Towards image retrieval by texture segmentation with genetic programming

This paper examines the feasibility of an approach to image retrieval from a heterogeneous collection based on texture. For each texture of interest (T), a T-vs-other classifier is evolved for small n times n windows using genetic programming. The classifier is then used to segment the images in the collection. If there is a significant contiguous area of T in an image, it is considered to contain that texture for retrieval purposes. We have experimented with sky and grass textures in the Corel Volume 12 image set. Experiments with a single image indicate that classifiers for the two textures can be learned to a high accuracy. Experiments with a test set of 714 Corel images gave a retrieval accuracy of 84% for both sky and grass textures. These results suggest that the use of texture could enhance retrieval accuracy in content based image retrieval systems.