臺灣博碩士論文加值系統

英文字元辨識已經有相當多的辨識方法，也有相當多的書籍提到，但字型一直都是一個影響辨識率的主因，辨識流程從文件的配置分析、文件的前處理，文字的分割，文字辨識及後處理結束，這中間每一個環節都可以影響辨識的結果，但除了文字辨識，字型幾乎不會影響其他的階段，而一般常見的英文字型就有50多種，現行的OCR (Optical Character Recognition) 也未必對每種字型都能適應。如同很多經典的二值化方法也都不能面對所有的情境一般，都能把每張圖做到前後景分離，讓背景被清除，僅留下我們要的前景。
人類是如何學習英文字怎麼書寫的呢?基本是從筆劃 (Stroke) 開始的，那筆劃又是什麼呢?它是一些不同方向的線條組成的，因此這個不同方向的線條即為一個重要的特徵。傳統的人工智慧網路，會直接利用前處理完後的圖形進行訓練，已得到想要的模型，但這種方式對應到很多不同的字型，因為各字型的特徵不明確，進而降低了網路學習的效率，也會影響後續的預測的準確度。
為了讓網路能更有效率的學習，本研究將導入一組人工特徵的概念，並利用更輕巧的人工智慧網路對特徵來進行學習，加快預測的速度，此網路我們稱為超輕量網路(Very-Lightweight Network , VLNet)，本研究的字元辨識系統將專注在單一字元的辨識，因此辨識流程中的資料分割及辨識完成的文字後處理，不在本研究的範圍。
使用人工特徵的擷取能將訓練的資料量降低，訓練的資料量降低，就是訓練的輸入降低加上網路的架構簡化，這些因素都能讓網路更小更有效率，計算的需求下降，非常適合於邊緣計算的運用。
最近的深層學習研究都是利用網路加深的方式，藉以提升辨識的準確率，但卻無法在一般的IoT (Internet of Things) 上使用，反而造成實作的不便性。本研究所提出的方法，不但不往這個方向發展，在特徵的擷取也使用更具辨識目標代表性的物理特徵，不採取深層學習的自動特徵擷取，此方法在於獲得更精確的特徵，用於學習，在這個過程中也降低的資料的長度，本研究是用的特徵為135個張量 (Tensor)，而LeNet使用了784個張量，AlexNet使用了196,608個張量，張量的多寡直接影響的運算的時間，因此在IoT的環境下，特徵的使用量，就顯得重要。
本研究的實驗結果，在未知字型的正確率為97.81%，優於LeNet的96.67%，AlexNet的91.15% 及MobileNet V3的95.77%。在已知字型的為99.1%，也優於LeNet的98.59%，AlexNet的95.48% 及MobileNet V3的98.31%。在樹莓派4B的運算時間上，每個字元的預測時間到達4.44至4.61毫秒之間。
應用方面，利用VLNet辨識工廠的出貨單，由精益科技所收集到的3072份樣本，人工分類為58類，分別進行頭部採樣，採取200個特徵，並與AlexNet與MobileNet V3進行實驗，並且比對辨識準確度及辨識速度的比較，準確度VLNet以98.28%領先了AlexNet的 93.68%及MobileNet V3的97.13%。速度上VLNet比MobileNet V3快了將近36倍，分類一張的平均速度是6.14 ms，比AlexNet快上272倍，整體而言，在樹梅派的實驗結果，準確度及速度都是最佳的，足以證明，VLNet除了可以使用在辨識文字，也能使用在文件分類，同樣有優異的表現。

English character recognition has a considerable number of identification methods, there are also quite a few books mentioned, but the font has always been a main reason affecting the accuracy recognition , the recognition process from the configuration analysis of the file, the pre-processing of the file, the division of the text, the end of the text recognition and post-processing, each link in the middle can affect the results of recognition, but in addition to text recognition, the font almost does not affect other stages, and the general common English text type has more than 50 kinds, the current Optical Character Recognition (OCR) may not be able to adapt to each font type. Just as many classic binary methods cannot face all situations, they can separate each picture from the front and back, so that the background is cleared, leaving only the foreground we want.
How do humans learn English and how to write? It basically starts with a stroke, but what is a stroke? It is composed of lines in different directions, so this line in different directions is an important feature. The traditional artificial intelligence network will directly use the pre-processed graphics for training, and the desired model has been obtained, but this method corresponds to many different fonts, because the characteristics of each font are not clear, which reduces the efficiency of network learning and affects the accuracy of subsequent predictions.
In order for the network to learn more efficiently, this study will introduce the concept of stroke features and use a lightweight artificial intelligence network to learn and speed up predictions, this network is called Very-Lightweight Network (VLNet). This study's character recognition system will focus on the recognition of individual characters. Therefore, the data segmentation during the recognition process and the post-processing of the recognized text are beyond the scope of this study. The use of stroke feature acquisition can reduce the amount of data of the graph, reduce the amount of data of the graph, that is, the input of training is reduced and the architecture of the network is simplified, these factors can make the network smaller and more efficient, and the demand for computing is reduced, which is very suitable for the use of edge computing.
Recent research in deep learning has been focused on increasing network depth to improve recognition accuracy, but this approach is not suitable for general Internet of Things (IoT) applications and can cause practical difficulties. The proposed method in this study not only avoids this direction but also uses physical features that are more representative of the recognition target for feature extraction, rather than automatic feature extraction in deep learning. This approach enables more precise feature extraction and reduces the length of the data used for learning. The proposed method uses 135 tensors for feature extraction, while LeNet uses 784 tensors and AlexNet uses 196,608 tensors. The number of tensors directly affects the computational time, making the amount of features used in IoT environments critical.
The experimental results revealed a 97.81% accuracy in recognizing unknown fonts, surpassing LeNet (96.67%), AlexNet (91.15%), and MobileNet V3 (95.77%). For known fonts, the accuracy achieved was 99.1%, outperforming LeNet (98.59%), AlexNet (95.48%), and MobileNet V3 (98.31%). In the operation time of the Raspberry Pi 4B, the prediction time for each character reaches between 4.44 and 4.61 milliseconds.
In application, VLNet was used to recognize shipping documents in a factory setting using 3072 samples collected by Plustek Inc.. The samples were manually categorized into 58 classes and 200 features were extracted from the header. Experiments were conducted using AlexNet and MobileNet V3 to compare recognition accuracy and speed. The results show that VLNet achieved 98.28% accuracy, surpassing AlexNet's 93.68% and MobileNet V3's 97.13%. VLNet was also nearly 36 times faster than MobileNet V3, with an average classification speed of 6.14 ms per image, which is 272 times faster than AlexNet. Overall, the experimental results on the Raspberry Pi 4B demonstrate that VLNet is the most accurate and fastest, and can be used not only for character recognition but also for document classification with outstanding performance.

中文摘要 -------------------------------------------------------------------------- vi
英文摘要 -------------------------------------------------------------------------- viii
誌謝 -------------------------------------------------------------------------- x
目錄 -------------------------------------------------------------------------- xi
表目錄 -------------------------------------------------------------------------- xii
圖目錄 -------------------------------------------------------------------------- xiii
符號說明 -------------------------------------------------------------------------- xiv
一、 緒論-------------------------------------------------------------------- 15
1.1 動機-------------------------------------------------------------------- 15
1.2 目的-------------------------------------------------------------------- 16
1.3 論文架構-------------------------------------------------------------- 17
二、 文獻探討-------------------------------------------------------------- 18
三、 研究內容與方法----------------------------------------------------- 23
3.1 資料收集-------------------------------------------------------------- 23
3.2 前處理----------------------------------------------------------------- 24
3.3 特徵擷取-------------------------------------------------------------- 26
3.4 一維卷積網路架構-------------------------------------------------- 30
3.5 VLNet輕量化網路工作流程-------------------------------------- 31
四、 實驗方法與結果----------------------------------------------------- 33
五、 研究應用-------------------------------------------------------------- 37
5.1 應用架構-------------------------------------------------------------- 37
5.2 應用方法-------------------------------------------------------------- 37
六、 結論-------------------------------------------------------------------- 48
參考文獻 -------------------------------------------------------------------------- 50
附錄一 -------------------------------------------------------------------------- 56
附錄二 -------------------------------------------------------------------------- 57

[1] Y. Yang, X. Lijia and C. Chen, "English character recognition based on feature combination," Procedia Engineering, vol. 24, p. 159–164, 2011.
[2] S.-W. Lee, D.-J. Lee and H.-S. Park, "A new methodology for gray-scale character segmentation and recognition," IEEE transactions on pattern analysis and machine intelligence, vol. 18, p. 1045–1050, 1996.
[3] I. B. Cruz, A. Dı́az Sardiñas, R. Bello Pérez and Y. Sardiñas Oliva, "Learning optimization in a MLP Neural Network Applied to OCR," in MICAI 2002: Advances in Artificial Intelligence: Second Mexican International Conference on Artificial Intelligence Mérida, Yucatán, Mexico, April 22–26, 2002 Proceedings 2, 2002.
[4] N. Sharma, B. Kumar and V. Singh, "Recognition of off-line hand printed english characters, numerals and special symbols," in 2014 5th International Conference-Confluence The Next Generation Information Technology Summit (Confluence), 2014.
[5] J. R. Quinlan, "Induction of decision trees," Machine learning, vol. 1, p. 81–106, 1986.
[6] L. Yuelong, L. Jinping and M. Li, "Character recognition based on hierarchical RBF neural networks," in Sixth International Conference on Intelligent Systems Design and Applications, 2006.
[7] E. d. A. Neves, A. Gonzaga and A. F. F. Slaets, "A multi-font character recognition based on its fundamental features by artificial neural networks," in Proceedings II Workshop on Cybernetic Vision, 1996.
[8] R. Arnold and P. Miklós, "Character recognition using neural networks," in 2010 11th International Symposium on Computational Intelligence and Informatics (CINTI), 2010.
[9] J. Bai, Z. Chen, B. Feng and B. Xu, "Image character recognition using deep convolutional neural network learned from different languages," in 2014 IEEE International Conference on Image Processing (ICIP), 2014.
[10] R. Ptucha, F. P. Such, S. Pillai, F. Brockler, V. Singh and P. Hutkowski, "Intelligent character recognition using fully convolutional neural networks," Pattern recognition, vol. 88, p. 604–613, 2019.
[11] P. W. Frey and D. J. Slate, "Letter recognition using Holland-style adaptive classifiers," Machine learning, vol. 6, p. 161–182, 1991.
[12] H. H. Heriz, H. M. Salah, S. B. A. Abdu, M. M. El Sbihi and S. S. Abu-Naser, "English Alphabet Prediction Using Artificial Neural Networks," 2018.
[13] J. Parkinson and B. Khurana, "Temporal order of strokes primes letter recognition," The Quarterly Journal of Experimental Psychology, vol. 60, p. 1265–1274, 2007.
[14] Y. Y. Tang, B. F. Li, H. Ma and J. Lin, "Ring-projection-wavelet-fractal signatures: a novel approach to feature extraction," IEEE Transactions on circuits and systems II: Analog and digital signal processing, vol. 45, p. 1130–1134, 1998.
[15] W. B. Lund, D. J. Kennard and E. K. Ringger, "Combining multiple thresholding binarization values to improve OCR output," in Document Recognition and Retrieval XX, 2013.
[16] N. A. Shaikh, Z. A. Shaikh and G. Ali, "Segmentation of Arabic text into characters for recognition," in Wireless Networks, Information Processing and Systems: International Multi Topic Conference, IMTIC 2008 Jamshoro, Pakistan, April 11-12, 2008 Revised Selected Papers, 2009.
[17] S. Khalid, T. Khalil and S. Nasreen, "A survey of feature selection and feature extraction techniques in machine learning," in 2014 science and information conference, 2014.
[18] K. K. M. Shreyas, S. Rajeev, K. Panetta and S. S. Agaian, "Fingerprint authentication using geometric features," in 2017 IEEE International Symposium on Technologies for Homeland Security (HST), 2017.
[19] C.-H. Chen and K.-W. Huang, "Digit Recognition using Composite Features with Decision Tree Strategy," International Journal of Interactive Multimedia and Artificial Intelligence, vol. In Press, p. 1, January 2022.
[20] C.-H. Chen, Z.-H. Huang and K.-W. Huang, "Recognition of Handwritten English and Digits Using Stroke Features and MLP," in 2022 Joint 12th International Conference on Soft Computing and Intelligent Systems and 23rd International Symposium on Advanced Intelligent Systems (SCIS&ISIS), 2022.
[21] C.-H. Chen, W.-J. Wu and K.-W. Huang, "Fast Digit Recognition Using Lightweight Neural Network Model," in 2023 International Conference on Consumer Electronics-Taiwan (ICCE-Taiwan), 2023.
[22] A. S. Tarawneh, A. B. Hassanat, D. Chetverikov, I. Lendak and C. Verma, "Invoice classification using deep features and machine learning techniques," in 2019 IEEE Jordan International Joint Conference on Electrical Engineering and Information Technology (JEEIT), 2019.
[23] P. Singh and S. Budhiraja, "Feature extraction and classification techniques in OCR systems for handwritten Gurmukhi Script–a survey," International Journal of Engineering Research and Applications (IJERA), vol. 1, p. 1736–1739, 2011.
[24] R. Verma and J. Ali, "A-survey of feature extraction and classification techniques in OCR systems," International Journal of Computer Applications & Information Technology, vol. 1, p. 1–3, 2012.
[25] A. Yang, X. Yang, W. Wu, H. Liu and Y. Zhuansun, "Research on feature extraction of tumor image based on convolutional neural network," IEEE access, vol. 7, p. 24204–24213, 2019.
[26] G. S. Lehal, "Optical character recognition of Gurmukhi script using multiple classifiers," in Proceedings of the international workshop on multilingual OCR, 2009.
[27] T. Kobayashi, A. Hidaka and T. Kurita, "Selection of histograms of oriented gradients features for pedestrian detection," in Neural Information Processing: 14th International Conference, ICONIP 2007, Kitakyushu, Japan, November 13-16, 2007, Revised Selected Papers, Part II 14, 2008.
[28] S. Singh, A. Aggarwal and R. Dhir, "Use of Gabor Filters for recognition of Handwritten Gurmukhi character," International Journal of Advanced Research in Computer Science and Software Engineering, vol. 2, 2012.
[29] A. Shawon, M. J.-U. Rahman, F. Mahmud and M. A. Zaman, "Bangla handwritten digit recognition using deep cnn for large and unbiased dataset," in 2018 international conference on Bangla speech and language processing (ICBSLP), 2018.
[30] T. A. Assegie and P. S. Nair, "Handwritten digits recognition with decision tree classification: a machine learning approach," International journal of electrical and computer engineering (IJECE), vol. 9, p. 4446–4451, 2019.
[31] V. Rajinikanth, S. Kadry, R. González-Crespo and E. Verdú, "A study on RGB image multi-thresholding using Kapur/Tsallis entropy and moth-flame algorithm," 2021.
[32] H. Hosseini, B. Xiao and R. Poovendran, "Google's cloud vision api is not robust to noise," in 2017 16th IEEE international conference on machine learning and applications (ICMLA), 2017.
[33] J. Memon, M. Sami, R. A. Khan and M. Uddin, "Handwritten optical character recognition (OCR): A comprehensive systematic literature review (SLR)," IEEE Access, vol. 8, p. 142642–142668, 2020.
[34] J. Kim, J. Kim, H. Kim, M. Shim and E. Choi, "CNN-based network intrusion detection against denial-of-service attacks," Electronics, vol. 9, p. 916, 2020.
[35] A. A. Barbhuiya, R. K. Karsh and R. Jain, "CNN based feature extraction and classification for sign language," Multimedia Tools and Applications, vol. 80, p. 3051–3069, 2021.
[36] W. Cheng, Y. Sun, G. Li, G. Jiang and H. Liu, "Jointly network: a network based on CNN and RBM for gesture recognition," Neural Computing and Applications, vol. 31, p. 309–323, 2019.
[37] J. H. Lee, K.-Y. Kim and Y. Shin, "Feature image-based automatic modulation classification method using CNN algorithm," in 2019 International Conference on Artificial Intelligence in Information and Communication (ICAIIC), 2019.
[38] V. Savitha, M. Karthick and T. Kalaikumaran, "Parasitic Egg detection from Microscopic images using Convolutional Neural Networks," Tamjeed Journal of Healthcare Engineering and Science Technology, vol. 1, p. 24–34, 2023.
[39] Y. LeCun, L. Bottou, Y. Bengio and P. Haffner, "Gradient-based learning applied to document recognition," Proceedings of the IEEE, vol. 86, p. 2278–2324, 1998.
[40] A. Krizhevsky, I. Sutskever and G. E. Hinton, "Imagenet classification with deep convolutional neural networks," Communications of the ACM, vol. 60, p. 84–90, 2017.
[41] K. Simonyan and A. Zisserman, "Very deep convolutional networks for large-scale image recognition," arXiv preprint arXiv:1409.1556, 2014.
[42] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke and A. Rabinovich, "Going deeper with convolutions," in Proceedings of the IEEE conference on computer vision and pattern recognition, 2015.
[43] Z. Zhong, L. Jin and Z. Xie, "High performance offline handwritten chinese character recognition using googlenet and directional feature maps," in 2015 13th international conference on document analysis and recognition (ICDAR), 2015.
[44] K. He, X. Zhang, S. Ren and J. Sun, "Deep residual learning for image recognition," in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016.
[45] A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. Andreetto and H. Adam, "Mobilenets: Efficient convolutional neural networks for mobile vision applications," arXiv preprint arXiv:1704.04861, 2017.
[46] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov and L.-C. Chen, "Mobilenetv2: Inverted residuals and linear bottlenecks," in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018.
[47] A. Howard, M. Sandler, G. Chu, L.-C. Chen, B. Chen, M. Tan, W. Wang, Y. Zhu, R. Pang, V. Vasudevan and others, "Searching for mobilenetv3," in Proceedings of the IEEE/CVF international conference on computer vision, 2019.
[48] S. Chen, Y. He, J. Sun and S. Naoi, "Structured document classification by matching local salient features," in Proceedings of the 21st International Conference on Pattern Recognition (ICPR2012), 2012.
[49] P. Sarkar, "Learning image anchor templates for document classification and data extraction," in 2010 20th International Conference on Pattern Recognition, 2010.
[50] S. Usilin, D. Nikolaev, V. Postnikov and G. Schaefer, "Visual appearance based document image classification," in 2010 IEEE International Conference on Image Processing, 2010.
[51] A. W. Harley, A. Ufkes and K. G. Derpanis, "Evaluation of deep convolutional nets for document image classification and retrieval," in 2015 13th International Conference on Document Analysis and Recognition (ICDAR), 2015.
[52] N. Audebert, C. Herold, K. Slimani and C. Vidal, "Multimodal deep networks for text and image-based document classification," in Joint European Conference on Machine Learning and Knowledge Discovery in Databases, 2020.
[53] M. Z. Afzal, A. Kölsch, S. Ahmed and M. Liwicki, "Cutting the error by half: Investigation of very deep cnn and advanced training strategies for document image classification," in 2017 14th IAPR International Conference on Document Analysis and Recognition (ICDAR), 2017.
[54] J. T. Townsend, "Theoretical analysis of an alphabetic confusion matrix," Perception & Psychophysics, vol. 9, p. 40–50, 1971.
[55] R. R. Selvaraju, A. Das, R. Vedantam, M. Cogswell, D. Parikh and D. Batra, "Grad-CAM: Why did you say that?," arXiv preprint arXiv:1611.07450, 2016.
[56] B. Zhou, A. Khosla, A. Lapedriza, A. Oliva and A. Torralba, "Learning deep features for discriminative localization," in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016.
[57] L. Chen, J. Chen, H. Hajimirsadeghi and G. Mori, "Adapting grad-cam for embedding networks," in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2020.
[58] R. Fu, Q. Hu, X. Dong, Y. Guo, Y. Gao and B. Li, "Axiom-based grad-cam: Towards accurate visualization and explanation of cnns," arXiv preprint arXiv:2008.02312, 2020.