MSIDetector: Detecting Multi-Scenario industrial defects using an adapted visual foundation model and dual thresholding discriminator

Industrial defect detection is crucial for enhancing product quality. The diverse nature of industrial scenarios has posed challenges for developing a unified defect detection model that can address multiple industrial scenarios. To meet this challenge, we started from a data perspective, developing three self-designed data acquisition apparatuses to capture defect samples from multiple industrial scenarios. Leveraging grounding DINO (GDINO), an open-set detection approach, we propose a unified defect detection method, MSIDetector. By integrating an industrial feature adapter and a context-aware dual-thresholding defect discriminator, we successfully incorporate prior industrial knowledge into the model. The performance of the MSIDetector was evaluated across four industrial scenarios, including a public defect dataset, achieving a total state-of-the-art mAP@50 of 76.2, surpassing the performance of FasterRCNN, YOLO-7X, and the visual foundation model GDINO, with quantitative increases of 13.93, 6.55, and 2.1, respectively. This report represents the first successful attempt to apply foundation models to defect detection.