MCWS-YOLO: A Steel Surface Defect Detection Method Based on State-Space Modeling

Abstract

To address the issues of weak global context modeling, insufficient multi-scale feature representation, and poor feature discrimination under complex backgrounds in existing steel surface defect detection methods, we develop MCWS-YOLO, a Mamba-integrated detection algorithm. First, two state-space feature extraction modules, C3k2-VM and C3k2-VMConv, are designed. C3k2-VM leverages the SS2D selective scanning mechanism to construct a long-range global receptive field with linear complexity, effectively enhancing global context dependency modeling. C3k2-VMConv integrates multi-scale convolutional branches, and while retaining the advantages of global modeling, it fully extracts local texture and edge details of defects, achieving synergistic enhancement of global semantics and local details through a hierarchical differentiated design. Second, to mitigate information loss during downsampling and optimize computational efficiency, weighted convolution (wConv2d) is introduced as the downsampling layer, which enhances feature retention in defect regions through a spatial weight matrix. Finally, a multi-scale fusion module, SPPF-Attn, is constructed by embedding bidirectional spatial attention and multi-scale channel attention into the spatial pyramid pooling structure, achieving the enhancement of key defect features and adaptive differentiated fusion of multi-scale features. Experimental results on the NEU-DET, GC10-DET, and APDDD datasets show that the mAP50 reaches 78.5%, 61.8%, and 59.8%, respectively, which are 2.9%, 2.3%, and 2.3% higher than those of the YOLOv11 baseline model, while the overall computational cost is reduced by approximately 27%. Balancing lightweight design and real-time performance, the proposed method shows strong generalization, enabling effective application in industrial surface defect detection.