基于C3D模型的视频分类技术  

Video classification technology based on C3D model

作　　者：孙毅 成金勇[1] 禹继国 SUN Yi;CHENG Jinyong;YU Jiguo(School of Computer Science and Technology,Qilu University of Technology(Shandong Academy of Sciences),250353,Jinan,Shandong,PRC)

机构地区：[1]齐鲁工业大学(山东省科学院)计算机科学与技术学院,山东省济南市250353

出　　处：《曲阜师范大学学报（自然科学版）》2020年第3期85-89,共5页Journal of Qufu Normal University(Natural Science)

基　　金：山东省科技重大专项(2019JZZY020124)。

摘　　要：目前,解决视频分类问题比较典型的方法是使用深度学习方法.该文设计了一种新的神经网络结构用于解决视频分类问题同时使用了交叉熵损失函数和一些减少神经网络过拟合的方法.网络结构采用了3D卷积神经网络结构,这是由于3D卷积神经网络相比2D卷积网络可以同时处理图像时域信息和图像空间信息,保留输入信息的时间特征.我们将视频文件通过各种手段,转化为图像帧的形式,放入该文设计的3D卷积神经网络中学习和训练,最后通过分类器对图像的的种类进行划分,得到每个数据分类概率的结果.与之前的C3D网络相比我们增加了网络的深度,优化了网络结构,并通过实验验证了改进的有效性.At present,the typical method to solve the video classification problem is using deep learning methods.In this paper,a new neural network structure is designed to solve the video classification problem.Video classification problem also uses the cross-entropy loss function and some methods to reduce the overfitting of the neural network.The network structure uses a 3D convolution neural network structure.The 3D convolution neural network can process image time domain information and image space information at the same time compared to the 2D convolution network,retaining the temporal characteristics of the input information.We convert the video file into the form of an image frame by various means,put it into the 3D convolution neural network.We designed to learn and train,and finally divide the type of image through the classifier to obtain the classification probability of each data.Compared with the previous C3D network,we increase the depth of the network,optimize the network structure,and verify the effectiveness of the improvement through experiments.

关 键 词：视频分类 卷积神经网络 时间特征 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]