Hybrid Technique for Fundus Image Enhancement Using Modified Morphological Filter and Denoising Net

Diabetic retinopathy is manually predicted by ophthalmologists using features such as variations in length and thickness of blood vessels, lesions like hemorrhages, exudes, and microaneurysms in fundus images. However, the images captured by the fundus camera sensor are prone to noise, undergo uneven illumination, and poor contrast. To effectively identify retinal diseases at an earlier stage, it is necessary to eliminate noise and improve image resolution. A hybrid model for denoising and enhancement is proposed that combines a modified morphological filter and a denoising net. The proposed multi-angle two-stage morphological filter and denoising net (MMFDNet) aids the recovery of a clean image and preserves features from the degraded retinal image. This discriminative learning denoising model removes various noise levels of speckle and additive white Gaussian noise (AWGN) from the fundus image. In Denoising Net, the noisy fundus image is restructured using a down-sampling operation, and the four sub-images are combined with a tunable noise level map, which increases the size of the receptive field by improving the speed and performance of the network. The denoised sub-images are grouped to get the latent image, and the morphological filter is used to improve the quality of the latent image. The effectiveness of this method is tested quantitatively and qualitatively with other existing methods. The databases used to analyze the MMFDNet are proprietary dataset, Messidor and STARE dataset. Compared to existing methodologies, the proposed enhancement technique improves peak signal-to-noise ratio (PSNR) by 2.29%, structural similarity index (SSIM) value by 9.53% and reduces the error by mean square error of 16.31% for Gaussian noise with a noise variance of 0.001. Similarly, improves PSNR by 6.60%, SSIM value by 0.33% and reduces the error by MSE of 47.81% for speckle noise with a noise variance of 0.001. The experimental results illustrate that the proposed MMFDNet outperforms the state-of-the-art methods.