Proposal of a CBIR Framework for Retinal Images Using Fusion Edge Detection and Zernike Moments

Automated segmentation of blood vessels in retinal images is a critical component in the detection of diabetic retinopathy (DR). In this work, an attempt has been made to develop a Content Based Image Retrieval (CBIR) framework for retinal images. Various edge detection methods such as LoG, Canny, Sobel, Prewitt and Roberts are employed on preprocessed retinal fundus images to segment blood vessels. The output images of individual methods are integrated into four combinations such as LoG-Canny, Sobel-Canny, Sobel-Prewitt and Roberts-Prewitt using D-S fusion technique. The low and high-order invariant Zernike moments are extracted from the segmented blood vessels in four combinations and are analyzed. Results show that the D-S fusion based LoG-Canny edge detection provides continuous vessel map and higher vessel width than the other three combinations. The high-order Zernike moments of this group shows significant differentiation between normal and abnormal images. The retrieval performance such as precision and recall for D-S fusion based CBIR system is found to be 93% and 81% respectively. This CBIR system could be useful for automated diabetic retinopathy detection in mass screening.