面向视觉特征表达的深度学习算法研究

发布时间：2018-04-01 06:37

本文选题：深度学习　切入点：自编码机　出处：《武汉大学》2017年硕士论文

【摘要】：近年来,随着成像技术的发展,图像、视频等数据越来越丰富,如何从海量数据中提取有效信息也成为了一个难题。深度学习已被证明是有效的解决途径。深度神经网络是一种多层次特征学习模型,能够自动从原始视觉数据中提取出较为抽象的特征,用于后续的图像分类、目标检测等工作。然而,深度神经网络的训练往往需要大量的标记样本,并且模型参数数量庞大,容易过拟合。为了解决这些问题,本文基于前人的模型,提出了改进的深度神经网络模型以及防止过拟合的方法,以提高深度神经网络在图像识别中的性能。本文的创新点如下:(1)提出一种新型的非监督神经网络模型——深度卷积降噪自编码机,从未标记的图像样本中学习有效的特征表达,从而改善了当前主流深度网络训练需要大量标记样本的问题。(2)提出了一种新的正则化方法——结构化去相关约束,能够有效地规范化神经网络模型,防止模型陷入过拟合,同时使得模型学习结构化和不冗余的特征表达,极大地提升了网络模型的特征学习能力和图像分类能力。
[Abstract]:In recent years, with the development of imaging technology, image, video and other data are more and more abundant. How to extract effective information from massive data has also become a difficult problem. Deep learning has been proved to be an effective solution. Depth neural network is a multi-level feature learning model. Abstract features can be automatically extracted from the original visual data for subsequent image classification and target detection. However, the training of depth neural networks often requires a large number of labeled samples and a large number of model parameters. In order to solve these problems, an improved depth neural network model and a method to prevent over-fitting are proposed based on the previous models. In order to improve the performance of depth neural network in image recognition, the innovation of this paper is as follows: 1) A new unsupervised neural network model-depth convolution noise reduction self-coding machine is proposed, which can learn effective feature expression in unmarked image samples. Therefore, the problem of large number of tagged samples is improved in the training of current mainstream deep network. A new regularization method, structured decorrelation constraint, is proposed, which can effectively standardize the neural network model and prevent the model from falling into overfitting. At the same time, it makes the model learning structure and non-redundant feature representation, greatly improve the network model feature learning ability and image classification ability.
【学位授予单位】：武汉大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP18

【相似文献】