Semantic-prior enhanced representation learning for weakly supervised video anomaly detection

Weakly supervised video anomaly detection (WsVAD) is a challenging task in which only video-level anomaly labels are available during training, while frame-level predictions are required at inference. Most existing WsVAD approaches rely on multi-instance learning (MIL) paradigms to handle the weak supervision setting. However, these methods often suffer from limited capability in capturing the diverse and context-dependent characteristics of anomalies, and may underutilize informative abnormal segments, resulting in suboptimal detection performance. In this work, we propose a novel Semantic-prior Enhanced Representation Learning (SERL) framework to enhance the effectiveness of WsVAD. Given an input video, we first design a Dynamic Temporal Encoder to model long-range temporal dependencies and short-term contextual cues, thereby facilitating more accurate anomaly localization. Subsequently, a Semantic Enhancement Learning module is introduced to refine feature representations by integrating semantic priors derived from abnormality-aware prompts. Finally, a Sample-Batch Selection strategy is developed to effectively identify and leverage informative abnormal snippets, thus improving the exploitation of weak supervision signals. Extensive experiments on the widely adopted benchmark datasets UCF-Crime, XD-Violence, and ShanghaiTech demonstrate that the proposed SERL framework achieves state-of-the-art results, confirming its effectiveness and robustness in addressing the WsVAD task.