计算机科学
异常检测
人工智能
模式识别(心理学)
噪音(视频)
特征提取
目标检测
计算机视觉
雷达跟踪器
信号处理
噪声测量
鉴定(生物学)
数据建模
信噪比(成像)
杂乱
数据挖掘
作者
Haojin Tang,Zhipeng Xiong,梁志坚,Yizhi Luo,Xiaofei Yang,Xin Lin,Dong Tang
标识
DOI:10.1109/tii.2026.3738819
摘要
Generalizable few-shot anomaly detection is crucial in industrial inspection, requiring models to generalize to unseen categories and detect diverse defects using only a few normal samples. Although vision–language models provide strong semantic priors and generalization capability, existing methods often rely on unidirectional vision–language interaction and single-paradigm visual modeling, limiting cross-category generalization and multimorphological anomaly representation. To address these issues, we propose a bidirectional visual–textual learning (BVTL) framework. First, we introduce a hybrid discrete–continuous visual feature modeling module that integrates a discrete-sensitive Transformer with a continuous-sensitive Mamba network to capture multimorphological anomalies. Second, a bidirectional semantic guidance mechanism is designed to enable textual prompts to guide visual representation learning, while the learned visual features in turn facilitate category-agnostic semantic anchor learning for enhanced cross-category generalization. Moreover, a consistency-guided vision–language collaborative decision strategy is incorporated to model consistency across visual and textual predictions, guide reliability estimation, and perform adaptive fusion by balancing the learned integration score with the boundary reference score. Finally, extensive experiments on MVTec AD, VisA, MPDD, and AITEX demonstrate that BVTL achieves superior or highly competitive performance on anomaly classification and segmentation tasks, exhibiting strong effectiveness and generalization in generalizable few-shot industrial anomaly detection.
科研通智能强力驱动
Strongly Powered by AbleSci AI