Hierarchical Ensemble of AutoEncoder for Restoration of Images Corrupted by Cumulative Combination of Noise

Sayarnil Ganguly, Sanjana Reddy Katham, Sanyam Agrawal, Soumen Sinha, Rahul Roy

Abstract: Denoising images corrupted with multiple types of noise is significantly challenging due to their variable noise distribution. These numerous types of noise have a cumulative effect on the image. Therefore, considering only a single or multiple distribution of noise would be ineffective for restoration. Moreover, estimation of the noise is difficult where there is heterogeneity in the combination of the noise. Hence in this paper, we propose a sequence of autoencoders to progressively restore the noisy images. Each autoencoder specializes in the restoration of images affected by a particular combination of noise. These trained autoencoders are arranged in a state-space graph such that all possible combinations of noises can be screened by the autoencoder or their ensembles. The root of the state space graph is the \(N^{th}\) level autoencoder as it is exposed to all kinds of noise. In contrast, the autoencoder at the leaf level is exposed to one type of noise. During restoration of images, the obtained state space tree is searched using a heuristics search algorithm to find the near-optimal ensemble of autoencoders which can denoise the combinations of noises present in an image. Experiments are conducted with diverse combinations of noises to demonstrate the efficacy of the proposed approach. Furthermore, the proposed method is compared with various state-of-the-art approaches. It was observed that the proposed approach had significant efficiency in handling the combination of noise and outperforms various state-of-the-art techniques.