In electrical engineering and automated manufacturing systems, high-precision, non-contact defect detection and geometric measurement are key to ensuring reliable equipment operation and product quality control. The transition from qualitative defect detection to quantitative three-dimensional (3D) characterization remains a significant challenge in automated industrial inspection. To address these issues, this paper proposes a 3D visual weak-texture defect detection and quantification method for industrial inspection applications in electrical automation. This method is based on a binocular vision perception framework, integrating pixel-level semantic segmentation and robust stereo matching to achieve precise defect localization and 3D metrology. The system first employs a Multi-Scale Edge-Aware Segmentation Network (MSEA-Net) to extract defect regions with exceptional boundary fidelity. Subsequently, a Deformable Convolution and Attention-guided Stereo Matching Network (DCASM-Net) is proposed to reconstruct dense and accurate 3D point clouds from surfaces with weak or repetitive textures, a common challenge in industrial settings. The 2D segmentation masks are precisely mapped to the 3D coordinate space, enabling the direct computation of key physical dimensions such as length, width, depth, and projected area. Comprehensive experiments on a dedicated industrial defect dataset demonstrate that our system achieves state-of-the-art performance in both defect segmentation (F1-score: 96.2%, IoU: 93.1%) and stereo matching (≥ 3px error rate: 2.1% in non-occluded areas). The proposed framework can serve as a visual perception and metrology module in electrical automation systems and intelligent manufacturing equipment, providing a reliable technical path for automated 3D defect detection in industry.
Xiong et al. (Wed,) studied this question.