In this paper we developed an intelligent method for the selection of statistical, textural and fractal features that characterize different regions of interest in eye-fundus images. Because the regions like optic disc, macula, exudates and hemorrhages are difficult to detect, an intelligent scheme for feature detection and classification is necessary. The method is based on a voting scheme that takes into account the values on the main diagonal of different confusion matrices. These matrices are generated based on clusters from sorted data-sets of feature values. Both the sliding box and fixed box techniques where used to divide the image into patches, in order to highlight the regions of interest and to obtain the unique signature of features for each of them. Two algorithms are proposed, one for the intelligent selection of features and the second for the testing of the accuracy of selected features. The results obtained on 100 test images proved the efficiency of the proposed method compared to other algorithms.
展开▼
