基于非监督卷积神经网络的运动估计研究

发布时间：2019-05-24 16:18

【摘要】：AlexNet的发布开启了计算机视觉的新纪元,计算机视觉领域开始走向深度学习的新时期。在这过程中对于使用卷积神经网络的运动估计的研究也逐渐展开。由于该机器学习系统复杂度较高,所以其对训练数据量的需求非常庞大。但是在现阶段能够针对运动估计问题训练卷积神经网络的大型标准数据库还并不存在。因此研究如何使用非监督学习的方法根据具体问题设计卷积神经网络和如何使用迁移学习将已有的成熟的卷积神经网络系统运用到解决运动估计问题中有重要的意义。本文主要的工作内容如下:(1)基于光流方程建立了成本方程,用于以非监督方式训练卷积神经网络。为了处理运动估计,精细设计了卷积神经网络(CNN)的体系结构,其中主要是针对卷积层中卷积核的数量和卷积核的大小进行设计和调整,并建立卷积层与卷积层间的结构关系。在卷积神经网络中引入了残差结构,以应对卷积神经网络中由于卷积层过多造成的梯度消散和训练时间过长的问题。再者针对光流的不足,设计了一种从粗糙到精细的模型,处理本文神经网络中无法对输入图片对中较大幅度运动进行计算的问题。最后通过实验验证该网络在运动估计问题中的有效性,和残差结构对卷积神经网络在运动过程中的作用。(2)为了能够提高本文中卷积神经网络的性能,本文使用了迁移学习的方法对该系统进行升级。因为在现实中物体一般是运动的载体,所以在本文中引入了VGG模型中的前两个卷积层和其相应的参数,对输入的图片进行“预处理”,并且使用一种非一般结构的卷积神经网络分别处理输入的图片对中的图片。在连接从VGG中引入的部分卷积层和用非监督学习训练的卷积神经网络时,加入包含3D卷积核的卷积层用于提取运动信息。对于光流无法对较大幅度进行运动计算的缺点,本文中使用课程学习的方法制定策略由浅入深的对卷积神经网络进行训练,提高其面向较大运动幅度计算的鲁棒性。最后在实验中,使用不同的训练策略训练卷积神经网络,检测和证明本文中所描述的方法的有效性。
[Abstract]:The release of AlexNet ushered in a new era of computer vision, and the field of computer vision began to move towards a new period of deep learning. In this process, the research on motion estimation using convolution neural network is also gradually carried out. Because of the high complexity of the machine learning system, the demand for training data is very large. However, at present, there is no large standard database which can train convolution neural network for motion estimation. Therefore, it is of great significance to study how to use unsupervised learning method to design convolution neural networks according to specific problems and how to use transfer learning to apply the existing mature convolution neural network systems to solve motion estimation problems. The main contents of this paper are as follows: (1) based on the optical flow equation, the cost equation is established to train convolution neural networks in an unsupervised manner. In order to deal with motion estimation, the architecture of convolution neural network (CNN) is designed and adjusted, which is mainly designed and adjusted for the number of convolution cores and the size of convolution cores in convolution layer. The structural relationship between convolution layer and convolution layer is established. The residual structure is introduced into the convolution neural network to deal with the problem of gradient dissipation and long training time caused by too many convolution layers in the convolution neural network. Furthermore, in view of the shortcomings of optical flow, a rough to fine model is designed to deal with the problem that the large motion in the input picture pair can not be calculated in the neural network. Finally, the effectiveness of the network in motion estimation problem and the effect of residual structure on convolution neural network are verified by experiments. (2) in order to improve the performance of convolution neural network in this paper, In this paper, the method of transfer learning is used to upgrade the system. Because in reality, objects are generally the carrier of motion, so in this paper, the first two convolution layers and their corresponding parameters in VGG model are introduced, and the input pictures are "preprocessed". A convolution neural network with non-general structure is used to process the pictures in the input picture pairs. When connecting the partial convolution layer introduced from VGG and the convolution neural network trained by unsupervised learning, the convolution layer containing 3D convolution kernel is added to extract the motion information. In order to solve the problem that the optical flow can not calculate the motion of a large range, this paper uses the course learning method to formulate a strategy to train the convolution neural network from shallow to deep, so as to improve the robustness of the convolution neural network for large motion amplitude calculation. Finally, in the experiment, different training strategies are used to train convolution neural network to detect and prove the effectiveness of the method described in this paper.
【学位授予单位】：南京邮电大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【相似文献】