EBI-YOLO: A Hardware-Efficient Detector with Cross-Spatial Attention for Steel Surface Defects

In order to cope with the challenges of the current steel surface defect detection algorithms with low recognition accuracy, large number of parameters in the model and huge computational overhead, We propose EBI-YOLO, an improved YOLOv8n algorithm for steel surface defect detection. We introduce the EMA attention mechanism at the end of the backbone network, which effectively enhances feature extraction capabilities. At the same time, a lightweight BiFPN structure is used to replace the Neck structure of the original YOLOv8n to enhance the model's performance in detecting multi-scale defects. The experimental results show that compared with the original YOLOv8n model, the EBI-YOLO proposed in this study improves the mAP50 metrics on the NEU-DET dataset by 6.5%, reduces parameters by 39%, and reduces computational cost (FLOPs) by 46.91%; the mAP50 also improves by 5.2% on the GC10-DET dataset. The model strikes a good balance between detection accuracy and complexity, demonstrating superior generalisation of the algorithm, which meets the requirements of practical industrial deployments and applications.