Research on Industrial Defect Machine Vision Detection Algorithm Based on Deep Learning and Spectral Fusion

Industrial surface defect detection is core to intelligent manufacturing quality assurance, guaranteeing stable product quality and reliable production. Conventional defect detection deep learning algorithms relying on Red-Green-Blue (RGB) images suffer low accuracy and weak robustness amid complicated industrial scenes. This paper puts forward Multi-Spectral Fusion Defect Detection Network (MS-FDNet), a novel industrial surface defect detection framework combining deep learning and spectral imaging. It adopts a spectral-spatial attention coupling module to fully fuse multi-band spectral information and exploit distinctive feature advantages of each band. Self-calibrated convolution coupled with spatial pyramid structure constructs a multi-scale feature extraction branch to strengthen the model’s perception of defects of varying sizes. A multi-task composite loss integrating Focal Loss and Dice Loss is designed to balance model memory occupation and detection precision and boost overall inference efficiency. Tests on the public NEU-det dataset reveal the proposed algorithm achieves a mean Average Precision at Intersection over Union threshold 0.5 (mAP@0.5) of 81.1%, a 4.6% improvement over the baseline while retaining fast inference. It runs at 67.9 frames per second with a tiny model volume of merely 4.22 MB, showing outstanding deployment practicability. The method greatly enhances defect identification performance and system stability under uneven illumination and cluttered backgrounds. This work offers an innovative industrial defect detection scheme with valuable theoretical research value and promising engineering application potential.