CRDE-net: improved multi-scale detection network for strip steel surface defects based on YOLOV5

Surface defects in strip steel can significantly impact its quality and performance. Therefore, this paper proposes an improved strip steel surface defects detection network model (CRDE-Net) based on the Yolov5 model. First, we designed an efficient aggregation network module using the idea of VOVNet to deepen the network's number of layers and improve the ability of network feature extraction. Secondly, we introduce Res2Net convolution to increase the multi-scale feature detection capability of the network, introduce DSConv convolution to replace the 3x3 convolution in the model, reduce the amount of calculation and parameters of the network, and reduce memory usage. In addition, adding the ECA attention mechanism to the model makes the network pay more attention to defect features, thereby improving detection accuracy and speed. Finally, CoordConv convolutions are introduced in the neck network to make the network spatially aware. Extensive experiments on the NEU-DET dataset show that the average accuracy of our algorithm reaches 83.7% mAP, compared with the baseline network, our accuracy is increased by 5.9%, and the number of computational volumes is reduced by 22.2%, which fully proves that our algorithm it has good detection performance.