FabricDefectNet: An AI-Driven Fabric Inspection System Utilizing Simulation Images as Reference Information

This paper introduces FabricDefectNet, an AI-driven visual inspection system for automating weaving inspection in the textile industry. Conventional AI inspection systems require many defective sample images for training, and the cost of data collection is a significant barrier to practical application, especially in today’s high-mix, low-volume production environment. To solve this problem, our system adopts a novel approach that uses fabric simulation images (generated from design data) not as direct training data, but as reference templates for comparison with the actual fabric. The core of the proposed system consists of three stages: (1) robust feature matching between the simulation image and the real fabric image using the state-of-the-art LightGlue matcher; (2) precise correction of non-rigid deformations (stretching, distortion, skew) in the fabric image using Thin Plate Spline (TPS) warping; and (3) pixel-wise anomaly detection by measuring cosine distances between the aligned patches. Evaluation experiments with a prototype system demonstrate that FabricDefectNet can accurately detect major defects such as color mismatches, pattern errors, and pitch differences without using any actual defect samples. This work demonstrates a practical AI tool that can detect unknown defects using only normal (design) data, which is expected to improve the efficiency and consistency of quality control in textile manufacturing.