Convolutional Residual Blocks With Edge Guidance for Image Denoising

Abstract

We propose an image denoising Convolutional Neural Network (CNN), comprises of alternative vertical and horizontal Sobel edge guided residual blocks. The input images are corrupted by additive white Gaussian noise. Initially, the first convolutional layer maps three channel (RGB) input noisy image into depth feature maps. Next, each residual block extracts vertical/horizontal edge using the Sobel edge mask. The edge guidance helps in preserving edges and textures in the given input image. Further, the input depth feature map along with corresponding edges are fed into simple <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$1\times 1,1\times 3,3\times$</tex> 1 and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$3\times 3$</tex> convolutional layers with residual connections. Usage of simple convolutional blocks reduces the overall network parameter count, while maintaining the better performance. The simple and efficient design of the proposed residual convolution block can easily be plugged into any other image reconstruction networks. Experimental results show that the proposed model is more effective in preserving textures and edges while removlng noise.