摘要
在工業的快速發展下，有許多的工業產品講究更高的精確度，又因為近幾十年來深度學習的發展更是各個產業的主流，瑕疵檢測結合深度學習成為了主要的未來趨勢。 但對瑕疵檢測而言，瑕疵資料樣本並不容易取得，而本論文藉由CycleGAN模型架構，修改其生成器模型及判別器模型。可以使模型僅藉由無瑕疵樣本作為訓練資料，也不需要做標籤動作就能達到瑕疵檢測的非監督式學習模型。建構一個可以將瑕疵影像修復的生成模型，再藉由將原始輸入影像與生成影像作比對區分出影像是否有瑕疵。 最後，實驗結果顯示，本論文所設計的模型架構是有效的。並且利用影像相似度的比較，驗證實驗的三組類別的準確率結果都有達到八成以上。
並列摘要
With the rapid development of industry, many industrial products pay attention to higher accuracy, and since the development of deep learning has become the mainstream of various industries in recent decades, defect detection combined with deep learning has become the main trend in the future. However, it is not easy to obtain defect data samples, so this thesis modified its generator model and discriminant model by using the CycleGAN model architecture. An unsupervised learning model can only achieve defect detection by using flawless samples as training materials and without the need for tag actions. A generation model can be built to repair the defective image, and then compare the original input image with the generated image to distinguish whether the image has defects or not. Finally, the experimental results show that the model architecture designed in this thesis is effective. Moreover, by comparing the image similarity, the accuracy of the three groups of the verification experiment has reached more than 80%.
參考文獻
延伸閱讀
- 陳彥良（2015）。基於異構計算系統之高效能自動表面瑕疵檢測〔碩士論文，中原大學〕。華藝線上圖書館。https://doi.org/10.6840/cycu201500331
- Lo, K. C. (2016). 非監督式特徵學習與多重解析度直方圖池化 [master's thesis, National Chiao Tung University]. Airiti Library. https://www.airitilibrary.com/Article/Detail?DocID=U0030-0803201714331820
- 錢鉦津（2012）。監督式學習型智能軟體的測試評估方法。品質月刊，48(11)，18-23。https://doi.org/10.29999/QM.201211.0005
- 雷承維（2020）。Unsupervised Deep Learning System for AOI Detection〔碩士論文，國立臺灣師範大學〕。華藝線上圖書館。https://doi.org/10.6345/NTNU202001296
- 陳昱任（2020）。An Anomaly Detection System Based on Unsupervised Learning in IoT Environment〔碩士論文，國立成功大學〕。華藝線上圖書館。https://www.airitilibrary.com/Article/Detail?DocID=U0026-0006202200000087