沖壓件的需求日益增多，現今許多工廠仍以人工目檢管控品質，長期下來目檢員視力會受到一定的傷害，而且每個人對瑕疵的判定也有所差異。為解決此問題，本研究設計一套專業的拍攝架構，我們先取出清晰的瑕疵影像以降低周遭金屬紋路的干擾，接著搭配本實驗室所研發的傅立葉域奇異值分解強化影像，再結合深度學習進行瑕疵檢出與分類。實驗結果顯示在原始圖像判定瑕疵正確率只有但在使用傅立葉域奇異值分解將圖像增強後，正確率來到 52.8%。在分類準確率方面，原始圖像的分類準確率只有 84%，影像增強後分類率來到 96%，由此可證明經過增強 的圖像對深度學習有很大的助益。本研究產出能夠有效避免目檢員因為長時間工作導致的誤判，以及視力受損的問題。

The demand for stamping parts is increasing. However, many factories still use manual visual inspection to control product quality. In the long term, visual inspectors will suffer certain damage to their eyesight, while the judgement on defect from various person is also different. In order to solve this problem, this research proposed a novel framework. We first captured clear and flawed images to reduce the interference of surrounding metal lines and then used singular value decomposition in the Fourier domain, which has been developed by our laboratory, to enhance defects. And then, deep learning is adopted to perform defect detection and classification. Experimental results show that the correct rate of defect detection in the original images is only 47%, but after using the proposed method to enhance the image, the correct rate reaches to 52.8%. In terms of classification, the accuracy of the original image is only 84%, and the classification rate after image enhancement has reached 96%, which proves that the enhanced image is of great help to deep learning. The result of this research can effectively avoid the misjudgment of visual inspectors due to long working hours and the problem of visual impairment.

中文摘要 I
AbstractII
致謝 III
目錄 IV
圖目錄 VI
表目錄 VIII
第1章 緒論 1
1.1 簡介 1
1.2 研究背景與動機 1
1.3 相關文獻回顧 2
1.4 硬碟殼沖壓件問題描述 3
1.5 章節概要 4
第2章 表面瑕疵拍攝架構 5
2.1 簡介 5
2.2 燈源選擇 5
2.3 拍攝架構設計 7
2.3.1 燈源角度不變，移動燈源與樣本之間的距離測試 7
2.3.2 燈源角度不變，移動燈源與樣本之間的距離測試 11
2.3.3 燈源角度不變，移動燈源與樣本之間的距離測試 14
2.3.4 硬碟殼沖壓件表面瑕疵最終架構 17
第3章 輪廓毛邊瑕疵拍攝架構 21
3.1 簡介 21
3.2 燈源選擇 21
3.3 拍攝架構設計 24
第4章 圖像增強 26
4.1 簡介 26
4.2 傅立葉域自適應奇異值分解 26
4.2.1 圖像增強 29
第5章 瑕疵檢測與分類 33
5.1 簡介 33
5.2 瑕疵檢測 33
5.3 瑕疵分類 36
5.3.1 Dropout 37
5.3.2 Data augmentation 39
第6章 實驗結果與討論 40
6.1 簡介 40
6.2 檢測以及分類系統規格 40
6.3 深度學習訓練與測試準備 41
6.4 瑕疵檢測結果與討論 43
6.5 瑕疵分類結果與討論 49
第7章 結論與未來工作 52
參考文獻 53

[1]Jing-Wein Wang, Ngoc Tuyen Le, Jiann-Shu Lee, and Chou-Chen Wang, “Color face image enhancement using adaptive singular value decomposition in Fourier domain for face recognition,” Pattern Recognition, vol. 57, pp. 31-49, Sep. 2016
[2]Liu Wei, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, Alexander C. Berg “SSD: Single Shot MultiBox Detector.” Lecture Notes in Computer Science (2016): 21–37. Crossref. Web.
[3]Alex Krizhevsky, Ilya Sutskever, Geoffrey E. Hinton.“ImageNet classification with deep convolutional neural networks,” Advances in Neural Information Processing Systems, vol. 25, no. 2, pp. 1-9, 2012.
[4]Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, Ruslan Salakhutdinov “Dropout: A Simple Way to Prevent Neural Networks from Overfitting,” Journal of Machine Learning Research 15 (2014) 1929-1958
[5]黎玉線，自適應奇異值分解之影像增強應用於人臉識別和指紋分類，國立高雄應用科技大學電子工程系研究所，博士論文，2015。
[6]涂勛城，硬碟殼沖壓件瑕疵檢測系統，國立高雄應用科技大學，光電與通訊工程研究所，碩士論文，2016