Defect Detection using DINO-ViT and Optimal Threshold Determination with Synthetic Data

This thesis proposes a defect detection model based on the PatchCore frame- work, combining with DINO-ViT as the feature extractor. The proposed method is mainly designed for scenarios lacking defective training data. DINO-ViT can extract highly semantic information and high-resolution feature maps, exhibiting strong generalization and applicability across various industrial defect detection datasets. It surpasses ResNet-based feature extractors and other state-of-the-art (SOTA) defect detection methods. In defect detection tasks, defining a threshold to detect defective pixel regions is often required, which typically relies on exten sive experience. This study proposes a method to determine the optimal threshold for converting image heatmaps into binary segmentation maps, thereby enhancing the practicality of the defect detection system. We demonstrate the effectiveness of the proposed model across multiple public datasets (MVTec AD, VisA, BTAD, MTD) as well as datasets collected from the production lines in Delta Electronics. The results show that our method has outstanding performance in defect detection accuracy, whether measured by image-level AUROC or pixel-level AUROC, even when the training dataset lacks defective samples and defects only appear in the test dataset. The results also validate the effectiveness of DINO-ViT in providing precise predictions and clear defect boundaries. Additionally, we used synthetic datasets to assist in finding the optimal threshold and explored the impact of hyperparameters for generating synthetic data on defect detection results.