在軋延鋁片的過程中，可能會造成輥軸表面些微的刮傷，因此在軋延固定數量的鋁片之後，輥軸需要被重新研磨以移除表面的鋁粉或刮痕，保證後續軋延的品質。以砂輪進行輥軸表面研磨的過程中，猛烈的撞擊會造成砂屑不定時掉落，掉落的砂屑在相對運動中會在輥軸表面形成鋸齒型與坑洞型等額外的缺陷。當輥軸本身表面缺陷過多時，鋁片被輥軸軋延時也可能因輥軸本身的缺陷而留下刮痕或坑洞，將嚴重影響鋁片產品品質，因此輥軸表面品質控管是很重要的一環。
在這篇論文中，我們與中鋁合作，基於卷積神經網路的架構，發展了針對輥面毛邊缺陷的檢測方法。此系統可供工廠工作人員現場使用，除了可對當前研磨品質提供量化的分析之外，亦可用作分析不同品質砂輪、研磨參數等之輔助參考。
許多常見的工業缺陷檢測方法，常用於影像背景規律或單純的資料集之上。由於輥軸表面影像背景複雜且變異性高，本篇論文以深度學習檢測演算法Faster R-CNN為基準，提出了四個修改，以提升缺陷檢測的準確度。首先在特徵提取的部分，針對長短變異性巨大的缺陷使用了特徵金字塔網路，並提出加權供網路的損失函數使用，提升對不同長短缺陷的檢測能力；針對區域提取網路中正負樣本的選取條件進行設計，可以提升整體架構中提出前景框的準確度；在第二階段先調整前景框位置再進行辨識，讓辨識時的輸入特徵更加準確；多解析度的辨識器針對長短不一的缺陷，有效提升缺陷辨識的精準度。
本論文提出的架構應用在輥面毛邊缺陷的資料集，鋸齒型缺陷與坑型缺陷在F1-score精準度上分別可達80.9%與85.0%，相較於一般的Faster R-CNN在鋸齒型缺陷的recall提升了11.5%、precision提升了7.4%，坑洞型缺陷的precision提升了10%。在整體效果上，鋸齒型缺陷提升了9.5%、坑洞型缺陷提升了4.9%。以此證明本篇論文提出的架構應用在輥面毛邊缺陷資料集相較於一般Faster R-CNN可以獲得較好的檢測率與精準度，期望未來應用在研磨現場時，可以為提升輥面研磨品質貢獻一份心力。

The process of rolling aluminum sheets may cause slight scratches on the surface of the roller. After rolling a fixed number of aluminum sheet, staffs need to regrind the roller to remove aluminum powder or scratches on the surface to ensure the quality for subsequent rolling. When grinding the surface of the roller with a grinding wheel, the excessive force will cause the sand to fall gradually. The falling sand abrades additional defects such as “Ridge-type defect” and “Pit-type defect” on the roller surface. If too many defects exist on the surface of the roller, the aluminum sheet rolled by the roller may also leave scratches or pits, which will seriously affect the quality of the aluminum sheet product. Hence, the quality control for the surface of the roller is an important issue.
In this paper, we collaborate with China Steel Aluminum Corporation (CSAC) to develop a detection method for burr defects on the roller surface based on convolutional neural network. This system assists on-site factory staff members in quantifying the current grinding quality and analyzing different quality grinding wheels and grinding parameters.
Many common industrial defect detection methods are often used on defect inspection with relatively regular or simple data sets. Due to the high variability background in the defect image on roller surface, we propose four modifications based on the deep learning detection algorithm, Faster R-CNN, to improve the performance of defect detection. First of all, the weighting feature pyramid network, which enhances the detection ability of defects in different lengths, is proposed for feature extraction. Second, designing for the selection conditions of the positive and negative samples in the region proposal network improves the accuracy of foreground box extraction. Third, the module of refining and classification R-CNN adjusts the position of the foreground proposal boxes, so that the input features for classification are more accurate. At last, the multi-resolution classifier effectively increases the performance of classification for defects in various lengths.
We implement the proposed architecture with the dataset of roller surface burr defects. In the experimental results, the performance of Ridge-type defects and Pit-type defects can reach 80.9% and 85.0% respectively. Compared with the general Faster R-CNN, the proposed network improves the recall of the Ridge-type defects by 11.5%, the precision of Ridge-type defects by 7.4%, and the precision of the Pit-type defects by 10%. In terms of the overall effect, we boost the accuracies of Ridge-type defects by 9.5%, and Pit-type defects by 4.9%. The experimental result depicts that the proposed architecture is reliable in the dataset of roller surface burr defects. With a better detection rate and accuracy compared to the general Faster R-CNN, it is applicable to the grinding site in the future for improving the grinding quality of roller surface.

[1] J. Long, E. Shelhamer and T. Darrell, “Fully Convolutional Networks for Semantic Segmentation,” in Conference on Computer Vision and Pattern Recognition, 2015.
[2] D. E. Rumelhart, G. E. Hinton and R. J. Williams, “Learning representations by back-propagating errors,” Nature, pp. 533-536, 1986.
[3] S. Ren, K. He, and R. B. Girshick el al., “Faster R-CNN: Towards Real-Time Object Detection with Region proposal networks,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 39, no. 6, pp. 1137-1149, 2017.
[4] W. Liu, D. Anguelov, D. Erhan et al., “SSD: Single Shot MultiBox Detector,” in the European Conference on Computer Vision (ECCV), pp. 21-37, 2016.
[5] J. Redmon, S. Divvala, R. B. Girshick, and A. Farhadi, “You Only Look Once: Unified, Real-Time Object Detection,” in the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 779-788, 2016.
[6] K. Simonyan and A. Zisserman, “Very Deep Convolutional Networks for Large-Scale Image Recognition,” in the International Conference on Learning Representations (ICLR), 2015.
[7] K. He, X. Zhang, S. Ren and J. Sun, “Deep Residual Learning for Image Recognition,” in Computer Vision and Pattern Recognition, 2015.
[8] G. Huang, Z. Liu, L. v. d. Maaten and K. Q. Weinberger, “Densely Connected Convolutional Networks,” in Computer Vision and Pattern Recognition, 2016.
[9] J. Deng, W. Dong, and R. Socher, et al., “ImageNet: A large-scale hierarchical image database,” in the IEEE Conference on Computer Vision and Pattern Recognition(CVPR), 2009.
[10] M. Everingham, L. V. Gool, and C. K. I. Williams, et al., “The Pascal Visual Object Classes (VOC) Challenge,” in the International Journal of Computer Vision, 2009.
[11] Elbehiery, H., A. Hefnawy, and M. Elewa. “Surface defects detection for ceramic tiles using image processing and morphological techniques,” 2005.
[12] Iivarinen, Jukka. “Surface defect detection with histogram-based texture features,” Intelligent robots and computer vision xix: Algorithms, techniques, and active vision. Vol. 4197. International Society for Optics and Photonics, 2000.
[13] Song, Kechen, and Yunhui Yan. “A noise robust method based on completed local binary patterns for hot-rolled steel strip surface defects,” Applied Surface Science 285: p. 858-p. 864, 2013.
[14] Z.H. Guo, L. Zhang, D. Zhang, “A completed modeling of local binary pattern operator for texture classification,” IEEE Transactions on Image Processing, 19, pp. 1657-1663, 2010.
[15] Sharifzadeh, M., et al. “Detection of steel defect using the image processing algorithms,” The International Conference on Electrical Engineering. Vol. 6. No. 6th International Conference on Electrical Engineering ICEENG 2008. Military Technical College, 2008.
[16] Gonzales, R. and Woods, R., “Digital image processing,” Rahall Graphics, India, 2003.
[17] K. Liu, H. Wang, H. Chen, E. Qu, Y. Tian, and H. Sun, “Steel surface defect detection using a new Haar–weibull-variance model in unsupervised manner,” IEEE Trans. Instrum. Meas., vol. 66, no. 10, pp. 2585–2596, Oct. 2017.
[18] S. Faghih-Roohi, S. Hajizadeh, A. Núñez, R. Babuska and B. De Schutter, “Deep convolutional neural networks for detection of rail surface defects,” 2016 International Joint Conference on Neural Networks (IJCNN), Vancouver, BC, pp. 2584-2589, doi: 10.1109/IJCNN.2016.7727522, 2016.
[19] Y. Wang, R. Yoshihashi, and R. Kawakami el al., “Unsupervised anomaly detection with compact deep features for wind turbine blade images taken by a drone,” IPSJ Transactions on Computer Vision and Applications, 2019.
[20] Cortes, Corinna, and Vladimir Vapnik. “Support-vector networks,” Machine learning 20.3 (1995): 273-297, 1995.
[21] Z. Zhao, B. Li, R. Dong, and P. Zhao, “A Surface Defect Detection Method Based on Positive Samples,” Pacific Rim International Conference on Artificial Intelligence(PRICAI), p473-481, 2018.
[22] He, Yu, et al. “An end-to-end steel surface defect detection approach via fusing multiple hierarchical features,” IEEE Transactions on Instrumentation and Measurement 69.4 (2019): 1493-1504, 2019.
[23] Girshick, Ross. “Fast r-cnn,” Proceedings of the IEEE international conference on computer vision. 2015.
[24] R. O. Duda, and Peter E. Hart., “Use of the Hough transformation to detect lines and curves in pictures,” Communications of the ACM, vol. 15, no. 1, pp. 11-15, 1972.
[25] Lin, Tsung-Yi, et al. “Feature pyramid networks for object detection,” Proceedings of the IEEE conference on computer vision and pattern recognition. 2017.
[26] He, Kaiming, et al. “Mask r-cnn,” Proceedings of the IEEE international conference on computer vision. 2017.
[27] M. Jaderberg, K. Simonyan, A. Zisserman, and K. Kavukcuoglu. “Spatial transformer networks,” in NIPS, 2015.
[28] Scherer, Dominik, Andreas Müller, and Sven Behnke. “Evaluation of pooling operations in convolutional architectures for object recognition,” International conference on artificial neural networks. Springer, Berlin, Heidelberg, 2010.
[29] “Non-Maximum-Suppression,” [Online]. Available: https://medium.com/@chih.sheng.huang821/%E6%A9%9F%E5%99%A8-%E6%B7%B1%E5%BA%A6%E5%AD%B8%E7%BF%92-%E7%89%A9%E4%BB%B6%E5%81%B5%E6%B8%AC-non-maximum-suppression-nms-aa70c45adffa
[30] Lin, Tsung-Yi, et al. “Focal loss for dense object detection,” Proceedings of the IEEE international conference on computer vision. 2017.
[31] Paszke, Adam, et al. “Automatic differentiation in pytorch,” 2017.
[32] Saad, David. “Online algorithms and stochastic approximations,” Online Learning 5 (1998): 6-3, 1998.
[33] Sutskever, Ilya, et al. “On the importance of initialization and momentum in deep learning,” International conference on machine learning. 2013.
[34] J. Davis and M. Goadrich, “The Relationship between Precision-Recall and ROC Curves,” Proc. Int’l Conf. Machine Learning, pp. 233-240, 2006.
[35] Sasaki Y. “The truth of the f-measure,” Teaching, Tutorial Materials version: 26th October, 2007.