MC-DDet: Lightweight Mini-Channel Detector for Steel Strip Surface Defect Detection

This paper addresses the persistent challenge of achieving efficient and accurate surface defect detection in steel strips by proposing a lightweight mini-channel detector (MC-DDet). First, a mini-channel feature extraction network (MCNet) is designed around a mini-channel feature extraction (MCFE) module. This module iteratively extracts features within a narrow channel range, refining the feature representation. Second, a lightweight feature enhancement (LFE) module is constructed by integrating squeeze-and-enhance axis attention (Sea) and a grouped feedforward network (GFFNet). This LFE module enhances the detector’s sensitivity to both global and local features, enabling the capture of intricate characteristics. Third, to improve the selection of high-quality anchor boxes and mitigate the foreground-background class imbalance, Focal-EIoU loss and Focal loss are adopted. Experimental results on the NEU-DET dataset demonstrate that MC-DDet achieves 80.0% mAP, outperforming the baseline while simultaneously reducing the parameter count by 40.0% and the computational load by 50.6%. Furthermore, evaluation on the GC10-DET dataset yields an accuracy of 71.8% mAP, highlighting the potential for broad application of MC-DDet in industrial settings. The implementation and related resources are available at: https://github.com/hanxinif/MC-DDet.