Probabilistic and Structural Approaches to Image Understanding, Classification and Retrieval, and Applications to Remote Sensing and Medical Image Analysis (CAREER Award)


Increased availability of the internet and advances in computer performance and storage systems improved the feasibility of many scientific applications. Advances in these applications also increased the amount of generated data and the need for its effective processing and storage. Automatic content extraction and efficient access to this content became highly desired goals in intelligent databases.

This project aims to develop automatic techniques for semantic processing of images, extraction of the relationships between them, efficient storage of these relationships in databases, improved access to the stored information, and comparisons between new information and past cases. Semantic processing of images will be done using a hierarchical decomposition. First, clustering and density estimation will be performed on color and texture features of pixels. Pixels will be classified based on the estimated densities. Then, segmentation techniques will be used to group the pixels and divide images into meaningful regions. Even when pixels and regions can be identified correctly, semantic processing requires knowledge about the relationships and relative arrangements of regions. The arrangements will be modeled using attributed relational graphs that encode relationships between regions.

The developed methods will be used to find interesting patterns in image databases, to study change detection using images of the same scene taken at different times, and to perform content-based searches. This work will enable us to analyze large image databases by their high-level semantic content instead of the limited representation supported by low-level feature vectors.



April 2005 - April 2010


TUBITAK Logo TUBITAK - Scientific and Technical Research Council of Turkey (Grant no: 104E074)


189,360 YTL (~US$126,000)