MFAM-Net:A Surface Defect Detection Network for Strip Steel via Multiscale Feature Fusion and Attention Mechanism

Detecting defects on the surface of steel strip is a critical step in the production field of the steel industry. However, due to the diversity of defect types, scales, texture structures and the irregular distribution of defects in strip steel images, existing detection methods are difficult to achieve accurate detection. To this end, we propose a steel strip surface defect detection network (MFAM-Net) based on an YOLOv5 model. Firstly, we design a multi-scale feature extraction module (MSC3) to capture more complete and comprehensive defect texture features for strip defect objects with large scale changes. Secondly, we add the coordinate attention module (CA) to the network, while the spatial pyramid pooling structure (SPPF) of the backbone network is also replaced to increase the network's ability to locate defect features. Third, we introduce a new detection branch to improve the network's detection performance for faults of various sizes. Finally, we introduce the Wise-IoU (WIoU) loss function to focus the detection network on ordinary quality anchor boxes, and enhance the network's regression capability for defect positions. Experimental results show that on the NEU-DET and GC10-DET datasets, our method achieves an mean average precision of 83.5% and 66.7%, respectively, which is 7% and 2.9% higher than the baseline network YOLOv5, and also outperforms other defect detection models.