Steel surface defect detection based on the lightweight improved RT-DETR algorithm

The existing object detection algorithms for steel surface defects detection have several drawbacks such as large model size, low efficiency, more parameters, and slow speed which limit their application in production. To solve these problems, in this paper a lightweight improved object detection algorithm, named LRT-DETR, is proposed based on the RT-DETR algorithm. Firstly, the BasicBlock modules in RT-DETR are replaced by the lightweight MobileNetV3 module. As a feature extraction backbone, it can fully capture the long-range feature interactions of steel defects and the local feature interactions. Secondly, a more efficient depthwise separable convolution (DWConv) and the VoVGSCSP structure are introduced into RT-DETR to refine the feature fusion network, which thereby enhances the feature extraction and fusion, as well as reduces the computational complexity of the algorithm and achieves a lightweight effect. Finally, a novel bounding box similarity measure MPDIoU is used to replace the loss function in the original network providing a more accurate measure of the similarity between bounding boxes to improve the accuracy of the model. Experimental results show that an average precision (AP) of the proposed LRT-DETR algorithm is 74.8%, 1.5% higher than the original algorithm, and the number of parameters and computational cost are reduced by 65.90% and 75.96%, respectively. Further experimental results demonstrate that the LRT-DETR algorithm is better than other compared algorithms in terms of detection speed and accuracy. Our algorithm can be used in the real-time and efficient detection of steel surface defects.