【作者】 张翼；

【导师】 吴洪；

【作者基本信息】 电子科技大学 ， 计算机软件与理论， 2013， 硕士

【摘要】 随着数字医学图像技术的发展，近十年来，医院每天采集的图像数量呈现出爆炸增长趋势。如何从这些大量的数据中检索出需要的信息是一个迫切需要解决的问题。通过把医学图像进行分类，可以有效的改善图像的检索性能。传统采用人工对图像进行分类的方法耗时耗力，因此利用计算机来对这样巨大的图像数据集进行自动分类成为一个重要的研究问题。目前，基于局部特征的图像分类技术因其具有良好的性能得到了广泛的应用，其分类流程主要包含如下环节：局部特征提取、词典构建、根据词典对图像进行编码、分类器的训练。本文首先对基于局部特征的图像分类现状进行了回顾，并对分类方法的一些基本理论做了介绍。比较了不同局部特征采样方式以及视觉单词分配方法对分类精度的影响。为了适应在更大规模数据集上的应用，分析和比较了几种能提高词典构建速度以及单词查询效率的改进K均值聚类算法。考虑到稀疏编码对局部特征更好的重构性，本文分析和比较了传统稀疏编码以及其改进的局部约束线性编码方式在医学图像分类任务中的性能。本文的主要贡献包括：1.比较了Patch和SIFT局部特征在几种不同采样方式以及视觉单词软分配和硬分配对分类精度的影响。实验表明采用稠密网格抽取SIFT特征以及视觉单词软分配的方式在医学图像分类任务中达到了最好的性能。2.分析和比较了传统K均值聚了算法、层次K均值聚类算法、近似K均值算法在词典构建上的速度，以及所构建的词典在查找效率上的区别。实验表明采用层次K均值算法和近似K均值算法构建的词典能极大的提高词典的查找效率，从而能适应大规模数据应用的要求。并且近似K均值聚类算法比层次K均值聚类算法构建的词典查找速度更快。3.分析稀疏编码以及稀疏编码的改进方法——局部约束线性编码，并与基于传统K均值的图像编码方法进行比较。实验表明稀疏编码和局部约束线性编码在支持向量机中使用线性核函数就能达到基于传统K均值编码方法使用非线性核函数的分类精度，并且局部约束线性编码的编码效率要远远高于稀疏编码。更多还原

【Abstract】 In the last ten years, there has been an explosion in the number of images acquired every day in modern hospital. How to retrieve the relevant information from the large amount of data is an urgent problem that need to solve. Medical image classification is a effective approach to improve the performance of image retrieval. Traditional manual medical image classification is time-consuming. Hence, automatic medical image classification become an important research problem.At present, local features based image classification has been widely used because of its good performance. Local features based medical image classification mainly consists of the follow steps:local features extraction, visual vocabulary construction, image encoding, and classifier training. At the beginning, this paper reviews the state of the art of medical image classification based on local feature, and introduce some basic theory of classification method. Then, we compare some different sampling methods of local features and assign method of visual words in medical classification. After that, for apply to lager dataset, several improved K-Means clustering algorithms are analyzed and compared, which can improve the efficiency of dictionary construction and query. Finally, sparse coding and one of its improved variants are analyzed and used for medical image classification.In this paper, the main contributions include:1.Different sampling methods for Patch and SIFT features are compared by experiments, and also are different assignment methods of visual words. The experimental results show that dense-grid sampled SIFT feature with soft assignment can get the best performance among the compared methods in medical image classification task.2. Several K-Means based clustering algorithms are analyzed, which are traditional K-Means, hierarchical K-Means and approximation K-Means. Their efficiency on dictionary construction and search are compared by experiments. The experimental results indicate that hierarchical K-Means and approximation K-Means can greatly improve the efficiency of dictionary construction and searching, thus can adapt to lager dataset. In addition, approximate K-Means can get higher search speed than hierarchical K-Means.3.Sparse coding and its variants methods--locality-constrained linear coding are analyzed and compared with traditional K-Means for image coding. The experimental results show that the two sparse coding methods can get better performance with linear SVM than K-Means with non-linear SVM.更多还原