隨著資訊科技和機器學習演算法的快速發展，工廠內如缺陷辨識等問題能夠使用大量資料訓練複雜的機器學習模型來取代。然而，在實際應用時，大型資料集的取得往往是困難或成本高昂的。
透過L2-loss與預訓練權重，本研究討論使用一個結合卷積神經網路-支持向量機與遷移學習的模型所帶來的潛在優勢，並提出一個三階段程序（包括資料前處理，訓練多種不同配置的模型，和模型評估與最佳模型選取）於缺陷之自動判別。經由一個真實個案之分析，本研究收集發光二極體引線框缺陷之數據進行判別與分類。執行結果顯示，使用本研究之模型能夠提升分類準確率高達18.03%，並縮短至少一半的訓練時間。最終，所選取的最佳的模型配置為使用ResNet50-V2的架構與預訓練之權重為基底，加上一層客製輸出全連接層，以及使用L2-loss為模型損失函數；這個模型可用來建置引線框缺陷的自動缺陷分類系統。

With the rapid development of information technology and machine learning algo-rithms, tasks like defect classification can be replaced with high-capacity machine learning models trained by large amount of data, but in real life applications, large da-tasets are often difficult or costly to obtain.
This study discusses the potential benefits of employing a hybrid convolutional neural network-support vector machine (CNN-SVM) combined with Transfer Learning via L2-loss and pre-trained weights. A three-phase procedure, including data prepro-cessing, training multiple differently-configured models, and performance evaluation and optimal model selection is proposed, and a LED lead frame dataset obtained from real world is used to evaluate the proposition. It is found that when utilizing the pro-posed system, classification accuracy can be boosted up to 18.03%, while reducing at least half of the training time. Eventually, the optimal configuration of ResNet50-V2 as the base model with pre-trained weights, a custom output fully connected layer, and L2-loss is chosen as the model to construct the automatic defect classification system for the LED lead frame problem.

[1] Bennett, M.H., Tobin, K.W. Jr., Gleason, S.S. (1995). Automatic defect classifica-tion: status and industry trends. Proc. SPIE 2439, Integrated Circuit Metrology, In-spection, and Process Control IX.
[2] Breaux, L., Singh, B. (1995). Automatic defect classification system for patterned semiconductor wafers. Proceedings of International Symposium on Semiconductor Manufacturing. doi: 10.1109/ISSM.1995.524362.
[3] Cheon, S., Lee, H., Kim, C.O., Lee, S.H. (2019). Convolutional neural network for wafer surface defect classification and the detection of unknown defect class. IEEE Transactions on Semiconductor Manufacturing, 32 (2), pp. 163-170.
[4] Chou, P., Rao, A., Sturzenbecker, M., Brecher, V. (1993). Automatic defect classi-fication for integrated circuits. Proc. SPIE 1907, Machine Vision Applications in Industrial Inspection, pp. 95-103.
[5] Chou, P., Rao, A., Sturzenbecker, M. (1997). Automatic defect classification for semiconductor manufacturing. Machine Vision and Applications. 9 (4), pp. 201-214.
[6] Deng, Y.S., Luo, A.C., Dai, M.J. (2018). Building an automatic defect verification system using deep neural network for PCB defect classification. 2018 4th Interna-tional Conference on Frontiers of Signal Processing (ICFSP). doi: 10.1109/ICFSP.2018.8552045.
[7] He, K., Zhang, X., Ren, S., Sun, J. (2015). Deep residual learning for image recog-nition. arXiv: 1512.03385.
[8] He, K., Zhang, X., Ren, S., Sun, J. (2016). Identity mappings in deep residual net-works. arXiv: 1603.05027.
[9] Hertel, L., Barth, E., Käster, T., Martinetz, T. (2015). Deep convolutional neural networks as generic feature extractors. 2015 International Joint Conference on Neural Networks (IJCNN). doi: 10.1109/IJCNN.2015.7280683.
[10] Kang, J., Park, Y.J., Lee, J., Wang, S., Eom, D.S. (2017). Novel leakage detection by ensemble CNN-SVM and graph-based localization in water distribution systems. IEEE Transactions on Industrial Electronics, 65 (5), pp. 4279-4289.
[11] Kang, S.B., Lee, J.H., Song, K.Y., Pahk, H.J. (2009). Automatic defect classifica-tion of TFT-LCD panels using machine learning. 2009 IEEE International Sympo-sium on Industrial Electronics. doi: 10.1109/ISIE.2009.5213760.
[12] Kingma, D.P., Ba, J. (2014). Adam: A method for stochastic optimization. arXiv: 1412.6980.
[13] Krizhevsky, A., Sutskever, I., Hinton, G.E. (2012). ImageNet classification with deep convolutional neural networks. Advances in Neural Information Processing Systems, 25, pp. 1106-1114.
[14] LeCun, Y., Bottou, L., Bengio, Y., and Haffner, P. (1998). Gradient-based learning applied to document recognition. Proceedings of the IEEE, 89 (11), pp 2278-2323.
[15] Lin, H.D. (2009). Automated defect inspection of light-emitting diode chips using neural network and statistical approaches. Expert Systems with Applications. pp. 219-226.
[16] Lin, H., Li, B., Wang X., Shu Y., and Niu, S. (2019). Automated defect inspection of LED chip using deep convolutional neural network. Journal of Intelligent Man-ufacturing, 30 (6), pp. 2525-2534.
[17] Ng, A. (2004). Feature selection, L1 vs. L2 regularization, and rotational invari-ance. In Proceedings of the Twenty-First International Conference on Machine Learning. pp.78.
[18] Ng, H.W., Nguyen, V.D., Vonikakis, V., Winkler, S. (2015). Deep learning for emotion recognition on small datasets using transfer learning. Proceedings of the 2015 ACM on International Conference on Multimodal Interaction, pp. 443-449.
[19] Pan, S.J., Yang, Q. (2009). A survey on transfer learning. IEEE Transactions on Knowledge and Data Engineering, 22 (10), pp. 1345-1359.
[20] Park, J.K., Kwon, B.K., Park, J.H. (2016). Machine learning-based imaging sys-tem for surface defect inspection. International Journal of Precision Engineering and Manufacturing-Green Technology, 3 (3), pp. 303-310.
[21] Simonyan, K. and Zisserman, A. (2014). Very deep convolutional networks for large-scale image recognition. arXiv: 1409.1556.
[22] Skumanich, A. (1999). Process and yield improvement based on fast in-line auto-matic defect classification. Proc. SPIE 3884, In-Line Methods and Monitors for Process and Yield Improvement. doi: 10.1117/12.361342.
[23] Sun, X., Liu, L., Li, C., Yin, J., Zhao, J., Si, W. (2019). Classification for remote sensing data with improved CNN-SVM method. IEEE Access, pp. 164507-164516.
[24] Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., and Rabinovich, A. (2014). Going deeper with convolutions. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). doi: 10.1109/CVPR.2015.7298594.
[25] Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z. (2016). Rethinking the inception architecture for computer vision. 2016 IEEE Conference on Computer Vi-sion and Patter Recognition (CVPR). doi: 10.1109/CVPR.2016.308.
[26] Tang, Y. (2013). Deep learning using linear support vector machines. arXiv: 1306.0239.
[27] Tuv, E., Guven, M., Ennis, P., Lee, D.H.L. (2018) Faster, more accurate defect classification using machine vision. IT@Intel White Paper. Retrieved from https://www.intel.com/content/dam/www/public/us/en/documents/best-practices/faster-more-accurate-defect-classification-using-machine-vision-paper.pdf.
[28] Yang, H., Mei, S., Song, K., Tao, B., Yin, Z. (2017). Transfer-learning-based online mura defect classification. IEEE Transactions on Semiconductor Manufac-turing, 31 (1), pp. 116-123.