RMP-DETR: a lightweight multi-scale detector with tri-feature dynamic perception for hot-rolled steel strip surface defects

Detecting surface defects in hot-rolled steel strip is vital for ensuring engineering quality control. However, current hot-rolled steel strip surface defect detection faces challenges such as the difficulty of balancing algorithm lightweightness and detection accuracy, as well as insufficient capability in extracting complex defect features. To solve the above problems, this paper proposes a lightweight hot-rolled steel strip surface defect detection algorithm RMP-DETR based on improved RT-DETR, aiming to achieve efficient and accurate multi-scale defect detection. First, the approach includes a reparameterized dual-stream heterogeneous network (RDHNet), which substantially lowers computational overhead while preserving detection accuracy. Secondly, a multi-scale dynamic gated block fusion (MDGS) is developed to dynamically weight multi-scale features, optimising the feature fusion process. Finally, a progressive blockwise spatial fusion (PBSF) module is introduced to further enhance the detection capability of adjacent defect areas and reduce feature confusion. Evaluations conducted on the NEU-DET dataset demonstrate a 2.7% rise in detection accuracy (mAP@0.5) for RMP-DETR compared to the baseline, with reductions of 47.2% and 47.5% in parameters and computational load (GFLOPs), respectively. Furthermore, the algorithm’s generalisation capability was validated, further illustrating its robustness and superior detection performance, making it more suitable for practical deployment and application.