Multi-scale Image Decomposition using a Local Statistical Edge Model

TL;DR

This paper introduces a novel multi-scale image decomposition method using a local statistical edge model and a Sub-window Variance filter, enabling effective detail enhancement and suppression with gradient preservation and high parallelization.

Contribution

The paper presents a new non-linear filter and local statistical edge model for multi-scale image decomposition tailored for detail enhancement, with an efficient, gradient-preserving, and parallelizable pipeline.

Findings

Effective detail enhancement and suppression demonstrated

Gradient preservation avoids artefacts in results

High parallelization accelerates processing

Abstract

We present a progressive image decomposition method based on a novel non-linear filter named Sub-window Variance filter. Our method is specifically designed for image detail enhancement purpose; this application requires extraction of image details which are small in terms of both spatial and variation scales. We propose a local statistical edge model which develops its edge awareness using spatially defined image statistics. Our decomposition method is controlled by two intuitive parameters which allow the users to define what image details to suppress or enhance. By using the summed-area table acceleration method, our decomposition pipeline is highly parallel. The proposed filter is gradient preserving and this allows our enhancement results free from the gradient-reversal artefact. In our evaluations, we compare our method in various multi-scale image detail manipulation applications with other mainstream solutions.