A New Classification Mechanism for Retinal Images

Abstract

In this paper, we propose a classification mechanism for retinal images so that the retinal images can be successfully distinguished from nonretinal images, egg yolk images for example. The proposed classification mechanism consists of two procedures: training and classification. The image features of retinal images and nonretinal images are extracted at the beginning of the training procedure to make sure the precision rate of the proposed classification mechanism is as high as possible while maintaining acceptable execution time of training procedure. In this paper, we design two classification mechanisms: one is pure SVM and the other is a hybrid that combines PCA and SVM mechanisms. Experimental results confirm that the accuracy rate of pure SVM is up to 96% for both 10-image and 20-image data sets. Moreover, PCA+SVM not only successfully reduces the features of images by using PCA but also maintains the accuracy rate above 90% for 10-and 20-image data sets.