Scholarlink

<p>Violence detection within intelligent video surveillance systems is a pivotal component of smart city infrastructure. However, deploying deep learning models in real-world scenarios is hindered by the prohibitive computational complexity of centralized, large-parameter architectures, which are ill-suited for resource-constrained edge devices. Furthermore, the opaque "black box" nature of these complex models limits the interpretability essential for trustworthy decision-making. To address these challenges, this paper proposes EdgeViolenceLite (EVL), a lightweight framework optimized for edge deployment. First, we introduce a novel nonlinear module that leverages kernel functions and frame-difference cues to replace the computation-intensive dot-product attention mechanism of traditional Transformers, significantly reducing computational overhead. Second, a knowledge distillation strategy is implemented to transfer learned representations into a compact, five-layer Convolutional Neural Network (CNN). This design achieves substantial model compression, and the avoidance of feature-flattening techniques facilitates more interpretable feature representations. Extensive experiments on the XD-Violence and UCF-Crime benchmarks demonstrate that the proposed method achieves a superior balance between high detection accuracy and operational efficiency, meeting the rigorous requirements for practical surveillance applications in edge-computing environments. The project code can be found at <a href="https://github.com/1846659840/EVL_Project">https://github.com/1846659840/EVL_Project</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

A lightweight framework for efficient violence detection via kernel-based attention and knowledge distillation

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Zirui Wang,
  • Kerui Weng,
  • Nansheng Lin,
  • You-xi Li

摘要

Violence detection within intelligent video surveillance systems is a pivotal component of smart city infrastructure. However, deploying deep learning models in real-world scenarios is hindered by the prohibitive computational complexity of centralized, large-parameter architectures, which are ill-suited for resource-constrained edge devices. Furthermore, the opaque "black box" nature of these complex models limits the interpretability essential for trustworthy decision-making. To address these challenges, this paper proposes EdgeViolenceLite (EVL), a lightweight framework optimized for edge deployment. First, we introduce a novel nonlinear module that leverages kernel functions and frame-difference cues to replace the computation-intensive dot-product attention mechanism of traditional Transformers, significantly reducing computational overhead. Second, a knowledge distillation strategy is implemented to transfer learned representations into a compact, five-layer Convolutional Neural Network (CNN). This design achieves substantial model compression, and the avoidance of feature-flattening techniques facilitates more interpretable feature representations. Extensive experiments on the XD-Violence and UCF-Crime benchmarks demonstrate that the proposed method achieves a superior balance between high detection accuracy and operational efficiency, meeting the rigorous requirements for practical surveillance applications in edge-computing environments. The project code can be found at https://github.com/1846659840/EVL_Project.