Classification and retrieval of natural scenes

Scene recognition is of great importance for content based image retrieval systems and image indexing applications to process very large databases. Our motivation is to model the contents of the natural scenes by representing local image using region description. The basic idea of semantic modeling is to describe local image regions into semantic concepts using low level features such as color and texture. These local image region descriptions are combined to a global image representation that can be used for scene categorization and retrieval. In this paper, an automated Ncut segementation has been used for regions segmentation, Co occurrence matrix and Local Binary Pattern based texture features is used for local image representation that allows access to natural scenes. A simple, Non-parametric K-Nearest Neighbor classifier has been used to support automatic image annotation of local image region into semantic classes such as water, sky, and trees. Extensive experiments on databases like 8 Scene categories COREL, shows that the proposed technique performs well in scene classification.