本研究旨在找到一種瑕疵檢測方法，對於計畫委託方工廠紡織的坯布影像具有良 好的瑕疵檢測能力，該胚布資料集具有具有紋理特徵、亮度域偏移及瑕疵稀缺等特性。 在工業異常偵測的相關研究中，許多模型的設計基於受限的實驗環境。例如，常用的 MVTec AD 工業資料集 [1]，這個資料集中包含 15 種不同的物件及紋理類別的圖像， 這些圖像均在相同的拍攝角度、距離及光線條件下拍攝。然而，在工廠實際應用中， 影像會具有多樣的變異性，可能沒辦法固定角度、距離及光線等變數。此時基於受限 實驗環境研發出來的模型，當要應用在實際工廠中時，可能會因為這樣的變數，使得 研究的成果沒辦法完美地轉移到實際應用。待測影像的亮度與模型訓練時的影像不同， 是工廠應用常遇到的域偏移問題。本研究深度探討胚布紋理資料集中亮度分佈偏移的 解決方案，提出了一種針對胚布紋理以及亮度域偏移影像的異常偵測方法，達到最先 進的的異常偵測能力。;In the field of industrial anomaly detection, many models are designed based on con- trolled experimental environments. For instance, the widely used MVTec AD industrial dataset includes images of 15 different object and texture categories.[1] However, these images share the same shooting angle, distance, and stable lighting conditions. In real factory applications, images exhibit a greater variety of variations, and models developed in restricted experimental settings may not perfectly transfer to practical applications due to this variability. One common distribution shift problem encountered in factory settings is the difference in brightness between the test images and the training data, as seen in the greige fabric images at factory. Therefore, this study aims to find a greige fabric defect detection method that maintains high detection performance even when there is a bright- ness distribution shift in the test images. We have developed a semi-supervised anomaly detection method that can handle brightness domain shifts. This image gradient-based semi-supervised anomaly detection approach has shown effective results in detecting de- fects in greige fabric and successfully addresses the issue of brightness distribution shift, achieving state-of-the-art anomaly detection capabilities.
