A lightweight small-defect detection algorithm for cold-rolled strip surfaces with high precision and efficiency

In order to address the issue of the discrepancy between the accuracy and speed of defect detection on the surface of cold-rolled steel strips, a lightweight algorithm for the identification of small target defects, designated Yolo-DS, is proposed. Firstly, an NMS-free multi-detector distillation training method is designed, which significantly reduces the inference latency and improves the detection accuracy through the joint optimisation of a pair of single and a pair of multiple heads. Secondly, a Sparse Channel Self-Attention (SCSA) module is introduced with the objective of achieving efficient global feature modelling through a reduction in the self-attention computations. Furthermore, the proposed Multilevel Feature Extraction Attention (MFEA) module enhances the model's capacity to extract nuanced defect details and structural information through a multilevel attention mechanism. The experimental results demonstrate that Yolo-DS attains a mAP of 92.8% on the self-constructed M-CRSD (Micro Cold Rolled Strip Defect) dataset with an inference speed of 118FPS and a latency of 2.2ms, thereby markedly enhancing the detection efficiency. It offers an optimal solution for real industrial production.