面向图像识别的稀疏模型研究

发布时间：2018-04-01 22:13

本文选题：图像识别　切入点：稀疏模型　出处：《北京邮电大学》2015年博士论文

【摘要】：随着移动互联网和数字多媒体等技术的不断发展,以图像为载体的交流方式越来越多,因此图像识别技术显得愈发重要。稀疏模型作为神经生理学、计算机科学和人工智能的交叉研究领域,近年来在图像识别领域内发展迅速,已成为图像处理和计算机视觉等领域内的研究热点和难点。首先,本文对稀疏模型中的字典学习问题、稀疏编码问题和稀疏表示的降维问题进行了深入研究,并提出一系列算法；然后,将这一系列算法代入基于稀疏模型的图像识别系统中,并通过实验验证了它们的有效性。本文的主要工作和创新点如下：1.基于图像识别中稀疏模型的结构化特点和字典的判别性假设,本文提出了一种块稀疏判别字典学习算法。首先,该算法利用块稀疏的特点确定稀疏表示中非零项的位置,简化了系数求解的步骤；其次,引入不同类别子字典的相关性约束,增加类间距离并增强字典的判别性；再次,本文提出在测试阶段用结构化稀疏编码算法代替传统稀疏编码算法；最后,本文给出该算法基于梯度下降的优化策略。在人脸数据集上的实验结果表明块稀疏判别字典学习算法和结构化稀疏编码算法能提高图像识别系统的性能。2.利用同类信号间的强相关性,本文提出了块约束低秩字典学习算法和低秩局部稀疏编码算法。首先,块约束低秩字典学习算法利用块稀疏的特点限制稀疏表示中非零项的位置,并引入低秩约束对样本间的相关性建模；其次,低秩局部稀疏编码算法将训练样本集加入测试样本的系数求解过程中,避免当测试样本过少时出现无法求解系数的问题；最后,本文还提出了块约束低秩字典学习算法和低秩局部稀疏编码算法的优化流程。通过人脸识别实验和物体识别实验证明块约束低秩字典学习算法和低秩局部稀疏编码算法能有效地提高算法的识别准确率。3.根据图像识别中稀疏表示的块结构化特点,本文提出了一种基于图的块稀疏降维算法。首先,提出块稀疏表示距离计算公式,该距离对稀疏表示中同类分量和不同类分量采用不同的度量方式,减少了稀疏表示中由于同类分量的非零项位置没有对齐带来的误差,同时增强了算法的判别性；其次,基于图的块稀疏降维算法使用块稀疏表示距离构建样本间的邻接矩阵,并以保留样本在一定邻域内互相之间的距离关系为准则进行非线性降维；最后,本文还给出了基于图的块稀疏降维算法的计算步骤。在人工模拟数据集和人脸数据集上的实验结果证明基于图的块稀疏降维算法能有效降低识别错误率。4.针对非线性降维算法的“新样本扩展问题”,本文提出了一种基于图的块稀疏线性降维算法,并将其与块约束低秩字典学习算法和低秩局部稀疏编码算法相结合,提出基于判别性块稀疏与低秩的图像识别算法。首先,将基于图的块稀疏降维算法线性化,提出基于图的块稀疏线性降维算法,使其更适合实际的图像识别系统；然后,将块约束低秩字典学习算法、低秩局部稀疏编码算法和基于图的块稀疏线性降维算法代入基于稀疏模型的图像识别系统中,得到基于判别性块稀疏与低秩的图像识别算法。该算法在人脸识别实验和物体识别实验上的结果验证了本文创新点的有效性。
[Abstract]:With the continuous development of mobile Internet and digital multimedia technology, based on image more and more ways to communicate, so the image recognition technology becomes more and more important. As a sparse model of neurophysiology, interdisciplinary research field of computer science and artificial intelligence, in recent years in the field of image recognition in the rapid development, has become a research hotspot and difficulty in the field of image processing and computer vision in this paper. Firstly, the problem of learning in the dictionary in the sparse model, studied the problem of dimensionality reduction problem of sparse encoding and sparse representation, and proposed a series of algorithms; then, it will be a series of algorithms based on sparse model into image recognition system, and their effectiveness. By experiments. The main work and innovations are as follows: 1. the structural features and image recognition dictionary sparse model discrimination based on sex Assumptions, this paper proposes a block sparse discriminant dictionary learning algorithm. Firstly, the algorithm uses the characteristics of block sparse sparse representation in determining non-zero position, simplifies the steps of solving the coefficient; secondly, correlation constraints introduce different class sub dictionaries, increase the distance between the class and enhance the discriminative dictionary; again. This paper proposes the use of structured sparse encoding algorithm to replace the traditional sparse encoding algorithm in the test stage; finally, this paper gives the algorithm optimization strategy based on gradient descent. In the face data sets. The experimental results show that the block sparse discrimination dictionary learning algorithm and structured sparse encoding algorithm can improve the utilization of the strong correlation between signals of similar performance of.2. image recognition system in this paper, a low rank constraint dictionary learning algorithm and low rank local sparse encoding algorithm. Firstly, a low rank constraint learning algorithm uses the block dictionary The characteristics of restricted sparse sparse representation of the nonzero entries in the position, and the introduction of low rank constraint on the correlation between the modeling samples; secondly, the process of solving the coefficients of low rank local sparse encoding algorithm will join the training sample set of test samples, when the test sample is too little to avoid be coefficients problem; finally, this paper also put forward the optimization process of the low rank constraint block dictionary learning algorithm and low rank local sparse encoding algorithm. Through the experiments on face recognition and object recognition experiments show that the low rank constraint block dictionary learning algorithm and low rank local sparse encoding algorithm can effectively improve the recognition accuracy of.3. according to the characteristics of block structured sparse image recognition in this paper said, proposes a dimensionality reduction algorithm based on block sparse graph. Firstly, the block sparse representation formula of distance, the distance of the sparse representation of similar and different components The component type measure using different methods, reduce the sparse representation because of similar components non zero position error is not aligned to bring, and enhances the discrimination of the algorithm; secondly, block sparse graph based dimensionality reduction algorithm using block sparse representation constructed from samples of the adjacency matrix, and to keep the samples in a certain in the neighborhood of the distance relationship between each other as a criterion for nonlinear dimensionality reduction; finally, this paper gives the calculation steps of dimensionality reduction algorithm based on block sparse graph. In the artificial data set and face data sets. The experimental results show that based on graph block sparse dimensionality reduction algorithm can effectively reduce the recognition error rate for nonlinear.4. dimensionality reduction algorithm "new extended sample", this paper proposes a dimensionality reduction algorithm based on sparse linear block diagram, and the low rank constraint block dictionary learning algorithm and low rank local sparse coding The combination of code algorithm, proposed an image recognition algorithm of distinguishing block sparse and low rank based. First, the block sparse graph based dimensionality reduction algorithm of linear dimensionality reduction algorithm is proposed based on sparse linear block diagram, the image recognition system is more practical; then, the low rank constraint block dictionary learning algorithm low rank, local sparse encoding algorithm and block sparse linear graph based dimensionality reduction algorithm based on image recognition system into sparse model, image recognition algorithm of distinguishing the block sparse and low rank based on the recognition algorithm. The experimental results verify the effectiveness of this innovation in face recognition experiments and objects.

【学位授予单位】：北京邮电大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：TP391.41

【相似文献】