Evaluation of Deep Learning Architectures for Image Denoising

Karthick Kumaran Ayyalluseshagiri Viswanathan

Abstract: Noise in the captured images has been a growing concern for various use cases not limited to mobile photography, low light imaging, drone capture, virtual reality headsets passthrough use cases etc., Recently with the advent of Apple Vision Pro and Meta Quest line of virtual reality (VR) headsets in the market, image denoising based research has got more importance to eliminate the noise from various sources, most importantly sensor noise, for improving the quality of passthrough applications. Traditional image denoising algorithms assume the noise to be Gaussian distributed but in practice the noise on the captured images can be significantly complex and so the traditional image filters fail badly for certain noise types. With the advancements in deep learning based neural networks it is now possible to remove the noise from the images so that the resulting image will be very close to the ground truth images. In this project I have implemented and evaluated two different deep learning architectures Autoencoders and U-Net for images affected with four different noise sources (Gaussian, Salt-and-Pepper, Speckle, Poisson) to denoise the images. Both the network models performance has been compared with PSNR (Peak Signal to Noise Ratio) metric and summarized the results.

Keywords: Deep Learning, Image Denoising, Autoencoder, U-Net, Convolutional Neural Network