Improved DRAEM: Enhance the Unsupervised AD in Defect Segmentation

Surface defect detection is a critical aspect of the manufacturing process, requiring accurate detection methods to ensure product quality. Traditional fully-supervised detection methods require extensive annotated data, leading to significant labeling costs. This study introduces a novel self-supervised surface defect detection method using synthetic pseudo defect samples, achieving pixel-level detection with only normal samples for training. Adopting a structure similar to DRAEM, this method concatenates the generative and discriminative networks, using both the input and output of the generative model to enhance the discriminative network's decision boundaries. Introducing a multi-head attention mechanism and a Group Aggregation Bridge (GAB) module for feature fusion significantly boosts the discriminative network's performance and segmentation accuracy. Extensive comparative experiments with a variety of similar anomaly detection algorithms have demonstrated the superior segmentation performance of this work.