Publication

Abstract : Diabetic retinopathy (DR) is a disorder affecting retinal blood circulation; it is precipitated by diabetes mellitus. Since DR can lead to complete blindness, its early detection is of utmost importance. The computational methods for detecting DR include segmentation, feature extraction, and classification. This study uses different feature descriptors—local directional pattern (LDP), local binary pattern (LBP), and histogram-oriented gradients (HOG) for extracting features by processing diabetic retinopathy images. A set of supervised learning classifiers—support vector machine (SVM), random forest (RF), and K-nearest neighbor (KNN)—are used to classify the retina images based on the features extracted. The study indicates that a combination of LBP and HOG coupled with RF is the most accurate form among the various candidate approaches and gives the best accuracy of 87.50%.