Foundations, Inference, and Deconvolution in Image Restoration

Image restoration is a critical preprocessing step in computer vision, producing images with reduced noise, blur, and pixel defects. This enables precise higher-level reasoning as to the scene content in later stages of the vision pipeline (e.g., object segmentation, detection, recognition, and tracking). Restoration techniques have found extensive usage in a broad range of applications from industry, medicine, astronomy, biology, and photography. The recovery of high-grade results requires models of the image degradation process, giving rise to a class of often heavily underconstrained, inverse problems. A further challenge specific to the problem of blur removal is noise amplification, which may cause strong distortion by ringing artifacts. This dissertation presents new insights and problem solving procedures for three areas of image restoration, namely (1) model foundations, (2) Bayesian inference for high-order Markov random fields (MRFs), and (3) blind image deblurring (deconvolution). As basic research on model foundations, we contribute to reconciling the perceived differences between probabilistic MRFs on the one hand, and deterministic variational models on the other. To do so, we restrict the variational functional to locally supported finite elements (FE) and integrate over the domain. This yields a sum of terms depending locally on FE basis coefficients, and by identifying the latter with pixels, the terms resolve to MRF potential functions. In contrast with previous literature, we place special emphasis on robust regularizers used commonly in contemporary computer vision. Moreover, we draw samples from the derived models to further demonstrate the probabilistic connection. Another focal issue is a class of high-order Field of Experts MRFs which are learned generatively from natural image data and yield best quantitative results under Bayesian estimation. This involves minimizing an integral expression, which has no closed form solution in general. However, the MRF class under study has Gaussian mixture potentials, permitting expansion by indicator variables as a technical measure. As approximate inference method, we study Gibbs sampling in the context of non-blind deblurring and obtain excellent results, yet at the cost of high computing effort. In reaction to this, we turn to the mean field algorithm, and show that it scales quadratically in the clique size for a standard restoration setting with linear degradation model. An empirical study of mean field over several restoration scenarios confirms advantageous properties with regard to both image quality and computational runtime. This dissertation further examines the problem of blind deconvolution, beginning with localized blur from fast moving objects in the scene, or from camera defocus. Forgoing dedicated hardware or user labels, we rely only on the image as input and introduce a latent variable model to explain the non-uniform blur. The inference procedure estimates freely varying kernels and we demonstrate its generality by extensive experiments. We further present a discriminative method for blind removal of camera shake. In particular, we interleave discriminative non-blind deconvolution steps with kernel estimation and leverage the error cancellation effects of the Regression Tree Field model to attain a deblurring process with tightly linked sequential stages