视觉印象深度学习算法研究

发布时间：2017-12-27 04:06

本文关键词：视觉印象深度学习算法研究　出处：《苏州大学》2016年博士论文　论文类型：学位论文

【摘要】：视觉印象是储存在人们记忆中的视觉信息,它是视觉认知过程中的一种重要形式。人类通过视觉获得的感官刺激,经由大脑的信息处理之后,形成有关认知客体的形象。这个形象以记忆的形式储存在脑海中,构成能够帮助人类理解与认知的视觉印象。视觉印象是人类大脑能够准确高效地完成各种各样复杂任务的基础。大脑在大部分时候直接借助于记忆中的视觉印象帮助视觉信息的处理,而不是通过盲目地计算。由于人类在认知客观事物时,八成以上的信息都来源于视觉并且在进行目标识别等任务中,人们总是通过已有的视觉印象去认知当前的事物,因此研究视觉印象机制对于模式识别、计算机视觉等领域具有重要的作用。深度学习通过堆叠单层模块构建一种深层的非线性网络,能够实现对复杂函数的无限逼近。深度学习能够以非监督的形式逐渐学习出不同抽象级别的特征,提供一种更具表现力的分布式表示方式。由于其自动学习特征的特性,使得相对于传统特征选择方法来说,深度学习节省了人工设计特征的代价。深度学习高度有效的特征提取方式,使其在目标识别等领域的应用都带来了突破性的结果。人类视觉皮层具有一个深度的结构。从模拟生物机制的角度,这成为支持深度学习技术的一个强有力的证据。从某种程度上,深度学习对人类逐层进行的认知过程进行了模拟。人脑的深度结构决定了视觉印象是对视觉信息逐步抽象的过程。这意味着逐层抽象的视觉印象能够与深度学习逐层提取的隐藏特征相对应。因而,将视觉印象机制与深度学习技术相结合是十分有潜力的研究方向。如何利用深度学习方法模拟人类视觉认知过程中产生的视觉印象是一个亟待解决的问题。这个问题要求设计的深度学习算法一方面要能够体现视觉印象的一些特点,另一方面也要能够完成视觉印象的相关功能。在深度学习的基础上,如何能够有效地提取出具有层次结构的视觉印象以及如何保证获得的视觉印象对微小扰动的鲁棒性,都是在设计深度学习算法时需要面临的主要问题。视觉印象深度问题是深度学习的核心问题之一,本文针对视觉印象深度问题的层次特征和稳定特征等方面进行研究,取得的成绩主要包括:第一,本文在视觉印象的基础上,提出了两种视觉印象模型:再认模型和泛化模型用来模拟人类视觉系统的认知过程。本文给出利用一个深度神经网络学习到的视觉印象如何能够被有效地迁移到其他视觉认知任务中。通过复用以非监督的方式训练得到的隐藏层,提出的算法能够在目标任务中大幅度地减少需要被标注图像样本的数量。实验证实了在源任务中估计的参数的确能够帮助网络在目标任务中提高目标分类的结果。第二,本文利用视觉印象给出一个用于训练拓扑深度神经网络的全新的方法。通过结合降噪自动编码器以及带有Hessian正则化项的收缩自动编码器,能够获得一个对输入数据的小幅度变化十分鲁棒的自动编码器。利用切传播算法来展示本文提出的方法如何能够捕获视觉印象的流形结构并且建立一个拓扑图册的图集。然后,利用学习到的特征去初始化一个深度网络,使用相对于其他模型更小的参数集合获得了更好的分类结果。第三,给出了本文开发的一个捕获视觉印象李群流形结构的新算法。通过设计单层李群模型,验证该表示学习算法如何能够被堆叠出一个深层的架构。另外,本文还设计了一个基于李群的梯度下降算法来解决神经网络权重的学习问题。实验结果表明本文提出的方法能够获得更加适用于深度网络的特征并且该特征的计算是十分有效的。综上所述,本文的创新点包括:(1)提出了李群视觉印象深度学习的表示新方法。(2)提出了视觉印象深度学习的度量方法,包括视觉印象深度学习的层次度量、拓扑度量和李群度量。(3)提出了视觉印象深度学习新算法。
[Abstract]:Visual impression is the visual information stored in people's memory. It is an important form of visual cognition. Human sensory stimuli obtained through the vision form the image of the cognitive object after the information processing of the brain. This image is stored in the form of memory in the mind, forming a visual impression that can help human understanding and cognition. Visual impression is the basis for the human brain to accomplish a variety of complex tasks accurately and efficiently. The brain helps the processing of visual information directly by means of the visual impression in memory, rather than by blind calculation. Because of human's cognition of the objective things, more than 80% of the information from the visual and object recognition tasks, people always through the visual impression to the cognition of things, so the study of the visual impression mechanism plays an important role in the field of pattern recognition, computer vision etc.. Deep learning constructs a deep nonlinear network by stacking a single layer module, which can achieve infinite approximation to complex functions. Deep learning can gradually learn the characteristics of different levels of abstraction in an unsupervised form, and provide a more expressive distributed representation. Because of the characteristics of its automatic learning features, it saves the cost of artificial design features compared to the traditional feature selection method. Deep learning is a highly effective feature extraction method, making it a breakthrough in the application of target recognition and other fields. The human visual cortex has a deep structure. From the point of view of the mimic biological mechanism, this has become a strong evidence to support deep learning technology. In a way, deep learning has simulated the cognitive process of human being. The depth structure of the human brain determines the visual impression is the process of the gradual abstraction of the visual information. This means that the visual impression of layer by level abstraction corresponds to the hidden feature extracted by layer by layer. Therefore, the combination of visual impression mechanism and deep learning technology is a potential research direction. How to use the depth learning method to simulate the visual impression produced in the process of human visual cognition is an urgent problem to be solved. This problem requires that the deep learning algorithm designed should embody some characteristics of visual impression, and on the other hand, it can complete the related functions of visual impression. On the basis of deep learning, how to effectively extract hierarchical visual impression and how to ensure the visual impression to gain the robustness of small perturbations are the main problems that need to be faced when designing deep learning algorithm. The visual impression of depth is one of the most important problems of deep learning, aiming at the problem of the visual impression of depth level features and stable characteristics, the main achievements include: first, based on the visual impression, puts forward two kinds of visual impression model: the recognition model and generalization model is used to simulate the process of human cognition the visual system. This paper presents how the visual impression learned by a deep neural network can be effectively migrated to other visual cognitive tasks. By reusing the hidden layer trained in an unsupervised way, the proposed algorithm can significantly reduce the number of image samples needed to be tagged in the target task. The experiment confirms that the estimated parameters in the source task do help the network to improve the result of the target classification in the target task. Second, this paper uses visual impression to give a new method for training topology depth neural network. By combining the noise reduction automatic encoder and the shrinking auto encoder with Hessian regularization term, we can get an automatic encoder which is robust to small changes in input data. The method of cutting propagation is used to show how the method proposed in this paper can capture the manifold structure of visual impression and set up an atlas of a topologic atlas. Then, using the learned features to initialize a depth network, the better classification results are obtained by using a smaller set of parameters relative to the other models. Third, a new algorithm for capturing the Li Qun manifold structure of visual impression is developed in this paper. By designing a monolayer Li Qun model, it is verified how the learning algorithm can be stacked out of a deep architecture. In addition, this paper also designs a gradient descent algorithm based on Li Qun to solve the learning problem of neural network weight. The experimental results show that the proposed method can be more suitable for the characteristics of the depth network and the computation of this feature is very effective. To sum up, the innovative points of this paper include: (1) a new method of expressing Li Qun's visual impression deep learning is proposed. (2) a measure of the depth learning of visual impression is proposed, including the hierarchical measurement, topological metric and Li Qun measure of the depth learning of visual impression. (3) a new algorithm for visual impression depth learning is proposed.
【学位授予单位】：苏州大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41;TP181

【相似文献】