Scene classification is an important task in order to understand the outdoor images. Classifying scenes is not a simple issue because of ambiguity, variability, occlusion and illumination problems. Outdoor images involve to type of regions: natural scenes and man made structures. The most specific characteristics of natural scenes are their color and texture features. On the other hand man made structures are not able to be represented by using basically color and texture information. Although some type of local features such as sift points are successful in recognizing man made structures, they are inefficient in obtaining color and texture dominant structures. Man made structures generally consist of regular line segments that have usually two major dominant colors. By combining the color and texture features with line features we perform image segmentation. We use color information in pixel based one class classification and extract meaningful regions for natural scenes. Then we use line color information to extract meaningful regions for man made structures by line clustering. We get significantly improved results in scene classification. In order to overcome visual polysemy we describe a relevance feedback technique. We provide an image retrieval system to the user and include the user contribution in classification process. In accordance with the user feedback the system learn a one class classifier and try to improve retrieval results.

DATE: 19 March, 2007, Monday@ 16:40
PLACE: EA 409