Abstract
This paper aims to propose a robust hybrid probabilistic learning approach that combines appropriately the advantages of both the generative and discriminative models for the challenging problem of diabetic retinopathy classification in retinal images. We build new probabilistic kernels based on information divergences and Fisher score from the mixture of scaled Dirichlet distributions for support vector machines (SVMs). We also investigate the incorporation of a minimum description length criterion into the learning model to deal with the common problems of determining suitable components and also selecting the best model that describes the dataset. The developed hybrid model is introduced in this paper as an effective SVM kernel able to incorporate prior knowledge about the nature of data involved in the problem at hand and, therefore, permits a good data discrimination. Our approach has been shown to be a better alternative to other methods, which is able to describe the intrinsic nature of datasets and to be of a significant value in a variety of applications involving data classification. We demonstrate the flexibility and the merits of the proposed framework for the problem of diabetic retinopathy detection in eye images.