<p>In the steel manufacturing industry, the detection and analysis of surface defects are crucial for ensuring high production quality and operational efficiency. Undetected defects can result in increased material wastage, elevated costs, and compromised product reliability. Although deep learning techniques have significantly advanced automated defect detection, their effectiveness often depends on access to large, meticulously labeled datasets and high-performance computing resources–requirements that are not always feasible in practical industrial settings. To address these limitations, this study introduces a lightweight Multi-Scale Attention Network (MSANet) designed to enhance defect detection performance while maintaining computational efficiency. The proposed MSANet architecture effectively captures and emphasizes defect-relevant features at multiple spatial scales. It begins by extracting multi-scale feature maps using SqueezeNet as the backbone. These maps are then refined through a purpose-built attention module that selectively focuses on regions most indicative of defects, enhancing the network’s discriminative capacity. A key innovation of MSANet lies in its feature fusion strategy, which integrates information from multiple scales to form a unified and enriched representation. This fusion not only boosts classification accuracy but also reduces reliance on heavy computational resources, making MSANet particularly suitable for deployment in real-world industrial environments. Experimental evaluations on the NEU surface defect dataset validate the effectiveness of the proposed approach. MSANet achieves an outstanding classification accuracy of 99.72%, significantly outperforming conventional models. Moreover, the network maintains over 92% accuracy even under data-scarce conditions, highlighting its robustness and strong generalization ability across a diverse range of defect types and operational scenarios.</p>