Super-resolution-guided dense enhancement segmentation network for defect detection of strip steel

Abstract Strip steel defect segmentation is critical for quality control in production. Deep learning-enabled automatic segmentation methods have advanced in recent years, but still struggle with segmentation tasks under low-resolution image conditions. To address the challenge, we propose a Super-Resolution (SR)-guided Dense Enhancement Network for segmenting strip steel defects in high-resolution SR space. Specifically, we propose a dual-branch joint learning architecture that includes compound SR and semantic segmentation branches, which simultaneously learn high-fidelity SR images and semantic hybrid features to embed SR fine-grained features across different scale spaces. To ensure efficiency and overcome the limitations of industrial computing resources, we design a hierarchical dense encoder and share dual-branch features in the low-level space. Additionally, we integrate an efficient Transformer into the capsule network and incorporate it into the segmentation branch to establish global feature dependencies, enhancing segmentation performance for categories with high local similarity. Finally, we employ a feature fusion module in the skip connection to balance deep semantic and spatial information within the network. We conduct extensive experiments and the results show that the proposed method outperforms the representative alternatives in detecting eight common defects, highlighting its potential as a reliable and efficient approach for metal surface defect detection. In order to verify the generalization performance of the model, experiments are carried out on another classical dataset SD-saliency-900 and a benchmark dataset mobile phone screen defect dataset, which shows that it also has superior performance in different scenarios. The source code is available at https://github.com/CDUTJ102/SDE-Net .