MRD-Net: An Effective CNN-Based Segmentation Network for Surface Defect Detection

Surface defect detection is a crucial component in industrial production. Many algorithms based on computer vision have been successfully applied to surface defect detection, but several problems need to be solved in practical implementation. First, surface defects of different products are diverse in shape, size, and location, which pose a challenge to the generalizability and accuracy of the algorithm. Second, the algorithm is required to be real time in practical detection. To address these problems, we propose a multiscale feature enhancement fusion (MFEF) and reverse attention (RA) network for surface defect detection, named MRD-Net, to achieve real-time and end-to-end defect segmentation. The framework first extracts multiscale feature maps from the pretrained MobileNetV2. Then, the MFEF module is proposed to enhance and fuse the feature maps of the deeper layers in the backbone to improve the detection performance. Through several skip connections, the adjacent branches are connected top-down. The RA module is applied in shallow branches to utilize deeper prediction information. Finally, the boundary refinement module is added to refine the object boundary and improve prediction accuracy. Our proposed method requires only a small number of defective samples and achieves a high detection accuracy. The experimental results on three datasets (14 industrial products) show that the proposed method outperforms the three state-of-the-art segmentation methods in terms of generalizability and accuracy, and it also achieves the requirement of real-time detection with a speed of 52 FPS for a
$352\times352$
image.