An Efficient Wavelet-Based Image Denoising Technique for Retinal Fundus Images

Abstract

Diabetic Retinopathy (DR) occurs as a complication of long-standing cases of diabetes, which may even lead to vision loss. Many automatic diagnostic systems developed in the recent past used fundus images to diagnose retinal diseases. Usually, retinal fundus images acquired from the fundus camera suffer from noise and various contrast issues that deteriorate the image quality. It becomes a tedious task for ophthalmologists to detect tiny blood vessels and other retinal abnormalities. The preprocessing can be performed to bring out the significant features and eliminate image noises, which substantially enhance the image quality and improve the disease detection accuracy rate. In this paper, we have used the median filter, wiener filter, and discrete wavelet transform (DWT) for denoising fundus images. The DWT shows better results than other denoising methods. To further improve the performance metrics and reduce the cost function, K-Singular Value Decomposition (K-SVD) has been used in a wavelet domain in which overcomplete dictionaries are created for sparse representation. This proposed method was termed as DWT_K-SVD. These denoising techniques are used to test on two Diabetic Retinopathy datasets, such as EyePACS and Messidor-2. To assess the performance, peak signal-to-noise ratio (PSNR), mean square error (MSE), and structural similarity index measure (SSIM) are the parameters used. The proposed DWT_K-SVD method achieved the best results without losing any image features when compared to the other techniques taken under consideration.