影像處理與瑕疵檢測在社會上的使用越來越廣泛，尤其是在推廣工業4.0的時代裡，不論是什麼產品都需要品質檢測，而本研究所使用的樣品則是過年經常看到的斗方，斗方本身因其本身顏色鮮豔，且使用的墨水常為油墨，導致瑕疵特徵非常容易被圖案和背景色給遮擋。因此本研究設計出了一個影像擷取架構，透過此架構能夠有效改善因光照造成的光線折射問題。而斗方本身難以觀測到的瑕疵特徵，本研究透過梯度陡降的概念，並搭配奇異值分解的技術，研發了一個針對顏色鮮豔影像的影像增強演算法，首先會將影像以RGB三通道影像進行分解，找出三色通道的平均值，透過本算法降低影像內亮麗的色彩並增強特徵的色彩，使瑕疵特徵更容易被觀測到。由於斗方的檢測以前沒人做過，本研究將使用我們所研發的另一個演算法進行比較，主要比較的項目有瑕疵的框選準確率和處理時間。從結果顯示，本研究所研發之演算法的準確率為96 %，處理時間約3秒/張，跟廠商所提供之人工檢測時間約10到15秒左右相比快了最少3倍，由此可證明，本論文所提出的影像強化算法的有效性。

The use of image processing and defect detection is increasingly widespread in society, particularly in the Industry 4.0 era where product quality inspection is critical. This study focuses on doufang, commonly seen during Chinese New Year, known for its vibrant colors and challenging defect detection due to ink and patterns. To address this, we designed an image capture framework to mitigate light refraction issues. We developed an image enhancement algorithm using gradient descent and singular value decomposition (SVD) techniques. This algorithm reduces bright colors and enhances feature colors, making defects easier to spot. Comparing it to another algorithm, our approach achieved a 96% accuracy rate with a processing time of around 3 seconds per image, at least three times faster than manual inspection. This underscores the effectiveness of our image enhancement algorithm.

中文摘要 I
Abstract II
致謝 III
目錄 IV
圖目錄 V
表目錄 VI
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 印刷品瑕疵問題描述 3
第2章 印刷品取像架構 5
2.1 光線於印刷品檢測的影響 5
2.2 圖案於印刷品檢測的影響 6
2.3 影像擷取架構設計 8
第3章 梯度陡降色彩強化演算法 12
3.1 簡介 12
3.2 奇異值分解 13
3.3 梯度陡降演算法應用於色彩強化 16
3.4 梯度陡降色彩強化演算法 25
第4章 實驗設計與結果討論 32
4.1 實驗設計 32
4.1.1 系統規格 35
4.1.2 資料擴充 36
4.1.3 YOLO v5 訓練和測試 37
4.1.4 實驗流程 38
4.2 結果與討論 39
第5章 結論與未來工作 48
參考文獻 49

[1]Jing-Wein Wang, Jiann-Shu Lee, and Wen-Yuan Chen, "Face recognition based on projected color space with lighting compensation," IEEE Signal Processing Letters, vol. 18, no. 10, Oct. 2011.
[2]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, 2016.
[3]Tuyen Ngoc Le, Duong Binh Giap, Jing-Wein Wang, and Chih-Chiang Wang, "Tensor-compensated color face recognition," IEEE Trans. on Information Security and Forensics, vol. 16, pp. 3339-3354, 2021.
[4]Ngoc Tuyen Le, Jing-Wein Wang, Chou-Chen Wang, and Tu N. Nguyen, “Novel framework based on HOSVD for ski goggles defect detection and classification,” Sensors, vol. 19, no. 5538, pp. 1-25, Dec. 14, 2019.
[5]Duong Binh Giap, Tuyebn Ngoc Le, Jing-Wein Wang, and Chia-Nan Wang, "Wavelet subband-based tensor for smartphone physical button inspection," IEEE Access, vol. 9, pp. 107399-107415, Jul. 26, 2021.
[6]Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner, "Gradient-based learning applied to document recognition," in Proceedings of the IEEE, vol. 86, no. 11, pp. 2278-2324, Nov. 1998.
[7]Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun, "Deep residual learning for image recognition," Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770-778, 2016.
[8]Kota Murahira and Akira Taguchi, "Hue-preserving color image enhancement in RGB color space with rich saturation," IEEE International Symposium on Intelligent Signal Processing and Communication Systems (ISPACS 2012) November 4-7, 2012.
[9]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。
[10]阮輝青，高階奇異值光線補償補償用於人臉識別，國立高雄應用科技大學電子工程系研究所，碩士論文，2017。
[11]吳家華，應用自適應能量分析於自動光學檢測，國立高雄科技大學，光電與通訊工程研究所，碩士論文，2017。
[12]王士逢，多重自適應奇異值分解之光線補償技術，國立高雄科技大學，光電與通訊工程研究所，碩士論文，2022。