Discriminative Indexing for Probabilistic Image Patch Priors

Abstract

Abstract. Newly emerged probabilistic image patch priors, such as Expected Patch Log-Likelihood (EPLL), have shown excellent performance on image restoration tasks, especially deconvolution, due to its rich expressiveness. However, its applicability is limited by the heavy computation involved in the associated optimization process. Inspired by the recent advances on using regression trees to index priors defined on a Conditional Random Field, we propose a novel discriminative indexing approach on patch-based priors to expedite the optimization process. Specifically, we propose an efficient tree indexing structure for EPLL, and overcome its training tractability challenges in high-dimensional spaces by utilizing special structures of the prior. Experimental results show that our approach accelerates state-of-the-art EPLL-based deconvolution methods by up to 40 times, with very little quality compromise.

Accelerating GMM-Based Patch Priors for Image Restoration: Three Ingredients for a <inline-formula> <tex-math notation="LaTeX">$100{\times}$ </tex-math> </inline-formula> Speed-Up

“…To the best of our knowledge, there are only two other approaches [39], [35] that have attempted to accelerate EPLL. Unlike our approach, these methods focus on accelerating only one of the steps of EPLL namely the Gaussian selection step.…”

“…In [39], the authors use a binary decision tree to approximate the mappingz i → k * i performed in step Gaussian selection. At each node k and level n of the tree, the patchz i is confronted with a linear separator in order to decide if the recursion should continue on the left or right child given by…”

“…Given that the main advantage of EPLL is that the same model can be used for any type of degradations, it is important that this property remains true for the accelerated version. Last but not least, the training of our search tree takes a few minutes while the training steps for the above mentioned approach take from several hours to a few days [39].…”

“…Recently, researchers have tried to improve the speed of EPLL by replacing the most time-consuming operation in the EPLL algorithm with a machine learning-based technique of their choice [39], [35]. These methods were successful in accelerating EPLL to an extent but did not consider tackling all of its bottlenecks.…”

Bejo: Behavior Based Job Classification for Resource Consumption Prediction in the Cloud

“…Without proper indexing, the usage of k nearest neighbors in large scale datasets makes these approaches time consuming, thus limiting their applications in practice. Fortunately, recent development in machine learning provides sophisticated nonlinear regression methods, which have been demonstrated to be effective in computer vision [9], [10] and speech recognition [11]. This sheds new light on more practical classification schemes on the resource consumption, but what learning algorithms are suitable for job classification on cloud computing platforms is still largely overlooked.…”