隨著現代工業的蓬勃發展，產品質量和生產良率逐漸受到關注，產品表面缺陷檢測的應用已成為生產過程中不可或缺的重要角色。本研究以線圈為實驗目標，並自行收集實驗訓練和測試數據。其中，將拍攝的圖像分割為224×224大小的圖像以進行訓練和測試。在這項研究中，使用不同的預訓練模型以原始資料庫進行訓練，但其結果不足以應用在真實產線上。藉由上述實驗，我們選擇最佳的訓練超參數進行訓練，以確保訓練參數不會引起識別結果的不穩定。通過對整個模型的重新訓練，我們首先選擇殘差神經網絡，以原始資料訓練的情況下，殘差神經網絡對原始數據庫的驗證準確性可以達到98％。但是容易發生過度擬合現像。此外，我們比較了殘差神經網絡的較短模型的分類能力。其次，在建立具有自行開發的軟體的資料庫之後，將殘差網路由不同數量的遮罩影像組成的訓練集進行進一步比較。與原始數據庫進行驗證，最佳識別能力是使用60個遮罩影像的實驗結果最佳。其驗證準確率可達95%，並大大減少了訓練期間驗證準確性的抖動（減少了過度擬合）。最後，通過調整模型最終輸出的預測值，結果表明模型在不同epoch訓練下大多可以提高對原始數據庫驗證的準確性，並且使用外來測試數據集的測試準確性也可以提高約5%。最後，實驗結果顯示，此系統可在正常環境中良好運作。

Due to the vigorous development of modern industries, the quality and productivity in production have gradually received attention, and the application of product surface defect detection has become an indispensable and important role in production. This research uses coils as experimental data, and collected experimental training and test data by ourselves. Among them, we segmented the film images into 224x224 size images for training and testing, and create an original database after the image is segmented. In this study, we first compared the advantages, disadvantages and performance of different pre-training models with the original database. As a result of the second experiment, the best training hyper-parameters are selected for training to ensure that the training parameters will not cause instability of the recognition results. After training all the parameters of the model, we found that the residual neural network performs is better, and the validation accuracy of the residual neural network to the original database is 98%. Secondly, after establishing a database with a self-developed program, the residual network is further compared with the training set composed of different numbers of masked images. The identification ability against the original database is the result of 60 masked images. Achieve 95% recognition rate and greatly reduce the chattering of its validation accuracy during training. Finally, by adjusting the one-hot value of the final output of the model, the results show that most of the accuracy of the original database can be improved, and the heterogenous test data can also be improved about 5% with the model at different epochs. In addition, we also compared the classification capabilities of shorter models. At the end of this study, some conclusions and future works are given.

[1] S. Wen, Z. Chen, and C. Li, “Vision-based surface inspection system for bearing rollers using convolution neural networks,” apply science, vol. 8, no. 12, pp. 2565-2589 2018.
[2] C. -Y. Lin, C. -H. Chen, and C. -Y. Yang, “Cascading convolution neural network for steel surface defect detection,” Advances in Artificial Intelligence, Software and Systems Engineering, vol. 965, pp. 202-212, 2019
[3] H. Y. T. Ngan, G. K. H. Pang, and N. H. C. Yung, “Automated fabric defect detection-A review,” Image and Vision Computing, Vol. 29, pp 442-458, 2011.
[4] C. -H. Chan, and G. K. H. Pang, “Fabric defect detection by Fourier analysis,” IEEE Transaction on Industry applications, vol. 36, no. 5, pp. 1267-1276, 2000.
[5] A. Ahmadian, and A. Mostafa, “An efficient texture classification algorithm using Gabor Wavelet,” Proceedings of the 25th Annual International Conference of the IEEE EMBS, pp. 930-933, 2003.
[6] S. Stan, G. Palubinskas, and M. Datcu, “Bayesian selection of the neighborhood order for Gauss–Markov texture models,” Pattern Recognition Letters, vol. 23, pp. 1229-1238, 2002.
[7] K. Fukushima, and S. Miyake, “A self-organizing neural network model for a mechanism of visual pattern recognition,” Competition and Cooperation in Neural Nets, vol. 45, pp. 267-285, 1982.
[8] A. Krizhevsky, I. Sutskever, and G. E. Hinton. “ImageNet classification with deep convolutional neural networks,” Neural Information Processing System, 2012.
[9] A. Oord, S. Dieleman, and B. Schrauwen, “Deep content-based music recommendation,” Neural Information Processing System, 2013.
[10] C. Ronan, and W. Jason. “A unified architecture for natural language processing: Deep neural networks with multitask learning,” Proceeding of the 25th international conference on Machine learning, pp. 160-167, 2008.
[11] S. Avraam, P. Nikolaos, and T. Anastasios, “Forecasting stock prices form the limit order book using convolution neural networks,” IEEE 19th Conference on Business Informatics, pp. 7-12, 2017.
[12] Q.Y. Wu, H. Wu, and X. Zhou, “Online transfer learning with multiple homogeneous or heterogeneous sources,” IEEE Transactions on Knowledge and Data Engineering, Vol. 29, no. 7, pp. 1494-1507, 2017.
[13] W. Dai, Q. Yang, and G.-R. Xue, “Boosting for transfer learning,” International conference on Machine learning, pp. 193-200, 2010.
[14] J. -T. Huang, J. Li, and D. Yu, “Cross-language knowledge transfer using multilingual deep neural network with shared hidden layers,” IEEE International Conference on Auoustics, Speech, and Signal Processing, pp. 7304-7308. 2013.
[15] V. Chouhan, S. K. Singh, and A. Khamparia, “A novel transfer learning based approach for pneumonia detection in chest X-ray images,” Applied Sciences, vol. 10, no. 2, 2020.
[16] JM-B Francisco., S. Fiammetta, MJ. Jose, U. Daniel, and F. Leonardo, “Forward noise adjustment scheme for data augmentation,” IEEE Symposium Series on Computational Intelligence, pp. 728-734, 2018.
[17] Zhong Z., Zheng L., Kang G., Li S., and Yang Y., “Random erasing data augmentation,” Computer Vision and Pattern Recognition, pp. 1708-1717, 2017.
[18] C. Shorten, and T. M. Khoshgoftaar, “A survey on image data augmentation for deep learning,” Journal of Big Data, vol. 6, no. 60, pp. 1186-1233, 2019.
[19] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” IEEE Conference on Computer Vision and Pattern Recognition, pp. 770-778. 2015.
[20] Z. Liu, C. Li, and Q. Zhao, “A fabric defect detection algorithm via context-based local texture saliency analysis,” International Journal of Clothing Science and Technology, vol. 27, no. 5, 2016.
[21] R. Stojanovic, P. Mitropulos, and C. Koulamas, “Real-time vision-based system for textile fabric inspection,” Real-time Imaging, vol. 7. pp. 507-518. 2001.
[22] X. Wu, K. Cao, and X. Gu. “A surface defect detection based on convolutional neural network,” Computer Vision Systems, pp. 185-194, 2017.
[23] S. Beura, B. Majhi, and R. Dash, “Mammogram classification using two dimensional discrete wavelet transform and gray-level co-occurrence matrix for detection of breast cancer,” Neurocomputing, vol. 154, pp. 1-14, 2014.
[24] W. C. Li, and D. M. Tsai, “Wavelet-based defect detection in solar wafer images with inhomogeneous texture,” Pattern Recognition, pp. 742-756, 2012.
[25] M. S. Reis, and A. Bauer, “Image-based classification of paper surface quality using wavelet texture analysis,” Computers and Chemical Engineering, pp. 2014-2021, 2010.
[26] Yong Li, and Jinbing Xu, “Electronic product surface defect detection based on a MSSD network,” Information Technology, Networking, Electronic and Automation Control Conference, pp. 773-777, 2020.
[27] Meenakshi Garg, and Gaurav Dhiman, “Deep convolution neural network approach for defect inspection of textured surface,” Journal of Institute of Electronics and Computer, vol. 2, pp. 28-38, 2020.
[28] Hailang Chen, “CNN-based surface defect detection of smartphone protective screen,” Journal of Physics: Conference Series, vol. 1616, 2020.
[29] L. Xiao, B. Wu, and Y. Hu, “Surface defect detection using image pyramid,” IEEE Sensor Journal, vol. 20, no. 13, pp. 7181-7188, 2020.
[30] Xiaoming Lv, Fajie Duan, Jia-Jia Jiang, Xiao Fu, and Lin Gan, “Deep active learning for surface defect detection,” Sensors, vol. 20, no. 6, pp. 1650-1661, 2020.
[31] Ying Liang, Ke Xu, and Peng Zhou, “Mask gradient response-based threshold segmentation for surface defect detection of milled aluminum ingot,” Sensors, vol. 20, no. 16, pp. 4519-4540, 2020.