. 3344. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/s20123344\u003c/span\u003e\u003cspan address=\"10.3390/s20123344\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eOta I, Yamamoto R, Sako S, Sagayama S (2007) Online Handwritten Kanji Recognition Based on Inter- stroke Grammar, Ninth International Conference on Document Analysis and Recognition (ICDAR 2007), Curitiba, Brazil, pp. 1188\u0026ndash;1192. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/ICDAR.2007.4377103\u003c/span\u003e\u003cspan address=\"10.1109/ICDAR.2007.4377103\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHu X, Inamoto M, Konagaya A (2019) Recognition of Kuzushi-ji with Deep Learning Method, in Proc. Annu. Conf. JSAI, vol. JSAI2019, 33rd, Session ID 4H2-E-5-01, p. 4H2E501, Jun. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.11517/pjsai.JSAI2019.0_4H2E501\u003c/span\u003e\u003cspan address=\"10.11517/pjsai.JSAI2019.0_4H2E501\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSolis AI, Zarkovacki J, Ly J, Atyabi A (2023) Recognition of Handwritten Japanese Characters Using Ensemble of Convolutional Neural Networks. arXiv e-prints. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.48550/arXiv.2306.03954\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2306.03954\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTsai C (2016) Recognizing Handwritten Japanese Characters Using Deep Convolutional Neural Networks, University Stanford, Stanford, CA, USA, pp. 405\u0026ndash;410\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eP\u0026eacute;rez JVT (2020) Recognition of Japanese handwritten characters with machine learning techniques, B.S. thesis, Multimedia Engineering, Universidad de Alicante\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSandler M, Howard AG, Zhu M, Zhmoginov A, Chen L-C, MobileNetV2: Inverted Residuals and Linear Bottlenecks, in 2018 IEEE/CVF Conference on Computer Vision and Pattern, Recognition (2018) pp. 4510\u0026ndash;4520\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKeras Applications Keras (2023) \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://keras.io/api/applications/\u003c/span\u003e\u003cspan address=\"https://keras.io/api/applications/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSuski PM (1931) The Phonetics of Japanese Language. Routledge Library Edition, vol 59. Routledge, London\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJinmeiyou, Kanjihyou Ministry of Justice, Japan, 2017. Accessed: Nov. 22, 2023. [Online]. Available:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.moj.go.jp/content/001131003.pdf\u003c/span\u003e\u003cspan address=\"https://www.moj.go.jp/content/001131003.pdf\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJouyou, Kanjihyou (2010) Agency for Cultural Affairs, Japan, Accessed: Nov. 22, 2023. [Online]. Available:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.bunka.go.jp/kokugo_nihongo/sisaku/joho/joho/kijun/naikaku/pdf/joyokanjihyo_2 0101130.pdf\u003c/span\u003e\u003cspan address=\"https://www.bunka.go.jp/kokugo_nihongo/sisaku/joho/joho/kijun/naikaku/pdf/joyokanjihyo_2 0101130.pdf\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJouyou Kanjihyou Japan Electronic Publishing Association (2023) \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.jepa.or.jp/ebookpedia/201701_3385/\u003c/span\u003e\u003cspan address=\"https://www.jepa.or.jp/ebookpedia/201701_3385/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePaxton S, Svetanant C (2014) Tackling the Kanji hurdle: investigation of Kanji learning in non-Kanji background learners. Int J Res Stud Lang Learn 3(3):89\u0026ndash;104\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKyouiku Kanji, NTT DOCOMO (2024) \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://dictionary.goo.ne.jp/word/教育漢字/\u003c/span\u003e\u003cspan address=\"https://dictionary.goo.ne.jp/word/教育漢字/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (accessed Jun. 4\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eShougakkou Gakushuu Shidou Youryou, Ministry of Education Culture Sports Science and Technology, Japan (2017) Accessed: Nov. 22, 2023. [Online]. Available: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.bunka.go.jp/kokugo_nihongo/sisaku/joho/joho/kijun/naikaku/pdf/joyokanjihyo_20101130.p df\u003c/span\u003e\u003cspan address=\"https://www.bunka.go.jp/kokugo_nihongo/sisaku/joho/joho/kijun/naikaku/pdf/joyokanjihyo_20101130.p df\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHoward AG et al (2017) \u0026lsquo;MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications\u0026rsquo;, arXiv [cs.CV]\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"Sriwijaya University","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":"MobileNet V2, ETL-9B, Handwritten Kanji Recognition, Kyouiku Kanji","lastPublishedDoi":"10.21203/rs.3.rs-7473659/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7473659/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eCharacter recognition has become a popular research topic in the field of pattern recognition and machine learning, including handwriting recognition, specifically \u003cem\u003ekanji \u003c/em\u003ehandwriting. This study performs handwriting recognition of \u003cem\u003ekyouiku kanji \u003c/em\u003egrade 1, which is the \u003cem\u003ekanji \u003c/em\u003erequired to be learnt by grade 1 elementary school students in Japan. This research uses ETL-9B dataset from Electrotechnical Laboratory (now AIST), uses CNN MobileNet V2 deep learning method that has been customized for mobile devices, and uses Android application as the user interface implementation. Based on the study results, the highest accuracy model was obtained with an accuracy of 96,6875% and a size of 27.4MB for the alpha 1.0 hyperparameter. It can be concluded that the CNN MobileNet V2 deep learning method has performed quite well in the process of recognizing handwritten \u003cem\u003ekyouiku kanji \u003c/em\u003egrade 1.\u003c/p\u003e","manuscriptTitle":"Kyouiku Kanji Grade 1 Recognition Using MobileNet V2 Based on Android","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-08-28 13:36:09","doi":"10.21203/rs.3.rs-7473659/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":"9c4a74c9-725e-46f6-a744-c3c3e791355f","owner":[],"postedDate":"August 28th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":53800601,"name":"Theoretical Computer Science"}],"tags":[],"updatedAt":"2025-08-28T13:36:09+00:00","versionOfRecord":[],"versionCreatedAt":"2025-08-28 13:36:09","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7473659","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7473659","identity":"rs-7473659","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}