DA-SAM: A Defect-Aware Segmentation Model for Industrial Product Inspection

The standard procedures for detecting industrial product defects typically involve collecting images, labeling defects, and training models through supervised learning. This process is often time-consuming and laborintensive. To address these challenges, we introduce the Defect-Aware Segment Anything Model (DA-SAM), a defect detection tool designed to simplify the conversion of box labels into segmentation annotations, thereby reducing labeling time and enabling efficient zero- or few-shot learning. First, we fine-tune the SAM model using LoRA and fine-grained bounding boxes to generate defect masks. The model is then trained on several public datasets. A two-stage training approach distinguishes defects from the background and addresses defect variations. Experimental results demonstrate that DA-SAM performs effectively on both the MVTecAD public dataset and the self-collected SPTD dataset.