LSEH-YOLO: a lightweight network for strip steel surface defect detection based on hybrid feature fusion and attention-guided head

Abstract Strip steel surface defect detection poses significant challenges due to small object omission, blurred features, and drastic scale variations, especially under the strict industrial demands of high accuracy, real-time performance, and lightweight deployment. To address these issues, we propose LSEH-YOLO, a lightweight detection framework integrating hybrid feature fusion and attention-guided mechanisms, with enhancements in backbone, attention, neck, and detection head.Specifically, the C3k2_LFEGM module introduces a three-stage design—local enhancement, gated fusion, and residual feedback—by integrating edge perception and window-level attention to boost feature response. The C2PSA-Mona module embeds Mona non-local guidance into multi-head attention to enhance inter-channel modeling. For the head, the Detect_AS module combines AFGC lightweight attention and directional Strip convolution to refine fine-grained localization. Furthermore, we propose the HB Neck, which reconfigures YOLOv11’s neck with a bidirectional feature flow, multi-strategy fusion (Fusion), lightweight upsampling (EUCB), and a structurally enhanced C3k2_SEH module.Experiments on NEU-DET and GC-DET datasets show that LSEH-YOLO improves mAP by 4.0% and 6.1% over the YOLOv11 baseline, while reducing parameters by 32.9%, FLOPs by 18.8%, and achieving 100.75 FPS. In the generalization evaluation on the GC-DET dataset, the model exhibited excellent accuracy and strong generalization performance.These results demonstrate that LSEH-YOLO effectively balances accuracy, speed, and efficiency, providing a practical solution for multi-scale defect detection in industrial settings.