Computer Vision News - September 2018

Aim & Motivation: Classical denoising methods used image prior modeling as a stepping stone for denoising. Although those methods achieve high denoising quality, they suffer from two main drawbacks: (a) they require a complex, computationally demanding, optimization problem for testing. (b) the non-convex models require manual selection of several parameters, leaving room for performance improvement. Advantages: The authors propose an end-to-end trainable feed-forward denoising convolutional neural network (DnCNN), taking advantage of the progress in deep learning methods (i.e. batch normalization, residual learning). The network doesn’t directly output the denoised image ^x, instead it is designed to predict the residual image (the noise itself). And rather than learning a discriminative model using an explicitly predicted image prior, it treats image denoising as a plain discriminative learning problem, that is, separating the noise from a noisy image. The authors propose a single network to solve three general image denoising tasks: blind gaussian denoising, SISR, and JPEG deblocking. Background: Image denoising, like every other field in computer vision, can be divided into two periods: the models developed prior to the deep learning revolution, and the models developed since :-). Prior to the deep learning revolution, image denoising explored various methods for modeling image priors, including gradient models, sparse models, Markov random field (MRF) models and nonlocal self-similarity (NSS) models, which are implemented in state-of-the-art methods such as BM3D, LSSC, NCSR and WNNM. As mentioned, these methods’ main disadvantage is a complex, computationally demanding optimization problem at the testing stage. TNRD (Trainable Nonlinear Reaction Diffusion model), despite being proposed in 2016 (after the deep learning revolution), still uses prior models. Developed by Chen et al., its critical point lies in the additional training of the influence functions. The effectiveness of the trained diffusion models is attributed to the following properties of the trained filters: Computer Vision News Research 5 Research Computer Vision News Super-resolution is the process of creating high-resolution images from low- resolution images, one of the best known approaches being SISR, where the purpose is the reconstruction of a high-res image from one single low-res image. SISR is challenging because without any data from a high-res image, the quality of the reconstructed image is necessarily limited. Moreover, SISR is an ill-posed problem, since an infinite catalogue of high-res images can be produced for every low-res image.