Cross Stage Partial Modulated Convolution Net for Leather Defect Detection

The precise, dependable, and smart examination of leather surface defect has emerged as a significant topic in the field of industrial inspection. To further improve the accuracy of leather defect identification, a cross-stage partial modulated convolution net (CSPMNet) is proposed, which is more lightweight and efficient while seeking to improve the computational performance. CSPNet is used as the backbone of the CSPMNet, whose convolutional blocks is replaced with convolution modulation modules, combining transformer with traditional convolutional techniques. Experimental findings indicate that the model in question attained an 86.33% Top-1 accuracy on our specially designed challenging data set for detecting defects on leather surfaces, which is superior to CSPDarkNet-53, CSPResNet-50, Res2Net-50, ResNeXt-50, EfficientNet-B3, RepVgg-A2, and DenseNet-201. The proposed CSPMNet also has good generalization ability. In the performance testing of the public dataset Cifar10, the Top-1 reached 86.19%, which also is better than the aforementioned comparative models. Our work provides a excellent solution for leather surface defect detection.