Advanced Restoration Techniques for Images and Disparity Maps

Abstract

With increasing popularity of digital cameras, the field of Computa- tional Photography emerges as one of the most demanding areas of research. In this thesis we study and develop novel priors and op- timization techniques to solve inverse problems, including disparity estimation and image restoration. The disparity map estimation method proposed in this thesis incor- porates multiple frames of a stereo video sequence to ensure temporal coherency. To enforce smoothness, we use spatio-temporal connec- tions between the pixels of the disparity map to constrain our solution. Apart from smoothness, we enforce a consistency constraint for the disparity assignments by using connections between the left and right views. These constraints are then formulated in a graphical model, which we solve using mean-field approximation. We use a filter-based mean-field optimization that perform efficiently by updating the dis- parity variables in parallel. The parallel updates scheme, however, is not guaranteed to converge to a stationary point. To compare and demonstrate the effectiveness of our approach, we developed a new optimization technique that uses sequential updates, which runs ef- ficiently and guarantees convergence. Our empirical results indicate that with proper initialization, we can employ the parallel update scheme and efficiently optimize our disparity maps without loss of quality. Our method ranks amongst the state of the art in common benchmarks, and significantly reduces the temporal flickering artifacts in the disparity maps. In the second part of this thesis, we address several image restora- tion problems such as image deblurring, demosaicing and super- resolution. We propose to use denoising autoencoders to learn an approximation of the true natural image distribution. We parametrize our denoisers using deep neural networks and show that they learn the gradient of the smoothed density of natural images. Based on this analysis, we propose a restoration technique that moves the so- lution towards the local extrema of this distribution by minimizing the difference between the input and output of our denoiser. Weii demonstrate the effectiveness of our approach using a single trained neural network in several restoration tasks such as deblurring and super-resolution. In a more general framework, we define a new Bayes formulation for the restoration problem, which leads to a more efficient and robust estimator. The proposed framework achieves state of the art performance in various restoration tasks such as deblurring and demosaicing, and also for more challenging tasks such as noise- and kernel-blind image deblurring. Keywords. disparity map estimation, stereo matching, mean-field optimization, graphical models, image processing, linear inverse prob- lems, image restoration, image deblurring, image denoising, single image super-resolution, image demosaicing, deep neural networks, denoising autoencoders