Region Selective Information Augmentation for Retinal Images

Abstract

This paper proposes a scheme for the image super-resolution (SR) of retinal fundus images obtained from the handheld imaging devices. Clinical information in fundus images is present in the blood vessels, the optic disk, the macula and the retinal lesions. The homogeneous regions in the fundus images do not carry much relevant diagnostic information. Therefore, a patch based region selective approach for retinal image SR is proposed for providing a relatively fast and accurate SR results. The contrast sensitivity index (CSI) is applied to coarsely measure the clinical relevance of the input low resolution retinal image patch. If the clinical relevance of the input patch is above a particular threshold, then sparse representation based SR is performed else time efficient bicubic interpolation is performed for the patch. The method attains a quality performance in terms of the peak signal to noise ratio (PSNR) and the structural similarity index measure (SSIM) value.