<p>Current steel surface defect detection methods face challenges such as excessive computational redundancy, limited multi-scale fusion capability, and inaccurate localization of small or irregular defects. To address these challenges, we propose a lightweight multi-scale fusion and localization-enhanced YOLO network (LMFL-YOLO). Firstly, we design a Lightweight Dual-Path Fusion (LDPF) module that reduces computational and memory overhead through partial convolution (PConv) group operations, performs parallel feature extraction using both simple and complex convolutions, and incorporates residual connections to preserve spatial details. This module enhances small object detection performance and accelerates inference. Secondly, to further improve the network’s multi-scale feature fusion capability, we design a Dynamic Multi-Scale Feature Fusion (DMFF) module in the neck of the network. This module improves cross-scale feature alignment through a Bi-level Routing Attention (BRA) mechanism, enhances relevant feature selection and suppresses noise, and refines features by incorporating residual connections, adjusting spatial and channel responses, and effectively fusing multi-scale feature maps. Finally, normalized Gaussian Wasserstein distance (NWD) is adopted to replace complete intersection over union (CIoU) as the regression loss function in the detection head, thereby improving the localization accuracy of small or irregular defects. In addition to single-device performance, the proposed module adheres to the core tenets of GPU parallelism and enables scalable data-parallel deployment in multi-camera industrial inspection systems, leveraging High-Performance Computing (HPC). We conduct comprehensive experiments on multiple datasets, which successfully demonstrate the superior performance and strong generalization capability of the proposed method for steel surface defect detection.</p>