A surface defect detection model for cold-rolled steel strips based on lightweight multi-scale contextual real-time detection transformer

Surface defects in cold-rolled steel strips, characterised by small sizes, diverse shapes and similar types, pose significant challenges to product quality control in high-speed rolling processes. To achieve real-time detection of such defects, this paper presents a novel Lightweight Multi-Scale Contextual Real-Time Detection Transformer (LMC-RTDETR) model. Initially, we designed a lightweight backbone network – Partial Channel Residual Fusion Network (PCRFNetwork) – to reduce model parameters and enhance feature extraction capability. Subsequently, an efficient encoder was developed, comprising two key modules: the Mixed Local and Multi-scale Attention module (MLMSA) and the Context-based Feature Fusion module (CAFF). The model’s feature representation capability is enhanced by the MLMSA module, while the CAFF module enhances the model’s capability to process contextual information across scales. Experimental results on a self-constructed cold-rolled steel defect dataset show that the presented model achieved a mean average precision (mAP) of 81.1% and a frame rate (FPS) of 163.4. Compared with the baseline Real-Time Detection Transformer (RT-DETR) model, it reduced GFLOPs by 13.5%, decreased model parameters by 34.5%, increased FPS by 1.2 and improved mAP by 4.2%. These findings demonstrate that the proposed LMC-RTDETR model enables the detection of cold-rolled steel strip surface defects efficiently and accurately, while meeting the requirements of industrial production environments.