SDMC-Mamba: Steel Surface Defect Detection Based on DDIM Enhanced Mamba with Multi-module Collaboration

Steel surface defects directly impact industrial product quality and production safety, making high-precision detection critical for high-quality steel production. Existing methods have clear limitations: CNNs, Transformers, and traditional Mamba struggle to effectively extract corresponding features and coordinate global-local feature extraction. To address these challenges, we propose SDMC-Mamba—a DDIM-enhanced Mamba with multi-module collaboration for high-precision steel surface defect detection. We design the Collaborative Vector Selective Scan (CVSS) module, which improves SparX-Mamba’s VSS block by fusing convolution and attention mechanisms to enhance collaborative extraction of local features and global information. Second, we propose the Hybrid Enhancement Branch (HEB), adopting a CNN-VSS parallel structure to assist the main branch in supplementing edge details and suppressing overfitting. Furthermore, we use DDIM for data augmentation, expanding the original NEU-DET dataset into a balanced 3600-image dataset. Experiments on the augmented dataset show SDMC-Mamba achieves 82.4% mAP@50 (2.9% higher than the state-of-the-art) and 0.464 mAP@95 (0.09 higher than the state-of-the-art), outperforming mainstream advanced models such as YOLO, RT-DETR, and SparX-Mamba.