Boosting Mamba-You Only Look Once (YOLO) with triple-aware enhancement for steel surface defect detection

Accurate and efficient surface defect detection is crucial for ensuring the structural integrity and quality of industrial steel products. However, existing detectors still struggle with sub-pixel, low-contrast defects and extreme variations in defect scale and shape. To address these challenges, we propose TriA-Mamba-You Only Look Once (YOLO), a triple-awareness enhancement of the Mamba-YOLO framework for steel defect detection. First, the input samples are processed using a discrete wavelet transform, which reduces background noise and enhances the representation of defect features. Next, the backbone is enhanced with a micro-feature-aware enhancement block, which comprises a Smooth-Mix Focus Stem (FS) block to minimize aliasing artifacts while preserving critical defect features, combined with a Selective Fusion Attention (SF-Att) block for refinement by selectively emphasizing discriminative micro-scale patterns. Within the neck, a context-aware feature enhancement module is implemented to tackle extreme scale and shape variance. It utilizes a receptive-field lightweight spatio-channel feature refinement (RLSC-FR) block to optimize local structure and receptive field characteristics, and an Attention-Augmented Visual State-Space (AA-VSS) block to dynamically adapt to defect scale and morphological variations. Finally, a path-importance-aware fusion module, using Norm-Weighted concatenation (NW-Concat), is developed for multi-path feature fusion, dynamically adjusting contribution weights to ensure balanced aggregation. Experimental results show that TriA-Mamba-YOLO achieves a mean average precision (mAP) of 81.3%, surpassing the original Mamba YOLO by 5.2% on the Northeast University-Defect Detection (NEU-DET) dataset, and achieves a balance between accuracy, computational efficiency, and robustness. These results highlight the effectiveness of the proposed approach for industrial surface defect detection applications.