稀疏表示在图像分类问题中的应用研究

发布时间：2018-05-18 10:09

本文选题：稀疏表示 + 重构样本　；参考：《安徽大学》2016年硕士论文

【摘要】：随着互联网与智能终端的迅速普及,图像数据迅速增长。由于图像数据未经全部标注,导致图像的分类存储和查找不方便,如何将海量的图像有效地进行分类,已经成为图像处理研究领域的一个热点方向。稀疏表示应用在图像分类问题中是图像处理中的一个重要分支。稀疏表示以其简洁模型迅速吸引了大量学者的注意,通过稀疏表示求解的系数向量,在每个子类字典中重构样本,然后记录测试样本减去所有重构样本的残差,找到最小的残差对应的子字典的类,作为测试样本的预测标签,这种图像分类的方法直观,对于有较好的字典样本,通过系数就能直观的观察到测试样本与哪些类相关。本文对稀疏表示在图像分类问题中的应用进行了深入研究,所做的主要工作和特点如下：(1)提出一种基于重构样本稀疏表示的图像分类算法。算法是通过产生一个重构的中间样本进行分类。针对现有的基于稀疏表示的图像分类中存在的不是所有信息都可以用来分类的问题,提出了一种将测试样本投影到训练样本上,丢弃训练样本中没有表示出来的信息,再用重构之后的样本来分类,解决了不能被训练样本表示的信息对分类的影响。为了取得更优秀的结果,实验中所有的训练样本和测试样本都先通过Gabor滤波器提取局部特征。实验验证了提出的算法具有有效性。(2)提出一种结合字典对和重构样本的稀疏表示应用于人脸图像分类的算法。对于图像的分类与识别来说,获得具有良好区分能力的字典是至关重要的。针对现有的基于稀疏表示的人脸图像分类中,求解具有区分能力字典的复杂度过高以及测试样本不能被字典全部表示出来的问题,提出了结合字典对学习与样本重构的稀疏表示用于图像分类。对于训练样本和测试样本在训练字典之前,利用Gabor滤波器提取不同方向,不同尺度的特征作为训练字典的原始输入,然后使用解析字典来辅助获得具有区分能力的综合字典,再将测试样本投影到获得的字典上,最后用基于综合字典上重构的样本来分类,减少了不能被字典表示的信息对分类的影响,提升分类的准确率。实验结果表明,在三个具有代表性的人脸数据集中,DPL ICSRIR方法能较好地分类人脸图像。
[Abstract]:With the rapid popularity of the Internet and intelligent terminals, image data is growing rapidly. Because the image data is not annotated completely, it is not convenient to store and find the image classification. How to classify the massive images effectively has become a hot topic in the field of image processing. Sparse representation is an important branch of image processing in image classification. The sparse representation has attracted the attention of a large number of scholars because of its concise model. Through sparse representation of the solved coefficient vector, the samples are reconstructed in each subclass dictionary, and the test samples are recorded to subtract the residuals of all the reconstructed samples. The method of image classification is intuitionistic to find the class of sub-dictionary corresponding to minimum residual error as the prediction label of test sample. For a good dictionary sample, we can intuitively observe which classes the test sample is related to through the coefficient. In this paper, the application of sparse representation in image classification is deeply studied. The main work and characteristics are as follows: 1) an image classification algorithm based on reconstructed sparse representation of samples is proposed. The algorithm is classified by generating a reconstructed intermediate sample. In order to solve the problem that not all the information can be used to classify the existing image classification based on sparse representation, a test sample is projected onto the training sample and the information that is not represented in the training sample is discarded. The effect of the information that can not be represented by the training samples on the classification is solved by using the reconstructed samples. In order to obtain better results, all the training samples and test samples are firstly extracted by Gabor filter. Experimental results show that the proposed algorithm is effective. (2) A new algorithm for face image classification is proposed, which combines dictionary pairs with sparse representation of samples. For image classification and recognition, it is very important to obtain dictionaries with good distinguishing ability. In the existing face image classification based on sparse representation, the complexity of the discriminative dictionary is too high and the test samples can not be completely represented by the dictionary. A sparse representation of learning and sample reconstruction is proposed for image classification. For training samples and test samples, before training dictionaries, Gabor filters are used to extract features of different directions and scales as the original input of the training dictionaries, and then analytic dictionaries are used to assist in obtaining comprehensive dictionaries with distinguishing ability. Then the test samples are projected onto the obtained dictionaries, and finally the samples reconstructed on the synthetic dictionaries are used to classify, which reduces the influence of the information that can not be represented by the dictionary on the classification, and improves the accuracy of classification. The experimental results show that the DPL ICSRIR method can be used to classify face images in three representative face datasets.
【学位授予单位】：安徽大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】