基于YOLO和ConvLSTM混合神经网络的暴力视频检测

VIOLENT VIDEO DETECTION BASED ON YOLO AND CONVLSTM HYBRTD NEURAL NETWORK

为解决现有暴力视频检测算法所存在的特征提取繁琐、时空特征信息利用少等问题,提出一种基于YOLO和ConvLSTM混合神经网络的暴力视频检测算法,通过1×1卷积改进DarkNet-53特征提取网络的全连接层,进而结合ConvLSTM根据时空信息序列化建模进行检测,解决了原本全连接层破坏图像结构、输入尺寸固定的问题,更好地保留了暴力视频的特征。经Hockey、RWF-2000和自定义的数据集实验的结果表明,该模型较其他传统模型的分类准确率更高。

In order to solve the problems of existing violent video detection algorithms,such as tedious feature extraction and less utilization of spatiotemporal feature information,a violence video detection algorithm based on YOLO and ConvLSTM hybrid neural network is proposed.The full connection layer of Darknet-53 feature extraction network was improved by 1×1 convolution,and combined with ConvLSTM,the detection was carried out based on spatiotemporal information serialization modeling.It solved the problem that the original full connection layer destroyed the image structure and the input size was fixed,and retained the characteristics of violent video better.The experimental results on hockey,rwf-2000 and user-defined datasets show that the classification accuracy of this model is higher than that of other traditional models.