Scholarlink

<p>Video anomaly detection remains a challenging task, particularly in identifying occluded, small-scale, and transient anomalies within complex contexts. Existing methods often overlook the enhancement of fine-grained target features and lack adaptive precise attention mechanisms, limiting their robustness and generalization ability. To address these limitations, we propose an adaptive feature refinement (AFR) method that integrates a small-object attention module (SAM) and small-object context-aware attention module (SCAM) into the Feature Pyramid Network of a clip-driven multi-scale instance learning architecture. This approach adaptively enhances the feature representation of key areas. Furthermore, we incorporate the Contrastive language-image pre-training (CLIP) model to enrich semantic information and improve cross-scene generalization. The SCAM module recalibrates the channel feature response to selectively enlarge object exception clues while suppressing irrelevant activations and the SAM module guides the model to focus on the discrimination mode of small-scale anomalies by integrating channel and spatial attention mechanisms. The semantic prior of the CLIP model further strengthens the expression ability of visual features. Extensive experiments on the UCF-Crime and XD-Violence datasets demonstrate that our AFR method outperforms state-of-the-art approaches, verifying its effectiveness and potential for real-world video anomaly detection tasks. The code is available at <a href="https://github.com/awsd123-wq/AFR/tree/master">https://github.com/awsd123-wq/AFR/tree/master</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Fine-grained video anomaly detection via adaptive feature refinement and semantic enrichment

  • 期刊论文

  • 发表日期：

    2026年1月16日

查看全文
  • Dezhi An,
  • Wenqiang Liu,
  • Jun Lu,
  • Shengcai Zhang

摘要

Video anomaly detection remains a challenging task, particularly in identifying occluded, small-scale, and transient anomalies within complex contexts. Existing methods often overlook the enhancement of fine-grained target features and lack adaptive precise attention mechanisms, limiting their robustness and generalization ability. To address these limitations, we propose an adaptive feature refinement (AFR) method that integrates a small-object attention module (SAM) and small-object context-aware attention module (SCAM) into the Feature Pyramid Network of a clip-driven multi-scale instance learning architecture. This approach adaptively enhances the feature representation of key areas. Furthermore, we incorporate the Contrastive language-image pre-training (CLIP) model to enrich semantic information and improve cross-scene generalization. The SCAM module recalibrates the channel feature response to selectively enlarge object exception clues while suppressing irrelevant activations and the SAM module guides the model to focus on the discrimination mode of small-scale anomalies by integrating channel and spatial attention mechanisms. The semantic prior of the CLIP model further strengthens the expression ability of visual features. Extensive experiments on the UCF-Crime and XD-Violence datasets demonstrate that our AFR method outperforms state-of-the-art approaches, verifying its effectiveness and potential for real-world video anomaly detection tasks. The code is available at https://github.com/awsd123-wq/AFR/tree/master.