Multi-scale based defect detection for automotive glass

Abstract

Small targets have low resolution and small size. Using the YOLOv3 model may not detect the target object. In view of this situation, this paper improves the YOLOv3 model and proposes a network structure with refined features. The channel feature refinement mechanism is introduced, and four detection scales are designed to prevent small objects from being submerged; and the spatial pyramid structure-Spp module is added to detect five defects of glass. The experimental results show that compared with the YOLOv3 model, the mAP value of the optimized model is 97.22%, and the mAP value of the original version is 92.31%, an increase of 4.91%. In the self-built data set, the accuracy of automotive glass defect recognition reached 97.66%, and the optimized model’s recognition effect on small targets was significantly improved. Compared with the traditional glass defect detection method, the method improves the detection accuracy and the recognition accuracy.