基于神经反应模型和稀疏表示的图像分类算法研究

发布时间：2018-02-11 17:28

本文关键词： 图像分类神经反应模型稀疏表示特征提取多层结构半监督学习鲁棒性　出处：《华中科技大学》2016年博士论文　论文类型：学位论文

【摘要】：图像分类是机器学习和计算机视觉领域中一个非常重要的研究内容,在国防安全、工业信息化、医学工程和互联网科技等方面均有广泛的应用。由于图像易受光照条件、拍摄角度、复杂背景、旋转变换或尺度放缩等因素的影响,从而对图像进行分类是一个非常具有挑战性的课题。因此,研究高效而具有鲁棒性的图像分类算法,具有重要的理论意义与实际应用价值。提取图像具有代表性的特征,是图像分类过程中的一个关键部分。为了从复杂的图像中提取有效的图像特征,以及克服训练数据不足的现实问题,本文基于神经反应模型和稀疏表示方法,着重研究了具有鲁棒性和区分性的图像分类方法,提出了一系列算法,并取得了较好的分类效果。论文的主要内容如下：首先,本文提出了一种基于稀疏表示的神经反应算法。该算法是一个多层的结构,试图模拟人类大脑皮层中的视觉处理机制,通过交替进行非负稀疏表示和最大化联合操作来构建。非负稀疏表示能够提取图像中的显著特征,而最大化联合操作使得模型对于平移等变换具有不变性。为进一步提高算法的性能,本文依据算法的结构特点,还设计了两种简单而有效的模板选取方法。通过仿真实验,验证了该算法能大幅提高原始的神经反应模型的性能,并在复杂图像的分类问题中取得较好的效果。然后,通过引入ELM (extreme learning machine)的理论，，本文提出了一种基于神经反应模型的多层结构算法。该算法主要包含两个阶段,即多层的ELM特征映射阶段和ELM学习阶段。其中,多层的ELM特征映射阶段是一个多层的结构,它是以递归的方式通过交替进行特征图构建和最大化联合操作来建立的。特别地,我们使用随机生成的输入权重构建特征图,且无需对权重参数进行优化调节,这使得模型的结构更加简单,有更高的计算效率。在ELM学习阶段,本文提出了一种基于弹性网正则化约束的优化模型,为ELM方法学习输出权重,并相应地设计了问题的优化算法。弹性网正则化约束能够为ELM方法学习更紧致和有选择性的输出权重,这将有益于ELM方法的特征学习过程。实验结果显示,与传统的深度学习方法相比,该算法能够在获得更高分类精度的前提下,花费更少的计算时间。最后,针对实际问题中训练数据不足的情况,本文提出了一种半监督的基于图正则化约束的稀疏表示算法。该算法能够充分利用已标记数据和未标记数据的信息,学习未标记数据的类别信息。学习得到的类别信息,不仅能够保持与原始数据相同的流形结构,而且具有较好的区分性,有利于对图像进行分类。为了充分挖掘未标记数据中隐含的结构信息,通过赋予邻域内未标记数据额外的权重,定义了一种新的类间相似度矩阵和类内相似度矩阵,从而可获取数据中更具区分性的特征。为了处理线性不可分的图像数据,本文还提出了一种基于图正则化约束的核稀疏表示方法,可在高维空间中对数据进行线性分类。相应地,本文为提出的稀疏表示方法及其核方法,设计了两种高效的优化算法。实验结果表明,算法在处理训练数据不足的图像分类问题时,仍然能够取得较好的效果。
[Abstract]:Image classification is a very important research field in computer vision and machine learning, in national security, information industry, has been widely used in medical engineering and Internet technology. The images are easily affected by illumination conditions, shooting angle, complex background, influence of rotation transform or scale factors, thus the image classification is a very challenging task. Therefore, the image classification algorithm based on efficient and robust, and has important theoretical meaning and practical value. With the representative feature extraction of image, is a key part of the image classification process. In order to extract the image feature effectively from complex in the image, and overcome the problem of insufficient training data, the neural response model and sparse representation method based on image, focuses on the robust and discriminative points Methods, put forward a series of algorithm, and achieved good classification results. The main contents of this paper are as follows: firstly, this article presents a neural response algorithm based on sparse representation. The algorithm is a multilayer structure, trying to simulate the human brain at the visual cortex of the mechanism, through alternating non negative the sparse representation and maximize the joint operation. To construct the non negative sparse representation can extract salient features of the image, and maximize the joint operation so that the model is invariant to translation. In order to further improve the performance of the algorithm, based on the structural characteristics of the algorithm, designs two kinds of simple and effective method. Through the template selection simulation results show that the algorithm can greatly improve the performance of the neural response of the original model, and achieved good results in the classification of complex images. Then, through the introduction of ELM (e Xtreme learning machine) theory, this paper proposes a multi-layer structure model based on the neural response algorithm. The algorithm mainly includes two stages, namely the multilayer ELM feature mapping stage and ELM learning stage. Among them, ELM feature mapping stage multilayer is a multi-layer structure, it is in a recursive way through alternate feature map construction and maximize the joint operation to establish. In particular, we use randomly generated input weights to construct the feature map and there is no need to optimize the adjustment of the weight parameter, which makes the model more simple structure and has high efficiency. The learning phase in ELM, this paper proposes an optimized model of elastic net regularization constraint based learning output weights for the ELM method, and the corresponding design optimization problem. The elastic net regularization constraint for ELM learning method is more compact and selective The output weights, which will feature learning is beneficial to ELM process. The experimental results show that compared with the traditional method, the algorithm can deep learning, while maintaining the higher classification accuracy, computational cost less time. Finally, according to the actual problems in the training data insufficiency, this paper proposes a semi supervised the sparse representation algorithm based on graph regularization constraint. The algorithm can make full use of data and unlabeled data information has been marked, unmarked categories learning information. Learning class information is obtained, which can not only keep the original data and the manifold structure of the same, but also has high discriminability to classify. In order to fully tap the image. The unmarked structure information hidden in the data, the unlabeled data additional weight to the neighborhood, we define a new similarity matrix between class and within class. The similarity matrix, which can obtain the data in more distinguishable features. In order to image data processing linear inseparable, this paper also proposes a representation method of kernel sparse graph regularization based on constraints, can be a linear classification of data in high dimensional space. Accordingly, the representation method and kernel method for the proposed sparse, two efficient optimization algorithm. The experimental results show that the algorithm in image classification problem of insufficient training data, still can achieve better results.

【学位授予单位】：华中科技大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】