MAFSRM: A Multiangle Feature Separation and Reconstruction Module for Industrial Defect Detection

In industrial manufacturing, the measurability of surface defects directly impacts the accuracy of quality assessment and the effectiveness of process control. While object detection-based methods can rapidly localize defects, they often fall short in supporting high-fidelity visual measurement due to several critical challenges: the strict sub-second timing required by high-speed production lines, the interference from complex backgrounds that distorts geometric and edge information, and the low salience and diversity of small-scale defects, which hampers the extraction of stable, measurement-relevant features. To address these issues, we propose a plug-and-play Multi-Angle Feature Separation and Reconstruction Module (MAFSRM), designed to enhance the quality and robustness of defect features from a measurement perspective. MAFSRM consists of three submodules: (1) Spatial Pyramid Pooling-Fast Separation and Reconstruction (SPPFSR) for fast, scale-adaptive feature extraction; (2) Weight Separation and Reconstruction (WSR) for isolating defect regions from background interference in the spatial domain; (3) Feature Separation and Reconstruction (FSR) for enhancing defect distinctiveness by refining channel-level feature representations. The module can be seamlessly integrated into mainstream single-stage detection frameworks, enabling easy deployment without altering the existing pipeline. Experiments on NEU-DET, GC10-DET, PCB, and a self-constructed Tire-DET dataset show that MAFSRM significantly improves both detection accuracy and feature stability, offering highly reliable inputs for downstream measurements of defect size and morphology.