Image Denoising Framework Employing Auto Encoders for Image Reconstruction

Auto Encoder (AE) can be used in denoising of images. It is a type of neural network that can reconstruct the input. An auto-encoder is to represent a (sparse) input dataset in a compressed form that retains the most relevant information such that it may be reconstructed at the output with minimal loss from the compressed representation. In this paper, deep AE, denoising AE, and variational AE are used. Any AE where an extra constraint is put on the bottleneck to have a low KL divergence from a Normal Distribution is a Variational AE. There are multiple ways in which Variational AE is used, but the most common one is generative. The decoders on the top of the bottleneck can be used to generate new data points. Maximum accuracy of 88.85% is observed using the Denoising autoencoder while 76.25% and 81.44% are observed for deep and Variational autoencoder, respectively. 8.3% and 14.18% accuracy improvement is observed in Variational and deep AE, respectively.