Skip to main content
eScholarship
Open Access Publications from the University of California

Efficient Methods for Image Denoising using Learned Patch Priors

  • Author(s): Parameswaran, Shibin
  • Advisor(s): Nguyen, Truong Q
  • et al.
Abstract

Cameras have become ubiquitous leading to an increase in the amount of video and image data captured by amateurs and professionals alike. Their ease of deployability makes them a great sensor for security applications as well. Hence, there is an ever-growing need to \textit{efficiently} process and enhance captured image and videos for improving the performance of subsequent computer vision algorithms or simply for aesthetic reasons. To address this need, we focus on creating efficient techniques for large scale image and video denoising with varying degrees of genericity.

We start by introducing a robust patch matching technique that increases the efficacy of denoising algorithms that build patch-specific filters. We show that using our matching criterion in multiple leading denoising algorithms provides additional performance gains over using default distance metrics. Next, we present a strategy to extend patch-based image denoising algorithms into a decompressed video denoising paradigm without increasing computational complexity. We leverage pre-calculated motion vectors present in a compressed video's bitstream to establish temporal correspondences, thus keeping the per-frame complexity of the video denoising algorithm equivalent to that of the corresponding image denoising method. Following this, we relax the patch-specific constraint on design of denoising filters leading to one of the fastest algorithms that uses \textit{targeted} local patch prior.

Recognizing that a \textit{targeted} patch prior could be a limiting factor for a wide variety of natural images, we develop an efficient denoising algorithm that uses a Gaussian Mixture Model (GMM) to model a \textit{generic} patch prior for image restoration.

It is two orders of magnitude faster than similar methods while providing a very competitive quality-vs-speed operating curve. The final work presented in this thesis improves upon GMM priors by proposing a more expressive distribution using Generalized Gaussian Mixture Models (GGMM) patch priors. We circumvent the prohibitive computational complexity of using GGMM patch priors for image restoration by introducing asymptotically accurate but computationally efficient approximations to the bottlenecks encountered in this formulation. Our evaluations indicate that the GGMM prior is consistently a better fit for modeling image patch distribution and performs better on average in image denoising task.

Main Content
Current View