Image Denoising based on Multi Domain Feature Inspired Convolutional Neural Network
Ramesh Kumar Thakur, Suman Kumar Maji · TENCON 2022 - 2022 IEEE Region 10 Conference (TENCON) · 2022
Blind Gaussian denoising is the process of removing noise from an image affected by additive white Gaussian noise (AWGN) of unknown variance. This paper proposes a multi domain feature inspired deep neural network for blindly (noise variance unknown) denoising AWGN corrupted images. The proposed network uses features of the noisy image in both spatial domain as well as frequency (discrete cosine transform) domain. Spatial and frequency domain features are concatenated to get a combined feature map from which we generate the residual noisy image. The residual noise is then subtracted from the input noisy image to obtain the final denoised image. By restricting the number of layers and trainable parameters in the proposed network, we ensure that it is lightweight in nature and provides fast denoising capability. Comparison with classical state-of-the-art methods and recent CNN based denoising approaches show that the proposed network gives superior denoising results both visually and quantitatively. Our code is uploaded at https://github.com/RTSIR/MDFBGDN.