EMamba: A MoE-enhanced state space network for robust steel surface defect detection

Steel surface defect detection is a critical component of intelligent manufacturing and industrial visual quality control. However, due to large variations in defect scale, elongated morphologies, strong directional continuity, and high inter-class visual similarity, existing detection methods struggle to achieve a desirable balance among global structural modeling, semantically adaptive representation, and computational efficiency. To address these challenges, this paper proposes an efficient steel surface defect detection framework, termed EMamba. By integrating state space modeling, conditional expert computation, and a hybrid multi-branch architectural design, the proposed framework systematically enhances detection performance in complex industrial scenarios. Specifically, we introduce a two-dimensional state space modeling mechanism and propose a Convolutional State Space Block (CSSBlock), which models long-range spatial dependencies and directional continuity with linear computational complexity. At the high-level semantic stage of the backbone network, an Efficient Sparse Mixture-of-Experts module (ES-MoE) is incorporated. Through a conditional Top-k routing mechanism, ES-MoE enhances the diversity and discriminability of feature representations. Furthermore, we propose a Mixed Star Network (MSN), which selectively configures high-order feature modeling modules in a hierarchy-aware manner within a multi-branch structure, achieving a fine-grained trade-off between representational capacity and computational efficiency. Building upon these components, a decoupled detection head is adopted to construct the unified EMamba detection framework. Extensive experiments on multiple steel surface defect datasets demonstrate that EMamba achieves competitive detection accuracy, structural defect perception capability, and inference efficiency compared with representative existing methods. These results indicate the effectiveness and practicality of the proposed hybrid modeling paradigm for steel surface defect detection.