Texture feature neural classifier for remote sensing image retrieval systems

Texture information is useful for image data browsing and retrieval. The goal of this paper is to present a texture classification system for remote sensing images aimed at the administration of large collections of those images. The proposed classifier is a hybrid system composed by an unsupervised neural network and a supervised one. Starting from a small portion of the image (pattern) the system should recognize the most similar class to a pattern in a database as well as to identify images that contain similar patterns. The texture feature vectors used to characterize the patterns are obtained from the images processed by a bank of Gabor filters. Experimental results using textures of the Brodatz album, multi-spectral and radar images are presented.