Precisely segmenting surface defects in steel strips is essential for enhancing product quality. Despite the potential improvements in defect segmentation accuracy and robustness offered by deep learning methods, including autoencoder, the following challenges still persist. Firstly, in industrial environments characterized by low-contrast defects against the background and high noise-to-signal ratios, current defect detection methods still face challenges inaccurately segmenting defects. Secondly, in industrial production, defects often follow a long-tail distribution, current defect detection methods exhibit lower accuracy in identifying defects in the tail- end categories. To tackle these challenges, a novel two-stage encoder multi-decoder network was introduced, comprising an initial defect detection stage and a category-specific refined stage. In the initial defect detection stage, the network's decoder employs global-local up-sampling modules to utilize deconvolution of multiple receptive fields for up-sampling feature maps. Subsequently, in the category-specific refined stage, the network initially separates the defect feature maps by employing a category separation module. It integrates prior information through a defect refinement module and a fusion module, fusing prior decoder features with corresponding one. Simulation experiments were conducted using the real-world strip steel defect dataset, and validation experiments were performed on the industrial imbalanced dataset collected from an actual project. The experimental results demonstrate the proposed method reliability in industrial production, with the segmentation mean intersection over union achieving 87.35% and 84.98% on these respective datasets.
