Industrial defect detection is critical for ensuring product quality and improving manufacturing efficiency, yet remains challenging due to strong texture interference, large-scale variation, irregular morphology, and weak boundaries, which jointly degrade accuracy and cross-dataset robustness. To enable real-time deployment under limited computation, we propose MSF \(^{2}\) -Net, a multi-scale feature fusion framework for industrial surface defect detection. MSF \(^{2}\) -Net integrates three complementary components: (1) a decoupled hierarchical fusion feature pyramid network that decouples shallow spatial fusion and deep feature fusion to inject fine details while aligning multi-level semantics; (2) a multi-scale contextual feature enhancement block that expands the effective receptive field and strengthens contextual representation via heterogeneous convolutions and structured residual fusion; and (3) a vertical–horizontal cross attention module that introduces directional inductive bias by modeling long-range dependencies along both axes for anisotropic defects. Extensive experiments on four benchmarks Fabric6056, Tianchi, NEU-DET, and PKU-Market-PCB demonstrate that MSF \(^{2}\) -Net achieves mAP \(_{0.5}\) of 94.5%, 49.1%, 78.3%, and 95.6%, respectively, showing consistent improvements over strong baseline detectors. With only 5.6M parameters and real-time inference capability, MSF \(^{2}\) -Net provides a practical solution for industrial surface inspection.
