SEDR-Net: A YOLOv11-Based Network for Conveyor Belt Surface Defect Detection in Complex Industrial Scenes

Belt conveyors are essential to continuous material transport systems, and reliable surface defect detection is therefore critical for safe and stable operation. In real industrial environments, defects such as tears, punctures, and localized damage are often small, elongated, and characterized by weak boundary contrast. Complex background interference further increases the difficulty of accurate and reliable detection for real-time defect detectors. To address these challenges, this paper proposes SEDR-Net (Structure-Edge and Detail Reconstruction Network), a YOLOv11n-based network, for this task. The Structural-Edge Fusion Block (SEFBlock), Channel-Spatial Collaborative Attention (CSCA), and Efficient Up-Convolution Block (EUCB) respectively enhance structural-edge representation, suppress redundant background responses, and recover local structures and boundary details. On a public conveyor-belt defect dataset, SEDR-Net achieves 90.8% Recall, with an mAP@0.5 of 92.4% and an mAP@0.5:0.95 of 58.4%, yielding improvements of 4.7, 3.8, and 7.1 percentage points over YOLOv11n, respectively. Meanwhile, SEDR-Net uses 2.42 M trainable parameters and maintains an inference speed of 134.2 FPS, indicating a favorable accuracy–complexity trade-off for real-time inspection. An independent external industrial test set further verifies the cross-scenario robustness and practical applicability of the proposed method under real mining conveyor-belt conditions.