YOLOv7-WDD: An Efficient Bi-directional Feature Aggregation Method for Workpiece Defect Detection

The efficiency of workpiece defect detection is closely related to ensuring product quality, production safety, and cost savings, making the improvement of workpiece defect detection efficiency of paramount importance as an indispensable part of industrial production processes. This paper introduces YOLOv7-WDD, an efficient algorithm for defect detection. Firstly, it introduces the Re-parameterized Bi-directional Concatenation (Rep-BiC) Module to merge multi-scale contextual features and reconstructs the Neck into a Re-parameterized Bi-directional Path Aggregation Network (RepBiC-PAN), which alleviates the problem of losing key features in traditional feature fusion structures. Secondly, it incorporates DECA (Deformable Coordinate Attention) to expand the model’s overall shape perception of defects in feature extraction and enhance the expression capability of channel and position features. Lastly, it refines the CIoU loss function used in bounding box regression to improve the detection performance for low-quality samples. Experimental analysis shows that for workpiece surface defect detection tasks using the NEU-DET dataset, the proposed method surpasses other similar algorithms in both accuracy and speed, achieving a mAP of 79.5% and a real-time frame rate of 238.1 FPS. Compared to the baseline YOLOv7, it achieves a 3.1% mAP improvement and a 5.5 FPS increase. The proposed method achieves high accuracy in identifying defects on workpiece surfaces while maintaining sufficient real-time performance.