基于子空间学习和稀疏编码的图像分类算法研究

发布时间：2018-01-05 01:03

本文关键词：基于子空间学习和稀疏编码的图像分类算法研究　出处：《北京交通大学》2016年硕士论文　论文类型：学位论文

【摘要】：随着多媒体和互联网技术的迅速发展,图像信息呈现出爆炸式的增长,如何在海量数据中快速寻找到所需要的数据已成为亟待解决的问题。图像分类技术通过对图像数据进行分析处理,自动检索分类出图片,满足人们的需求。因此,对图像分类的研究具有十分重要的理论意义和应用价值。本文使用图像的SIFT特征和像素特征,分别从高维数据降维、寻优子空间以及重构误差分类的角度对图像分类进行了细致深入的研究,主要工作包括：(1)图像分块可以更好的利用图像像素间的空问信息,获得更多的细节。因此,我们提出一种基于分块子空间的图像分类算法。并针对传统的数学推导无法获得分类问题的精确解的问题,提出了通过粒子群算法求解的子空间寻优模型。此方法使用粒子群算法对分块的子空间模型进行寻优处理,最终得到一个优化的子空间,并实现图像分类；(2)为了克服高维数据对于图像分类带来的分类精确度低、计算复杂度高的问题,提出了一种基于稀疏编码的正交平滑子空间的图像分类算法,将图像从高维空间变换到低维空间。此算法首先提取图像的SIFT特征,对此特征进行稀疏编码,再对编码矩阵进行最大池特征提取,最终获得对图像的一个高维特征描述,最后通过正交平滑子空间算法将这些高维的特征向量在正交平滑子空间中降维处理。实验验证了所提算法的有效性；(3)基于稀疏编码的图像分类算法,在训练字典的过程中并没有充分利用选取的训练数据的类别信息,这也就使得字典是无监督的。为了充分利用已知的先验信息,提出了一种有监督的重构误差分类算法。通过在字典训练过程中加入类别信息的方法获得更有效的字典描述,并采用重构误差最小化准则对图像进行分类,实验结果表明,该算法可以有效的改善分类性能。
[Abstract]:With the rapid development of multimedia and Internet technology, image information shows an explosive growth. How to quickly find the needed data in the massive data has become an urgent problem. Image classification technology through image data analysis and processing, automatic retrieval and classification of images. Therefore, the research of image classification has very important theoretical significance and application value. In this paper, the SIFT feature and pixel feature of image are used to reduce the dimension from high-dimensional data. The optimization subspace and reconstruction error classification are studied in detail. The main work includes: 1) Image segmentation can make better use of spatial information between pixels. Therefore, we propose an image classification algorithm based on block subspace, and aim at the problem that traditional mathematical derivation can not obtain the exact solution of the classification problem. A subspace optimization model based on particle swarm optimization (PSO) algorithm is proposed in this paper. The PSO algorithm is used to optimize the sub-space model of blocks, and finally an optimized subspace is obtained and the image classification is realized. 2) in order to overcome the problems of low classification accuracy and high computational complexity brought by high-dimensional data for image classification, an image classification algorithm based on sparse coding in orthogonal smooth subspace is proposed. The image is transformed from high-dimensional space to low-dimensional space. Firstly, the SIFT feature of the image is extracted, the feature is sparse coded, and then the maximum pool feature is extracted from the coding matrix. Finally, a high-dimensional feature description of the image is obtained. Finally, these high-dimensional feature vectors are reduced in the orthogonal smoothing subspace by orthogonal smoothing subspace algorithm. The experimental results show the effectiveness of the proposed algorithm. 3) the image classification algorithm based on sparse coding does not make full use of the class information of the selected training data in the process of training dictionary. This makes dictionaries unsupervised in order to take full advantage of known prior information. A supervised reconstruction error classification algorithm is proposed to obtain a more effective dictionary description by adding category information into the dictionary training process, and the reconstruction error minimization criterion is used to classify images. Experimental results show that the proposed algorithm can effectively improve the classification performance.
【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】