Local wavelet pattern based glaucoma detection from fundus images

Abstract

An image descriptor known as the local wavelet pattern (LWP) is presented in this thesis to characterize retinal fundus images for glaucoma detection. The LWP is obtained for every pixel of the fundus image by encoding the relationship of the center pixel with the wavelet decomposed local neighbors. Another similar local image descriptor known as the local binary pattern (LBP) only takes into account the association of the center pixel with its surrounding pixels. In the presented technique, the correspondence among the neighboring pixels is first established using wavelet decomposition and then its association with the transformed center pixel is considered. In order to range match the wavelet decomposed local neighbors and the center pixel, a center pixel transformation method is introduced. This thesis studies: a) encoding the information between local neighboring pixels us ing wavelet analysis and b) calculating LWP using local wavelet decomposed vector and transformed center pixel vector. The LWP method is tested over two retinal fundus image databases for classification accuracy. We also weighed up the proposed LWP method with other local image descriptors, and the results suggest that LWP is far superior for retinal fundus image classification.