Abstract
Using deep learning-based methods to detect surface defects in strip steel can reduce the impact of human factors and lower costs while maintaining accuracy and efficiency. However, the main disadvantages of this method is the inability to tradeoff accuracy and efficiency. In addition, the low proportion of valid information and the lack of distinctive features result in a high rate of missed detection of small objects. In this paper, we propose a lightweight YOLOX surface defect detection network and introduce the Multi-scale Feature Fusion Attention Module (MFFAM). Lightweight CSP structures are used to optimize the backbone of the original network. MFFAM uses different scales of receptive fields for feature maps of different resolutions, after which features are fused and passed into the spatial and channel attention modules in parallel. Experimental results show that lightweight CSP structures can improve the detection frame rate without compromising accuracy. MFFAM can significantly improve the detection accuracy of small objects. Compared with the initial YOLOX, the mAP and FPS were 81.21% and 82.87Hz, respectively, which was an improvement of 4.29% and 12.72Hz. Compared with existing methods, the proposed model has superior performance and practicality, verifying the effectiveness of the optimization method.