Denoising Convolutional Neural Network for Fundus Patches Quality

Image quality is one of the factors that supporting a system to perform tasks such as image classification, segmentation, and recognition. Better tasks learning needs image quality improvement. In this study, image quality improvement do by denoising Convolutional Neural Network (dnCNN). The dnCNN network include the convolution layer, rectified linear unit, batch normalization, and final regression layer. DnCNN layer configuration consists of 58 layers. The denoising image process have five steps: data access, noise image, configuration of dnCNN layer, denoised image, and assess image quality with Peak Signal-to-Noise Ratio (PSNR), Structure Similarity (SSIM), and Mean Square Error (MSE) Experiment uses public data fundus image from MESSIDOR and Retina Image Bank. Results show, average value of PSNR noisy image 20.75, PSNR denoised image 31.8, SSIM noisy image 0.92, SSIM denoised image 0.99, MSE noisy image 0.0084, MSE denoised image 0.0007.