<p>Steel surface defect detection is critical for industrial quality and safety. This study addresses the key challenge of achieving real-time and accurate steel surface defect detection under industrial conditions, specifically tackling the problems of unstructured defect morphology, extreme multi-scale variations, and complex noise interference. To address these challenges, we propose CMH-Net, a lightweight one-stage detection model incorporating a cascaded Haar wavelet downsampling module for enhanced feature extraction of unstructured defects and a multi-scale Mamba-Like Linear Attention (MLLA) module for effective multi-scale feature fusion. The model is further optimized with a hybrid loss function to improve localization accuracy and robustness against noise. Through these integrated mechanisms, CMH-Net achieves efficient and accurate defect detection with minimal parameter overhead. Moreover, these proposed modules are designed as versatile plug-and-play components, allowing integration into other network backbones. Experimental results on two public steel surface defect datasets demonstrate that CMH-Net outperforms state-of-the-art methods in terms of mAP50 (NEU-DET: 81.4%, GC10-DET: 71.5%). With only 3.6M parameters and 6.4ms inference time, our model achieves an optimal trade-off between detection accuracy and efficiency. Furthermore, CMH-Net also achieves an mAP50 of 76.2% on the Wood dataset. It reveals that our work also provides a foundation for extending experiments to other industrial detection tasks with satisfactory improvements.</p>