Synthesis-guided unsupervised anomaly detection in industrial images with large language model-driven analysis

초록

Surface anomaly detection is crucial in industrial imaging, requiring accurate identification of deviations from expected patterns. Traditional reconstruction-based models often fail to detect unseen anomalies due to their reliance on learned normal distributions. To overcome this, we propose Synthesis-Guided Unsupervised Anomaly Detection in Industrial Images with Large Language Model-Driven Analysis (SUADA), which frames anomaly detection as a discriminative task rather than a reconstructive one. Our approach generates artificial anomalies by combining Perlin and random noise for diverse anomaly generation. We further employ a transformer-driven attention mechanism within a hierarchical feature extraction framework to enhance anomaly localization and segmentation. The attention mechanism captures long-range dependencies, particularly across skip connections, enabling selective feature fusion for improved robustness against scale variations and efficient parameterization. Unlike conventional methods, SUADA directly localizes anomalies without complex post-processing. Furthermore, SUADA integrates an anomaly analysis module powered by Large Language Models (LLMs), which provides human-interpretable descriptions of detected anomalies, detailing their shape, location, and severity. By analyzing contours, spatial distributions, and shape characteristics, this module translates anomaly features into actionable insights, making industrial anomaly detection more accessible to non-expert users. SUADA achieves state-of-the-art performance on the MVTec dataset, with AUROC (Image) 96.01%, AUROC (Pixel) 86.87%, AP (Image) 96.86%, and AP (Pixel) 56.92%, while maintaining a high inference speed of 82 FPS, making it suitable for real-world industrial applications. These results demonstrate the effectiveness of our approach in enhancing neural computing for anomaly detection. By combining unsupervised learning, synthetic data generation, and language-based analysis, SUADA offers a scalable framework adaptable to diverse industrial inspection tasks.

키워드