Deep Convolutional Neural Network with Edge Feature for Image Denoising

Abstract

Image denoising is a classical challenge in computer vision and has attracted a large amount of research in the past few decades in attempts to find new approaches to denoise various types of images. The endeavors have been even more striking with the recent inception of deep learning. Deep learning has a well-known strength in its accuracy and effectiveness but with the main drawback of very long converging time during deep learning network training. Moreover, current state-of-the-art denoising techniques still lack the edge feature, one of the crucial attributes for sharp images. Therefore, in this paper, we propose a novel Deep Convolutional neural network with Edge Feature (DCEF) for denoising Additive White Gaussian Noise (AWGN). However, as the deep learning model training takes too much time to converge, we also propose an adaptive learning rate using a triangle technique that allows much faster converging time comparing to state-of-the-art approaches. Our DCEF demonstrates that it outperforms existing state-of-the-art approaches in terms of average PSNR scores in $\upsigma$ = 15 and 25 by 0.2 and 0.3, respectively, while achieving high MS-SSIM scores and using much fewer iterations to converge.