基于互信息的多通道联合稀疏模型及其组织病理图像分类

Mutual Information-based Multi-channel Joint Sparse Model for Histopathological Images Classification

针对传统联合稀疏模型中共有分量与独有分量都采用相同的字典进行特征表示,导致编码系数判别性低的问题,提出一种基于互信息的多通道联合稀疏模型,并将其应用于组织病理图像的分类.该模型通过K均值对样本特征进行聚类,分别得到R,G与B通道的字典;其次利用样本特征与3个字典之间的互信息,剔除弱相关原子且构造了1个共有字典与3个独有字典,以此为基础建立了多通道联合稀疏模型;同时引入图像的空间信息,结合空间金字塔匹配模型对不同层次的图像特征进行联合稀疏编码,利用编码系数训练SVM分类器.实验结果表明,该模型具有更好的特征表示能力,大大提高了编码系数的判别性,获得了较好的分类性能与较强的鲁棒性.

In the traditional joint sparse model, a dictionary was used for feature representation of the commonor unique components. This leads to the low discrimination of the sparse coding coefficients. In thispaper, mutual information-based multi-channel joint sparse model is proposed for histopathological imageclassification. The training samples are clustered into R, G and B channel dictionaries by using K-means. Byexploring the mutual information between training samples and three dictionaries, the irrelevant atoms aredeleted, meanwhile, a shared dictionary and three unique dictionaries constructed. Simultaneously,multi-channel joint sparse model is designed based on the shared dictionary and three unique dictionaries.Furthermore, in order to represent image feature of different levels, the spatial pyramid matching is used tothe multi-channel joint sparse coding. Finally, the joint sparse coding coefficients are used to train the SVMfor histopathological images classification. The experimental results show that the proposed model haspower feature representation ability and improve greatly the discrimination of coding coefficients. Thus thebetter classification performance and the power robustness can be obtained with compared to the traditionalmodels.