Beyond vision-only anomaly detection: DualAD-LLM by leveraging large language model with dual-detector model for leather defect inspection

Document Type

Article

Publication Date

8-25-2026

Abstract

Anomaly detection in industrial leather inspection has become increasingly important in computer vision and manufacturing, as it enables the identification of subtle surface defects that critically impact material quality and market value. Although advanced models such as EfficientAD and PatchCore have achieved promising results, their independent operation often neglects complementary information, leading to suboptimal decision boundaries. To address this limitation, a novel anomaly detection framework is proposed that integrates EfficientAD and PatchCore through a Large Language Model (LLM)-based score fusion strategy, namely Dual Anomaly Detection with LLM Fusion (DualAD-LLM). In the proposed approach, anomaly scores from both models are considered, and the LLM generates a fused probability-like score that captures nonlinear dependencies between detection methods. This fused score is further concatenated with the original features and classified using KNN, SVM, and neural networks under a 5-fold cross-validation protocol. Experiments conducted on a self-collected leather patch dataset comprising 250 defective and non-defective samples demonstrate that the LLM-based fusion consistently enhances performance, achieving up to 97.2% accuracy and F1-score. Comprehensive analyses confirm both the discriminative strength and interpretability of the proposed method. To the best of our knowledge, this is the first study to introduce LLM-based feature generation for anomaly detection in leather inspection. These findings suggest that the proposed framework offers a promising and explainable strategy for advancing automatic quality control in the leather industry.

Keywords

Anomaly detection, EfficientAD, Large Language Model (LLM), PatchCore, Score fusion

Publication Title

Expert Systems with Applications

ISSN

0957-4174

DOI

10.1016/j.eswa.2026.132521

Volume

324

Share

COinS