High-Frequency Dual-Branch Network for Steel Small Defect Detection

Abstract

Strip surface defect detection is pivotal in the steel industry for improving strip production quality. However, there is still a big gap between the existing working and the detection of small defects in strip steel in practical applications. In this paper, we propose the SSD-YOLO model, which is designed specifically for detecting small defects on strip steel surfaces. Given the challenge of feature extraction due to the small defect size, it utilizes a dual-branch feature extraction and channel-level feature fusion to enhance the expression capability of small defects. Moreover, it integrates a multiscale high-resolution detection module to achieve precise segmentation, thereby improving the overall detection accuracy of the model. The experimental results illustrate that the SSD-YOLO model, as proposed, attains a 98.0% mean average precision (mAP) and operates at 66 frames per second (FPS) when evaluated on the SSDD (Steel Small Defect Dataset). In comparison with YoloV8s, the SSD-YOLO achieves a significant improvement in accuracy, with an increase of 19.9%. The inference time and performance of our SSD-YOLO is well balanced, making it suitable for real-world deployment.