Research on paint defect detection algorithm based on improved YOLOv5s

The process of detecting and processing images for paint defects generated during the automobile manufacturing production process is more complex. Existing paint defect target detection exists with a large number of parameters, low detection and recognition accuracy, and leakage problems. It is proposed to improve the paint defect target detection algorithm based on the YOLOv5s algorithm. The lightweight backbone MobileNetv2 is utilised to replace its backbone network, and the Conv convolution in MobileNetv2 is replaced with the full-dimensional dynamic convolution (ODConv) to improve the detection accuracy while reducing the computational effort. The FPN+PAN-based neck part introduces the inverted residual mobile module (iRMB), predicts the head structure of small and medium target layers, and convolves the original Convolution with iRMB-Conv. By using this method to improve the network feature extraction ability and reduce the number of parameters, we effectively increase the detection ability of medium- and small-sized targets. The experimental results demonstrate improvements over the original algorithm in terms of detection precision P, recall rate R, average precision mean, and defect-specific type detection precision mAP0.5. The improved model not only ensures a smaller model, but the algorithm also effectively improves detection precision.