Deep Convolutional Autoencoders for Deblurring and Denoising Low-Resolution Images

In this paper, we implement machine learning methods to recover higher-dimensional signals from lower-dimensional, noisy, and blurry measurements. In particular, rather than utilizing optimization-based reconstruction methods, we use fully-connected multilayer perceptron (MLP) architectures and convolutional neural networks (CNN). In addition, we consider two different loss functions based on mean squared error and a Huber potential to train our models. Numerical experiments on the Street View House Numbers dataset show that while fully-connected MLPs are faster to train, reconstructions using CNNs are much more accurate.