Defect detection algorithm for steel surface based on improved YOLOv5

Steel is frequently used in manufacturing equipment, and detecting steel surface flaws is critical to the proper operation of steel equipment in manufacturing workshops. In the process of industrial steel defect detection, there are problems such as low detection accuracy of multi-category defects, low detection speed and high missed detection rate of small targets. An improved YOLOv5 algorithm has been proposed to address the aforementioned problems. This algorithm adds the SE attention module to the backbone network of YOLOv5, which improves the feature representation ability. The bidirectional feature pyramid network (BIFPN) is then utilized to increase the model's feature fusion capabilities and detection accuracy. Finally, a small object detection layer is introduced to the model prediction layer for small objects in the data set to lower the missed detection rate of small objects. The algorithm proposed in this research increases object detection accuracy through experimental verification on the NEU-DET data set, when compared to a variety of object detection algorithms. Compared with the original model of YOLOv5s, the speed is increased by three frames, and the mAP value is increased by 5.1%.