Small target defect detection on steel surface based on multi-scale feature fusion

Aiming at the problem of insufficient detection accuracy for small target defects on steel surfaces, which often leads to missed detections and false alarms, this paper improves upon the YOLOv8n algorithm to propose a more precise and efficient algorithm named HM-YOLO. To enhance the model’s sensitivity towards small-size defects, an additional tiny object detection layer is incorporated into the model, significantly improving the feature representation capability for small objects. However, this enhancement may also amplify noise and relatively lack high-level semantic information, thereby limiting its overall detection performance improvement. To fully exploit the potential of the tiny object detection layer, we introduce the MPDIoU loss function, which helps optimize boundaries between different defect types. Additionally, a novel multi-scale feature fusion method called Channel-Spatial High-screening Feature Pyramid Network (CSH-FPN) is proposed. This network uses high-level features as weights, adjusting the importance of different channels and spatial locations via the CBAM attention mechanism. It fuses filtered low-level feature information with high-level features, thus enhancing the model’s feature expression ability. Compared to the original YOLOv8n model, HM-YOLO achieves a 6.83% increase in mAP@0.5 while reducing the number of parameters by 33.89%. When compared to the YOLOv11n model, HM-YOLO shows a 3.23% improvement in mAP@0.5 and a reduction of 23.17% in the number of parameters.