摘要
针对常见的暴力视频分类模型参数量大、分类准确率较低等问题,提出了一种融合轻量型网络MobileNetV3-Small与注意力机制的暴力视频分类方法 M-AttBiLSTM(MobileNetV3-Small Attention-BiLSTM)。首先采用数据增强方法对数据集进行预处理;接着利用MobileNetV3-Small网络提取视频帧中的暴力特征,结合Attention机制与BiLSTM网络进一步对视频帧的时间序列信息进行建模并对特征信息进行融合;最后将融合得到的视频特征利用Softmax分类器对视频进行分类。实验结果表明,提出的M-AttBiLSTM方法在降低模型参数量的情况下,在Hockey Fight和Violent Flows两个公开数据集上的准确率达到了99.84%和96.53%,优于已有研究,验证了其在暴力视频分类领域的有效性及可行性。
To address the problems of large number of parameters and low classification accuracy of common violent video classification models,a violent video classification method M-AttBiLSTM(MobileNetV3-Small Attention-BiLSTM),which combines the lightweight network MobileNetV3-Small with the Attention mechanism,is proposed.Firstly,the data set is pre-processed by data enhancement method,then the violent features in the video frames are extracted using MobileNetv3-Small network,and the time series information of the video frames is further modeled and fused with the attention mechanism and BiLSTM network.Finally,the fused video features are classified using Softmax classifier.The experimental results show that the proposed M-AttBiLSTM method achieves 99.84%and 96.53%accuracy on two public datasets,Hockey Fight and Violent Flows,with a reduced number of model parameters,which is better than the existing studies and verifies its effectiveness and feasibility in the field of violent video classification.
作者
李娜
白涛
王迎超
马楠
LI Na;BAI Tao;WANG Yingchao;MA Nan(College of Computer and Information Engineering,Xinjiang Agricultural University,Urumqi 830052;Xinjiang Agricultural Informatization Engineering Technology Research Center,Urumqi 830052)
出处
《计算机与数字工程》
2025年第6期1711-1716,共6页
Computer & Digital Engineering
基金
新疆维吾尔自治区重点研发项目“多维度信息精准推送体系关键技术研究与应用”(编号:2017B01006-1)
2022年自治区高校基本科研业务费项目“农业大数据交换共享与可视化平台构建”
乌鲁木齐市科学技术计划项目(编号:Y16330001)资助。
作者简介
李娜,女,硕士研究生,研究方向:视频内容理解;白涛,男,硕士,副教授,研究方向:农业大数据、数据挖掘;王迎超,男,硕士研究生,研究方向:图像处理;马楠,女,硕士研究生,研究方向:农业信息化。
