Removing Impulsive Noise from Color Images via a Residual Deep Neural Network Enhanced by Post-Processing

Sahar Sadrizadeh, Hatef Otroshi-Shahreza, Farokh Marvasti

Abstract: Impulsive noise is a common type of noise that affects gray-scale/color images and videos. In this paper, we present a residual fully Convolutional Neural Network (CNN) to remove impulsive noise from color images in an end-to-end fashion. We train our residual CNN model on a customized dataset which contains noisy images with different impulsive noise density. The proposed dataset omits the need for multiple models for different noise densities. Moreover, we employ a multi-term loss function to train our model. One of the terms of the proposed loss function imposes the sparsity of the impulsive noise in the observation domain. To the best of our knowledge, this term has not been employed as a loss function for training a denoising CNN. Finally, we employ an iterative post-processing stage to further improve the performance of our method. Simulation results demonstrate that the proposed approach outperforms other notable algorithms in the literature. Furthermore, our method is quite fast, especially when implemented on a GPU-equipped system.