| 陈杰,李展,颜普,徐恒,杨某闪.基于3D SE-Densenet网络的视频暴力行为识别改进算法[J].安徽建筑大学学报,2023,31(): |
| 基于3D SE-Densenet网络的视频暴力行为识别改进算法 |
| Improved Video Violence Recognition Algorithm Based on 3D SE-Densenet Network |
| 投稿时间：2021-12-14 修订日期：2022-01-10 |
| DOI： |
| 中文关键词: 暴力行为识别 深度学习 Densenet SENet |
| 英文关键词: violence identification deep learning Densenet SENet |
| 基金项目:国家自然科学基金项目（61901006,62105002,62001004）；安徽省自然科学基金项目（1908085QF281,2008085MF182）；安徽高校协同创新项目（GXXT-2019-007）;安徽省住房城乡建设科学技术计划项目（2020-YF22） |
|
| 摘要点击次数: 5346 |
| 全文下载次数: 1965 |
| 中文摘要: |
| 针对传统暴力行为识别算法精度不高和三维卷积神经网络参数多的问题，本文提出一种基于3D SE-Densenet网络的视频暴力行为识别改进算法。采用3D Densenet模型提取视频的时空特征信息，SENet（Squeeze-and-Excitation Networks）按照时空特征的重要性程度来进行加权处理，根据加权的时空特征实现对视频中暴力行为的识别。实验结果表明，本文提出的3D SE-Densenet方法在Hockey Fights Dataset和Movies Dataset上识别准确率分别达到了99.1%和100%，可较准确的识别出暴力行为，准确率要高于传统的方法。 |
| 英文摘要: |
| Due to low accuracy of traditional violence recognition algorithms and too many parameters of 3D convolutional neural network, a video violence recognition algorithm based on 3D SE-Densenet network is proposed in this paper. The spatiotemporal feature information in video blocks is extracted by 3D Densenet model, SENet (squeeze and exception networks) performs weighted processing according to the importance of spatiotemporal features, and realizes the recognition of violence in video according to the weighted spatiotemporal features.The recognition accuracy of our 3D SE-Densenet method has reached 99.1% and 100% on hockey fights dataset and movies dataset respectively. The experimental results show that our method can better identify video violence and the accuracy is higher than the traditional method. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
