Abstract
Defect detection is vital for product quality in industrial production, yet current surface defect detection technologies struggle with diverse defect types and complex backgrounds. The challenge intensifies with multi-scale small targets, leading to significantly reduced detection performance. Therefore, this paper proposes the EPSC-YOLO algorithm to improve the efficiency and accuracy of defect detection. The algorithm first introduces multi-scale attention modules and uses two newly designed pyramid convolutions in the backbone network to better identify multi-scale defects; Secondly, Soft-NMS is introduced to replace traditional NMS, which can reduce information loss and improve multi-target detection accuracy by smoothing and suppressing the scores of overlapping boxes. In addition, a new convolutional attention module, CISBA, is designed to enhance the detection capability of small targets in complex backgrounds. In the end, we validate the effectiveness of EPSC-YOLO on NEU-DET and GC10-DET datasets. The experimental results show that, compared to YOLOv9c, [Formula: see text] increases by 2% and 2.4%, and [Formula: see text] increases by 5.1% and 2.4%, respectively. Meanwhile, EPSC-YOLO demonstrates superior accuracy and significant advantages in real-time detection of surface defects on products compared to algorithms such as YOLOv10 and MSFT-YOLO.