EMSD-YOLO: A Surface Defect Detection Model Based on Feature Extraction and Fusion

Accurate identification of industrial defects plays a critical role in maintaining product quality during manufacturing. Nevertheless, the wide range and intricate nature of defect types present major obstacles for current detection models. To tackle these issues, we introduce a new model named EMSD-YOLO. This framework incorporates a Selective Feature Fusion Pyramid Network (SFFPN), which enhances the detection of surface defects through multi-scale feature integration. Furthermore, it employs Hierarchical Dual-Stream Convolutions (HDSConv) to boost computational performance. A comprehensive comparison across three widely used datasets-NEU-DET, GC10-DET, and Tianchi-confirms the advantages of EMSD-YOLO. Specifically, it achieved top-tier results on the NEU-DET dataset, recording a leading mAP50 score of 0.783 and an F1 score of 0.81, surpassing other lightweight models such as YOLOv10-S and YOLOv11-S in both accuracy and processing speed. On the GC10-DET and Tianchi benchmarks, EMSD-YOLO also exhibited strong generalization and robustness, effectively tackling real-world complexities. Overall, EMSD-YOLO bridges critical gaps in precision, speed, and flexibility, positioning itself as an advanced approach to surface defect detection.