<p>Violence is a pressing issue worldwide, motivating the deployment of surveillance systems to ensure public safety and prevent criminal activities. Manual monitoring of massive video streams is laborious and error-prone, which has accelerated research into automated violence detection methods. This paper introduces CNeXt-DANet, a deep learning framework that combines the pretrained ConvNeXt backbone with a dual-attention mechanism to enhance the detection process. The backbone efficiently extracts discriminative features for representation. First, the channel attention module highlights the most informative channels by refining their feature weights. Next, the spatial attention module emphasizes important regions by aggregating contextual information across all channels. The sequential application of these attentions strengthens correlations and improves generalization. Extensive experiments on benchmark datasets validate its effectiveness, where CNeXt-DANet outperforms state-of-the-art models and attains 99.52% accuracy on RLVS and 99.23% on RWF-2000. In addition to high accuracy, detailed ablation and comparative analysis validate the reliability of CNeXt-DANet. The model’s efficiency and generalization make it suitable for practical real-time surveillance applications.</p>