Industrial defect detection replaces manual visual inspection with an efficient, accurate, and reproducible automated intelligent system, directly enhancing quality management in intelligent manufacturing and providing clear engineering value for production safety and corporate competitiveness; however, their production processes often generate various defects. Detecting these defects is challenging due to issues such as high inter-class similarity among defect types, weak semantic information in complex small objects, and the presence of multi-scale defective targets, all of which compromise product quality and safety. To overcome these challenges, we introduce MFDH-Net, an innovative industrial defect detection network. The framework begins with a Dual-domain Feature Extraction Network (DFE-Net), which employs a full-domain local sensing module to capture both global and local defect information bidirectionally. Next, a Multilevel Feature Aggregation Network (MFA-Net) is designed, incorporating a full-level fusion mechanism for multiscale features and an adaptive feature weighting strategy along bidirectional propagation paths, enabling deep feature interaction both within and across layers. Additionally, a Spatial Semantic Fusion Module (SSFM) is introduced to achieve spatial and pixel-level alignment, effectively mitigating semantic gaps in the feature pyramid that degrade detection performance. Finally, a Cross-aware Decoupling Head (CDH) extracts fine-grained defect representations, while a Cross-aware Attention Module (CAM) enhances sparse, defect-related features through horizontal and vertical dual attention weighting in spatial dimensions, thereby improving recognition of tiny defects in complex industrial backgrounds. The experimental results show that 94.3%, 96.6%, 71.7 and 98.9% of mAP@.5 are obtained on steel, PCB, GC10-Net and Automobile Surface Defect datase, respectively, and they reach 52FPS, obtaining SOTA performance, which provides reliable technical guarantee for intelligent manufacturing quality control.
Zhang et al. (Wed,) studied this question.