基于序列化卷积神经网络的行为识别研究与实现

发布时间：2018-06-26 03:26

本文选题：行为识别 + 卷积神经网络　；参考：《电子科技大学》2017年硕士论文

【摘要】：在最近几年来,行为识别技术对我们的日常生活影响越来越多,在本文中采用目前很热门的深度学习的方法来进行行为识别研究与应用。深度学习在很多领域的优异的表现吸引着各个领域的研究人员进行研究,而卷积神经网络(CNN)在计算机视觉领域应用最为成功,自然吸引了大量的研究人员,卷积神经网络(CNN)相对于传统的神经网络具有参数少、平移不变性等优点,另外CNN也是一种端到端的网络,不需要手动的设计特征,训练该网络仅仅只需要大量的训练样本,然后对样本进行标注,表明样本属于哪一类,之后再把样本输入到网络中,进行迭代使卷积神经网络得到分类的能力,进而就能获得不错的识别率。在本文中主要讲述的利用一些优秀的结构搭建的一些神经网络并对这些神经网络的性能进行一些验证和分析。本文主要内容包括:(1)针对原始3DCNN网络的不足,对其网络的结构优化的进行了优化,网络的结构中融入了MLP卷积层,该层中有多层感知机,该层在不增加网络深度的前提下,使得该神经网络在特征表征能力与抽象泛化能力方面有了很大的提升。另外,改进了下采样的操作,在时间维度上也进行下采样操作,将网络在空间下采样的优点推广到了时间维度,使得网络在时间维度上具有不变形,也大大的减少了时间维度上的计算量,大大的提高了网络的性能。另外,采用了Relu激活函数替代原有的双正切tanh函数,从饱和的激活函数变为了非饱和的激活函数,使得网络在训练过程中的训练时间大大的减少,增加了网络的收敛速度,另外通过relu激活函数的特征图很稀疏,使得泛化能力很强。另外,改进的网络输入减少了4个通道信息,增加了一个通道的设计的堆叠光流信息,减少了不必要的输入,减少了网络的计算量,提高了网络的性能。(2)本文中重点阐述了一个多任务递归卷积神经网络模型,在本章中所设计的多任务递归卷积神经网络结合了多任务学习理论,融入了改进的VGG网络结构与LSTM结构建立一个全新的更深的网络结构,为了缓和该神经网络在训练的过程中出现过拟合的情况,加入了Dropout技术,使网络能够更好的提取长时间序列化视频的特征,与多任务识别的能力。(3)最后,在本文中主要描述了一个基于Super 3D-CNN网络的行为识别系统,该行为识别系统具有快速部署,不需要对并对现有的监控系统进行特殊的改造,另外使用GPU加速的方法对识别算法进行加速,可以使该行为识别系统达到实时性的要求。
[Abstract]:In recent years, behavior recognition technology has more and more influence on our daily life. In this paper, we use the popular in-depth learning method to conduct behavior recognition research and application. The excellent performance of deep learning in many fields has attracted researchers in various fields, while the application of convolutional neural network (CNN) in computer vision is the most successful, which naturally attracts a large number of researchers. Compared with traditional neural networks, convolutional neural networks (CNN) have the advantages of less parameters and invariant translation. In addition, CNN is also an end-to-end network, which does not require manual design features, and only requires a large number of training samples to train the network. Then the samples are labeled to show which kind of samples belong to, and then input the samples into the network, iterate to make the convolution neural network classification ability, and then can obtain a good recognition rate. In this paper, we mainly discuss some neural networks based on some excellent structures, and verify and analyze the performance of these neural networks. The main contents of this paper are as follows: (1) in view of the shortcomings of the original 3DCNN network, the network structure is optimized. MLP convolution layer is incorporated into the network structure. There are multi-layer perceptron in the layer, which does not increase the depth of the network. The neural network has been greatly improved in feature representation and abstract generalization. In addition, the downsampling operation is improved, and the down-sampling operation is also carried out in the time dimension. The advantages of the network sampling in space are extended to the time dimension, which makes the network have no deformation in the time dimension. Also greatly reduces the time dimension computation, greatly improves the network performance. In addition, the reu activation function is used to replace the original bitangent tanh function, which changes from the saturated activation function to the unsaturated activation function, which greatly reduces the training time of the network during the training process and increases the convergence speed of the network. In addition, the characteristic graph of relu activation function is very sparse, which makes the generalization ability very strong. In addition, the improved network input reduces the 4 channel information, increases the stack optical flow information of the design of one channel, reduces the unnecessary input, and reduces the network computation. The performance of the network is improved. (2) in this paper, a multitask recursive convolution neural network model is described. In this chapter, the multitask recursive convolution neural network is combined with the multitask learning theory. The improved VGG network structure and the LSTM structure are integrated to establish a new and deeper network structure. In order to ease the situation that the neural network has been fitted during the training process, Dropout technology is added. So that the network can extract features of long-time serialized video better, and the ability of multi-task recognition. (3) finally, this paper mainly describes a behavior recognition system based on Super 3D-CNN network, the behavior recognition system has rapid deployment. There is no need for special modification of the existing monitoring system, and the GPU acceleration method is also used to accelerate the recognition algorithm, which can make the behavior recognition system meet the requirement of real-time.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【相似文献】