Computer Vision-Based Anomaly Diagnosis in Knitted Fabrics: A Graph-Theoretic Approach to Stitch Defect Localization

Automated quality inspection remains a critical challenge in textile manufacturing, where subtle fabric defects can significantly impact product value and performance. Existing computer vision systems struggle to simultaneously achieve precise stitch-level defect localization and robust generalization across diverse knitting patterns and deformations. This paper presents a novel hybrid architecture combining dual-branch convolutional neural networks with edge-conditioned graph neural networks, where visual features inform topological graph representations of stitch connectivity patterns for comprehensive anomaly detection. Experimental results demonstrate superior performance with 91.3% detection accuracy (vs. 85.7% for commercial systems), 92.8% overall recall, with strong performance on critical defect categories, and real-time processing at 20 FPS, while reducing the overall false positives rate to 2.3%, and up to 8.9% improvement over the baseline 3D CNN on connectivity defects. The proposed framework establishes a new paradigm for fabric inspection by integrating computer vision with structural graph analysis, achieving both high precision and practical deployment efficiency for industrial quality control applications.