结合视觉语义的乳腺X线图像检索

Mammogram retrieval combining visual semantic

针对乳腺X线图像检索过程中底层特征不能有效表达高层语义,提出一种结合视觉语义的图像检索算法。引入图半监督学习框架提取查询图像的类别归属度,作为视觉语义;考虑类别分布的不平衡性,在语义提取过程中加入类标签正则化;设计结合视觉语义和底层特征的相似度度量准则。在乳腺X线分块图像数据库上的实验结果表明,该算法能够有效提取图像的视觉语义,性能优于基于单一底层特征的检索算法。

As low level features can not reflect the high level semantic in mammogram retrieval, an image retrieval algorithm is proposed combining visual semantic. Firstly, a graph-based semi-supervised learning frame is applied to obtain the membership degree of the query images as the visual semantic. The class label regularized method is added to semantic extraction procedure due to the imbalance of class distribution. A similarity measurement is designed combining visual concept and low level features afterwards. The results in mammographic patches database demonstrate that the proposed algorithm can extract the visual se- mantic effectively, and achieve better performance than other retrieval algorithms based on single low level feature.