A real-time defect detection algorithm for steel plate surfaces: the StarNet-GSConv-RetC3 detection transformer (SSR-DETR).

Abstract

In view of the problems in industrial steel plate surface defect detection, such as high model complexity, insufficient recognition of small targets, and inefficient hardware deployment, this study proposes the StarNet-GSConv-RetC3 detection transformer (SSR-DETR) lightweight framework. The framework comprises a StarNet backbone network and an innovative star operation optimization structure to reduce computational complexity while enhancing feature representation capabilities. In the feature fusion stage, the RetBlock CSP bottleneck with 3 convolutions (RetC3) module with an explicit attenuation mechanism is designed to enhance the extraction of geometric features of microscopic defects by combining two-dimensional spatial priors, and grouped spatial convolution (GSConv) is used to optimize the aggregation of multiscale features. Experiments show that the algorithm achieves a mean average precision (mAP) of 88.2% and a classification accuracy of 92.0% on the Northeastern University steel surface defect (NEU-DET) dataset, which is 4.8% and 3.7% higher than the baseline model, respectively. Meanwhile, the model's computational load and size are reduced by 59.5% and 47.8%, respectively. Actual deployment tests show that this algorithm operates at 98.1 frames per second (FPS) on personal computer platforms and at 40.3 FPS, with a single-frame processing time of 24.8 ms, on the RK3568 embedded system, fully meeting the comprehensive requirements of industrial scenarios.