A Spatial–DCT Feature Fusion Network for Copper Strips and Plates Surface Defect Segmentation

Instance segmentation of surface defects is one of the research hotspots in the field of image segmentation. Due to limitations such as restricted receptive fields or the loss of fine-grained details, traditional neural network models still struggle to achieve sufficiently high-segmentation accuracy for surface defects. To meet the demand for high precision segmentation of surface defects on copper strips and plates in industrial quality inspection, this paper proposes a feature fusion segmentation network, termed DSFFNet. First, a dual-branch structure is designed in DSFFNet to fuse spatial-domain features with discrete cosine transform (DCT)-domain features, thereby obtaining richer feature information. Second, a 2D-DCT frequency feature extraction module is developed to more effectively capture the edge information of targets. Third, a triplet attention mechanism is introduced into the backbone network to form an attention-centric network. Finally, a bidirectional fusion module and a multi-scale fusion network are designed to capture finer-grained feature information. Comparative experiments conducted on the KUST-SEG-Dataset demonstrate that DSFFNet achieves 94.66% ± 1.07% (mask)mAP50 and 95.38% ± 0.06% (box)mAP50, outperforming several classic image segmentation methods. Furthermore, generalization experiments on the public NEU-Seg dataset yield a (mask)mAP50 of 86.27% ± 0.01%. The generalization results indicate that DSFFNet is robust to datasets with similar defect types.