Surface Defect Detection With Channel-Spatial Attention Modules and Bi-Directional Feature Pyramid

The YOLOv5 network architecture prioritizes speed and efficiency, but this may limit its ability to capture intricate details of complex objects. To solve the problems of insufficient feature extraction ability and incomplete feature fusion in the YOLOv5 single-stage detection network, we propose a YOLOv5 algorithm based on an improved bidirectional feature pyramid network (BiFPN). The FPN is modified into BiFPN with recursive feature fusion, bidirectional connections, and alignment of multi-scale features. It can utilize different levels of feature information to improve the ability of feature expression, thereby improving model detection accuracy and efficiency. This method incorporates CBAM attention mechanism, which can adaptively adjust the feature map while considering both channel and spatial features, thus more comprehensively expressing information in the feature map. For small targets, due to their small size, they often have high similarity with background and unclear features that make it difficult for traditional loss functions to optimize accurately. Using SIoU as a loss function can improve the recognition rate for small object. The experimental results show that our strategy significantly improves the performance of YOLOv5 on the NEU-DET dataset, reaching 77.5% mAP at 92 Frame Per Second(FPS), which is 3.1% higher than unimproved YOLOv5 network and 19.6% higher than SSD algorithm. Moreover, F1 score has also been improved while showing strong generalization ability on GC-DET dataset as well. Its detection speed is also superior to other algorithms indicating that this improved algorithm can quickly and efficiently detect surface defects in steel materials while significantly improving detection performance.