当前位置: 首页>博士论文>资源详情
基于稀疏与低秩模型的图像表达与分类研究
中文摘要

图像表达与分类是模式识别、计算机视觉以及图像处理等领域的重要研究内容,同时也是诸多计算机视觉与图像处理课题如图像理解、目标检测与跟踪以及图像检索等的研究基础。传统的机器学习方法在处理高维度、结构复杂的图像数据时面临着困难,稀疏与低秩模型是近年来涌现的研究热点,已在计算机视觉与机器学习领域得到成功应用。本文以子空间学习、图半监督学习以及超限学习机为载体,围绕稀疏与低秩模型在图像表达与分类中的应用开展研究工作,论文的主要工作和创新点如下: (1)基于鲁棒加权群稀疏表示的特征提取。 传统的稀疏表示存在无监督、随机选择以及鲁棒性不强等问题,这将降低基于稀疏表示的特征提取方法提取特征的鉴别力,影响分类准确率。针对这些问题,本文首先提出一种有监督的鲁棒加权群稀疏表示模型,该模型通过最小化l₁范数约束的表示误差以及加权l〓范数约束的表示系数,能够有效利用数据的标签信息,得到对噪声鲁棒的群稀疏表示。该表示可自适应地确定样本的类内、类间近邻及彼此间的相似度,利用该表示构造类内图与类间图,在图嵌入框架下学习低维鉴别子空间,提取高维图像数据的低维鉴别特征。在物体以及人脸图像数据集上的实验结果表明,与同类型方法相比,本文方法对噪声和遮挡更为鲁棒,利用本文方法提取的特征能够取得较高的分类准确率。 (2)基于块对角低秩稀疏表示的特征提取。 稀疏表示可用于揭示数据的局部几何结构关系,而难以有效发掘数据的全局结构特性,这将影响基于稀疏表示的特征提取方法提取特征的鉴别力以及最终的分类准确率。针对这一问题,本文提出一种块对角低秩稀疏表示模型,该模型能够揭示数据的局部几何以及全局多子空间的结构特性,同时块对角约束的引入可有效利用数据的标签信息。以样本重建为准则,由该模型得到的表示系数能够准确反映同类样本之间的相似性以及异类样本之间的差异性,获得具备块对角结构的低秩稀疏表示。进一步利用该表示指导低维空间的学习,使得同类数据紧凑聚集,异类数据相互分离,得到高维数据富有鉴别力的低维特征。与同类型方法相比,本文方法能够提取数据更具鉴别力的特征,取得较高的分类准确率。 (3)基于标签与局部约束的低秩图学习。 针对现有的图学习方法未有效利用数据的标签与局部信息这一问题。本文在图学习过程中引入低秩约束,同时利用带标签数据的标签信息以及数据间的局部信息来指导图的构建,提出一种标签与局部约束的低秩图学习模型。该模型能够有效地利用带标签数据的标签信息,而数据局部信息的引入使得构造的图具备稀疏性。因此,构造的图兼有低秩性、标签引导以及局部保持性(稀疏性)的特点。本文进一步将该低秩图学习模型应用于图半监督分类问题中。实验结果表明,由该模型得到的图在标签以及局部信息的引导下,能够准确揭示样本间的相似关系,使得标签信息能够高效地从带标签数据传播至无标签数据,完成半监督分类任务。 (4)基于神经元剪枝的鉴别超限学习机及其分层学习。 该部分研究了将超限学习机应用于图像特征学习与分类的三个问题。①超限学习机网络的泛化能力提升问题。提出一种鉴别超限学习机模型,该模型通过引入一个非负的标签松弛矩阵将传统超限学习机中严格固定的目标标签矩阵松弛为可学习的灵活标签矩阵,该策略能够为网络输出权重矩阵的学习提供更大的自由度,尽可能地扩大不同类别样本之间的距离以充分挖掘样本内蕴含的鉴别信息。实验结果表明,该策略有助于提高网络的泛化能力以及分类准确率。②超限学习机网络结构的设计问题。以鉴别超限学习机为基础,提出一种基于神经元剪枝的鉴别超限学习机模型。将行稀疏约束引入鉴别超限学习机输出权重矩阵的学习中,可以区分不同隐层神经元在信息处理过程中的重要程度,通过剪除无价值的神经元可自适应地确定隐层所需神经元,得到更为紧致的网络。实验结果表明,该模型可以在保持甚至提高分类准确率的同时,降低网络隐层神经元的存储需求以及预测时间复杂度。⑧基于多层超限学习机的特征学习与分类问题。基于标签松弛策略和行稀疏约束,提出一种用于图像特征学习与分类的基于神经元剪枝的多层鉴别超限学习机模型。首先利用行稀疏约束,构造了一种基于神经元剪枝的超限学习机自编码器,该自编码器可自适应地确定隐层所需神经元,通过逐层学习、堆叠该自编码器进行无监督的深度特征学习。在对特征进行有监督的分类模型学习时,采用标签松弛策略学习输出权重矩阵,有助于提高整个网络的泛化能力以及分类准确率。 关键词:图像表达,图像分类,稀疏模型,低秩模型,超限学习机

英文摘要

Image expression and classification are two of the most important issues in computer vision, pattern recognition and image processing. They are also the fundamental modulars of many computer vision and image processing tasks, such as image understanding, object detection and tracking, and image retrieval, which can directly affect the performances of successive works. The traditional machine learning methods are confronted with difficulties in dealing with high dimensional and complex data. Sparse and low rank models have recently emerged as advanced techniques in the field of computer vision and machine learning. Based on subspace learning, graph based semi-supervised learning and extreme learning machine, we apply sparse and low rank models for the problems of image expression and classification. The main contributions and results of this thesis are summarized as follows: (1)Image feature extraction based on robust weighted group sparse representation Traditional sparse representation model suffers from the problems of unsupervision, random selection and weak robustness. These problems might results in less discriminative feature obtained by sparse representation based feature extraction methods, based on which unsatisfied classification performance might be achieved. On these problems, a supervised robust weighted group sparse representation model is proposed by minimizing the combination of l₁ norm regularized representation fidelity and weighted l〓 norm regularized representation coefficients. The model can harness the label information to get stable and robust group sparse representation, and adaptively determine the intra-class and inter-class neighbors and the similarity relationships for each sample. Low dimensional subspace is learnt using the intra-class and inter-class graphs derived from the group sparse representation to get discriminative low-dimensional subspace, based on which low-dimensional and discriminative feature of high dimensional image data can be calculated. Experimental results show that the proposed method is more robust to noise and occlusions, and the feature extracted by our method can achieve promising classification accuracy in comparison with related methods. (2)Image feature extraction based on block-diagonal constrained low rank and sparse representation Spare representation can only reveal the local geometry structure of data, and cannot discover the global structure relationship among data. The drawback might potentially degrade the discriminability of obtained feature and classification accuracy. For the problem, a block-diagonal constrained low rank and sparse representation model is developed by taking the merits of both sparse and low rank representation models. The developed model can effectively discover both local geometry and global multi-subspace structures in data. Besides, the introduction of the block-diagonal constraint can utilize the label prior. Under the principal of sample representation, the representation coefficient obtained by the model can highlight interclass differences and enhance inter-class similarities to get block-diagonal low rank and sparse representation. The representation is further utilized to learn low dimensional subspace, where data points from the same class gather and data points from different classes separate. Compared with other methods, the feature extracted by our method is more discriminative with higher classification accuracy. (3)Label and locality constrained low rank graph learning On the problem that label and locality information is not well exploited in existing graph learning methods, a graph learning method is presented by fusing low rank constraint, label information and locality information of data. The model can effectively harness the precious label information, and the introduction of locality information will make the graph be sparse. As a result, the optimized graph has the traits of low rank, label guiding, and locality preserving (sparsity). The proposed model is further applied for graph based semi-supervised classification problem. Experimental results show that the obtained graph can well reveal the affinity relationships of samples, and the label information can effectively propagate from the labeled data to the unlabeled ones to fulfill the task of semi-supervised classification. (4)Neuron-pruning based discriminative Extreme Learning Machine and its hierarchical learning This section studies three problems of Extreme Learning Machine (ELM) for the problems of image feature learning and classification. ① The improvement of generalization ability of ELM network. A discriminative Extreme Learning Machine model (DELM) is developed. DELM introduces a nonnegative label slack matrix to relax the strict and fixed target label matrix of traditional ELM to be a leamable and flexible label matrix. The strategy can provide more freedom for the learning of output weights matrix to enlarge the distance between transformed inter-class samples.Experimental results show that this strategy can better discover the discriminative information in data, and improve the generalization ability of the obtained network with higher classification accuracy. ② Structure design of ELM network. Based on DELM, a neuron pruning based DELM model is developed by restricting the output weights matrix to be row spare with l〓 norm regularization. The strategy can distinguish the importance of different neurons in information processing and prune the worthless ones. As a result, the neurons needed in hidden layer can be adaptively determined for a more compact network. Experimental results show that the proposed model can reduce both the storage needs for neurons and the prediction time with comparative or even better classification performance. ③ Feature learning and classification based on multilayer ELM. A multilayer extreme learning machine is presented by adopting row sparse constraint and label relaxation strategy. An extreme learning machine based auto-encoder is developed with row sparsity constraint to adaptively determine the neurons needed in hidden layer. With this manner, the auto-encoder is learnt and stacked layer by layer to get deep neural network for unsupervised feature learning. For the supervised classification of obtained feature, the label relaxation strategy is employed to calculate the output weight matrix. Experimental results show that the obtained network is light and compact with promising classification performance. Key words: image expression, image classification, sparse model, low rank model, extreme learning machine

作者相关
主题相关
看过该书的人还在看哪些书