High resolution image classification with features from wavelet frames

Scale and orientation are fundamental parameters of visual processing. The methods of multiresolution feature extraction and enhancement of orientation features have been proposed for texture classification of a high resolution satellite image. A texture is characterized by a set of enhanced channel variances estimated at the output of the wavelet frame. Features from the wavelet frame are used as input to the error backpropagation algorithm for training, and classification. The results of the proposed method with multilayer perceptron (MLP) classifier are compared with the results of the maximum likelihood classifier (MLC). The method is that texture segmentation divides the image into “homogeneous” regions where local texture properties are approximately invariant. The measurements remain reasonably constant in a region where the texture is considered to be homogeneous. The orientation of texture elements and their frequency contents seem to be important clues for discrimination. The proposed method is illustrated with the aid of examples on texture of high resolution satellite image of Seoul, Korea