A Steel Surface Defect Detection Algorithm Based on YOLOv7-Tiny

In the modern manufacturing industry, the surface defect detection of steel is very important to ensure product quality. Based on YOLOv7-tiny algorithm, this study improves the real-time and accuracy of steel surface defect detection through a series of innovative improvements. First, we replace the original ordinary convolution structure by introducing Omni-dimensional dynamic convolution module, enhance feature extraction to boost model detection accuracy. Change the activation function from Leaky ReLU to SiLU, and enhance the gradient flow of the model, improve detection performance. In addition, by introducing SimAM module and improving the performance of detection headers through TSCODE, the model's ability to recognize small size targets is enhanced. After that, using MPDIoU instead of CIoU as the loss function, which simplified the calculation process and the efficiency and accuracy of defect detection are improved. Finally, we verified the model through the NEU-DET data set. The experimental results show that, with these improvements, the mAP@0.5 of the model is significantly improved, reaching 0.768 in the end, and the reasoning speed is also up to 34.5 fps, which proves the speed and accuracy of the improved model in real-time detection of steel surface defects.