SMC-YOLO: Surface Defect Detection of PCB Based on Multi-Scale Features and Dual Loss Functions

作者：Weibin Kong, Zhifei Zhang, Ting-Lin Zhang, Lei Wang, Zi-Yao Cheng, Mo Zhou · 发表于：IEEE Access · 年份：2024 · DOI：10.1109/access.2024.3434559 · 被引用次数：19 · 研究领域：Industrial Vision Systems and Defect Detection、Advanced Neural Network Applications、Infrastructure Maintenance and Monitoring

The detection of surface defects on printed circuit board (PCB) plays a vital role. However, current defect detection methods face significant challenges, such as frequently misidentifying non-defective areas as defects, low defect recognition capability, and diminished accuracy in identifying minor defects. To address these challenges, a novel defect detection network based on the YOLOv7-tiny network framework is proposed (SMC-YOLO). In this paper, we have improved the (Spatial Pyramid Pooling, Cross Stage Partial Channel) SPPCSPC module by adding three additional convolutional layers and introducing an extra feature channel below the max-pooling layer. These enhancements help in capturing fine-grained features and preserving crucial information, which is essential for detecting tiny defects on PCB. Additionally, we have modified the connection method between the max-pooling layers, changing from a parallel to a serial connection to reduce information loss. On this basis, incorporating the multi-head self-attention (MHSA) mechanism at the SPPCSPC module output enhances the network’s ability to focus on critical features across various scales, effectively capturing and utilizing detailed information, thereby improving the detection performance for small defects. Furthermore, the introduction of the content-aware reassembly of features (CARAFE) lightweight upsampling operator restores fine-grained details like defect edges and textures. It also preserves contextual information...