Surface Defect Detection of Strip-Steel Based on an Improved PP-YOLOE-m Detection Network

Surface-defect detection is crucial for assuring the quality of strip-steel manufacturing. Strip-steel surface-defect requires defect classification and precision localization, which a challenge in real-world applications. In this research, we propose an improved PP-YOLOE-m network detecting surface defects. First, data augmentation performed to avoid overfitting problem improve model’s capacity generalization. Secondly, Coordinate Attention embedded CSPRes structure backbone network’s feature extraction capabilities obtain more spatial location information. Thirdly, Spatial Pyramid Pooling specifically replaced Atrous neck network, enabling multi-scale broaden its receptive field gain information globally. Finally, SIoU loss function accurately calculates regression over GIoU. Experimental results show that AP, AP50, AP75, respectively, achieved 44.6%, 80.3%, 45.3% defects on NEU-DET dataset by 2.2%, 4.3%, 4.6% network. Further, our method has fast real-time can run at 95 FPS single Tesla V100 GPU.