MDRNet: efficient surface defect segmentation network based on CNN and Transformer

In industrial production, accurate segmentation of surface defects is crucial for ensuring product quality. However, the variability in defect size and shape, along with low contrast and inter-category similarity, limits the effectiveness of existing segmentation models in complex environments. These models often struggle with scale and shape variations, experience significant performance decline under lighting changes or low-contrast conditions, and lack sufficient category discrimination. Thus, precise surface defect segmentation remains a significant challenge. To address this, we propose MDRNet, an efficient surface defect segmentation network that integrates the advantages of CNNs and Transformers. First, we integrate multi-scale channel attention (MSCA) with the Transformer to design an efficient MSCA-Transformer module for feature extraction. This module focuses on local information while effectively establishing long-range dependencies, ensuring accurate multi-scale defect representation under low contrast. Second, we adopt Dense Connected Shuffle Group Convolution to efficiently integrate critical information from deep and shallow layers. Finally, we introduce a novel Region Attention module that highlights important local features while suppressing redundancy, thereby enhancing the ability to capture inter-category differences. MDRNet achieves the best segmentation performance on four public industrial surface defect datasets. Even in challenging scenarios with low contrast and large defect scale variations, MDRNet robustly delivers precise predictions of defect locations, categories, and contours, fully demonstrating its efficacy.