Strip steel surface defect detection plays a vital role in industrial quality control. However, the low contrast, multi-scale nature, and blurred boundaries of real-world defects, combined with the real-time demands of industrial environments, pose significant challenges for automatic detection. To address these issues, we propose a novel and efficient end-to-end detection model based on the Mamba architecture. The model adopts an encoder–decoder structure, where a Parallel Local-Global Interaction Module (PLGIM) is introduced in the encoder to fuse local and global features via Local and Global Mamba branches. A learnable class token is used to guide and fuse feature representations. Furthermore, we design a Multi-scale Edge-aware Guidance Module (MEGM), which enhances defect region perception through a multi-branch structure and residual-guided mechanism. Experiments on the real-world ESDIS-SOD dataset demonstrate that the proposed method achieves leading performance across various saliency detection metrics while maintaining a favorable balance between model accuracy, computational cost, and inference speed.