YOLO-SDD: A Scattered Defect Detection Model for Metal Plate Surface Based on Improved YOLO

Aiming at the existing metal plate surface scattered defect detection model with complex structure and poor detection effect of small targets, the paper proposes an improved YOLOv5 metal plate surface scattered defect detection model, YOLO-SDD. The YOLOv5 model is improved by CBAM attention mechanism to reduce the interference of irrelevant background information from both channel attention and spatial attention. A C2f module is substituted for the C3 module in model’s backbone structure to enhance the model’s ability to detect small targets. The activation function is modified to improve the accuracy and recall of model detection based on solving the gradient disappears problem. For the evaluation of scattered defect information, the local images of defects are sequentially subjected to grayscale transformation, Gaussian filter denoising, inverse binarization and defect contour annotation. The information such as the area of defective region, the location of defects and the number of flaws per square centimeter are calculated to meet the visualization needs of industrial production. Furthermore, the experimental results illustrate the accuracy of the proposed model reach $\mathbf{9 2. 8 \%}$ on the NEU-DET and 74.3% on the TCA-DET, and improve by 1.4% and 12.2% over the original model, indicating that the improved model, YOLO-SDD, has a strong universal applicability.