基于舌图像集合的距离度量学习模型及舌象分类方法研究

项目名称： 基于舌图像集合的距离度量学习模型及舌象分类方法研究

项目编号： No.61471146

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 无线电电子学、电信技术

项目作者： 张宏志

作者单位： 哈尔滨工业大学

项目金额： 87万元

中文摘要： 与现有的计算机舌诊研究不同，本项目拟将舌象分类的数据单元由传统的单幅静态舌图像扩展为舌图像集合。舌图像集合中包括对同一舌体所采集的舌象视频，或在不同时期、角度，甚至光照波动条件下采集的多幅舌图像。与单幅静态图像相比，舌图像集合包括更丰富的舌象分类和诊察信息，因此有助于大幅提高舌象分类的性能。然而舌图像集合中也包含了一定的扰动因素，且传统的基于点到点的距离度量学习方法无法用于舌图像集合的分类，因此对相关研究提出了全新的挑战。为此，本项目拟开展专门针对舌图像集合的分类与建模研究，发展稳健的舌图像集合距离度量学习模型与算法；基于舌象视频，发展的高性能舌有形特征分割与提取方法，解决特定稀缺舌象类别的小样本问题；最终，构建基于舌图像集合的高性能舌象分类模型，从而为舌诊信息化研究提供一种全新的量化和客观化分析手段，促进舌诊在教学、研究和临床等应用中的推广和普及，具有重要的学术价值。

中文关键词： 舌图像集合分类；舌象视频分析；距离度量学习；舌象特征分析；流形学习

英文摘要： Different from the existing research on computer tongue diagnosis (TD), we use tongue image set (TIS), instead of single static image, as the basic data unit for classification. Included in tongue image set are videos of the same tongue, or images which are captured at different time, from different view, and even under variable lighting conditions. Compared with single static image, a TIS includes more information for classification and inspection and therefore contributes to the performance of tongue image classification. Nevertheless, a TIS also contains certain outliers, and that traditional point-to-point distance metric methods are not suitable for TIS classification, which hence proposed new challenges to us. For this purpose, we intent to carry out specific research on TIS classification and modeling, and develop robust distance metric learning methods for TIS. Based on tongue video, high performance segmentation and extraction algorithms of tangible lingual characters are investigated. Plus, small sample size problem for certain tongue image classes is also addressed by employing tongue video. Subsequently, a high-performance model is configured up for TIS classification, which offers a novel quantifiable and objectification method for informatization research on Tongue diagnosis. The study will accelerate the promotion and popularization of TD in education, research and clinical.

英文关键词： tongue image set classificaiton;tongue video analysis;distance metric learning;tongue feature analysis;manifold learning