Extending Denoising AutoEncoders for Feature Recognition

Abstract

Image rendering techniques involve the addition of noise on the rendered samples. The characteristics of Monte Carlo rendered noisy images makes it difficult to denoise using conventional methods. The noise induced in this case is spatially sparse. Using traditional methods of denoising and feature recognition is time consuming for such images as they use the entire image space as their search space. This results in unnecessary calculations that can be avoided and therefore reduce the processing time significantly. A recurrent convolutional neural network that operates on varying spatial resolutions, also known as the auto encoder decoder structure perform very well on these rendered images. The partitioning into encoder and decoder phases lets the network operate on continuously decreasing and increasing spatial domains respectively. This can be coupled with classification techniques to incorporate feature recognition of noisy images. In this paper the pretrained autoencoder layers are supported with additional softmax and sigmoid layers to enable feature recognition capabilities. � 2018 IEEE.