Retinal image segmentation based on weighted local detectors and confusion matrix

The paper presents a method for accurate detection and localization of important regions from retinal images like: optic disc, macula, exudates and hemorrhages. To this end, the image is locally decomposed in sub-images (patches) and then it is processed based on the fusion of different information types: first order statistics, textural, fractal and spectral. Two multilayer processing networks are considered: one for the feature selection and class representative establishment, in the learning phase, and another for voting scheme, based on local detectors, in the classification phase. Taking into account the results from associated confusion matrices, in order to increase the classification accuracy, different weights were assigned to local detectors. It was tested 140 images from different public databases (40 for the learning phase and 100 for the classification phase). The experimental results indicate a good accuracy for all analyzed regions of retinal images.