Study of gabor and local binary patterns for retinal image analysis

In this paper selection of proper feature for retinal vascular tissue segmentation is studied. Different features have been proposed for retinal vessel detection. One of the most famous features adapted is Gabor wavelet. Due to multi resolution property of Gabor, combination of scales can be used to extract features. However, similar features in feature vector would increase the possibility of inter-correlation and not an apt result would be achieved. Also, local binary pattern (LBP) is studied. LBP is a powerful feature for texture analysis. Although LBP itself is not as good as Gabor for vessel detection, it will be showed enhances the result of segmentation. In order to select the best feature vector, a hierarchical feature selection method is proposed. At the fist step a set of candid features is selected which later reduces to the set of best possible features. It is shown that the combination of some resolutions of local binary pattern and Gabor, including the inverted green channel would structure the best feature vector.