Learning Concepts in Morphological Image Databases

加州大学河滨分校（University of California-Riverside）获得一项资助，开发形态学图像数据库的正式框架，该框架可自动从图像中提取适当的视觉特征，允许用户反馈，利用积累的元知识，有效地学习不同昆虫物种的视觉概念，有效地操纵数据库实体，显著提高图像检索性能。这些技术将通过使用各种定量性能评估措施在多个数据库上进行科学实验来验证。该项目涉及一个跨学科团队，生物学家和计算机科学家之间的密切合作，并与宾夕法尼亚大学和加州大学河滨分校的其他几位对这项工作感兴趣并从事昆虫研究的生物学家进行互动。该项目开发了用于提取形态信息的新技术，包括：(1)在半监督学习的健全数学框架中，在多个抽象层次上使用局部/全局特征；(2)使用关联反馈和从多个用户长期学习；(3)基于局部斑块表示的相似性分析和斑块之间的关系，这将提供几何形态计量学，并最终导致对特征的理解；(4)新颖的空间索引结构，考虑了不确定性，可以处理大量数据。该系统识别的物种形态特征将有助于在图像和视频数据库中进行快速、自动的物种识别和搜索。该系统的原型将与UCR现有的图像数据库服务器MorphNet相结合。

The University of California-Riverside is awarded a grant to develop a formal framework for morphological image databases, which automatically extracts appropriate visual features from images, allows user feedback, exploits the accumulated meta knowledge, effectively learns visual concepts for different insect species, efficiently manipulates database entities, and significantly improves the image retrieval performance. The techniques will be validated by performing scientific experiments on multiple databasesusing a variety of quantitative performance evaluation measures. The project involves an interdisciplinary team and a close collaboration between biologists and computer scientists and interactions with severalother biologists from the University of Pennsylvania and the University of California at Riverside who are interested in this effort and who work with insects. The project develops novel techniques for the extraction of morphological information including: (1) use of local/global features at multiple levels of abstraction in a sound mathematical framework of semi-supervised learning; (2) use of relevance feedback and long-termlearning from multiple users; (3) analysis of similarity based on local patch-based representation and relations between the patches which will provide geometric morphometrics and ultimately will lead to theunderstanding of character; and (4) novel spatial indexing structure that account for the uncertainty and can handle large amounts of data. Morphological features recognized by the system would be an aid to rapid, automated species identification and search in image and video databases. A prototype of the system will be coupled to MorphNet, an existing image database server at UCR.