Transformed-Domain Robust Multiple-Exposure Blending With Huber Loss

Ryo Matsuoka,Masahiro Okuda,Shunsuke Ono

doi:10.1109/access.2019.2951817

Abstract

Pixel-domain weighting methods for multiple-exposure blending can efficiently remove noise and under-/over-exposed pixels simultaneously in high dynamic range (HDR) image generation. Various types of noise such as non-Gaussian noise, e.g., Poisson, impulse noise, and pixel saturation, are often superimposed to multiple-exposure images taken with a high ISO setting in a low-light condition. Because almost all existing methods assume Gaussian noise, these methods cannot sufficiently reduce these types of noise. To achieve high-quality HDR image generation in such difficult conditions, we propose a novel multiple-exposure blending method in which image blending is performed in a wavelet domain so as to enhance the denoising performance. In addition, the Huber loss function is utilized as a fidelity measure in blending to make the method robust against outliers. We also introduce an efficient algorithm based on a primal-dual splitting method for solving our optimization problem. The experimental results demonstrate the advantages of the proposed method over several conventional methods.

Highlights

The objective of high dynamic range (HDR) imaging is to represent the amount of light in a scene with a broad dynamic range, and it is applied to various fields such as computer graphics, medical imaging, in-vehicle cameras, and surveillance systems
In this paper, we propose a novel multiple-exposure blending technique for HDR image generation based on wavelet decomposition and the Huber loss function
The proposed exposure blending is performed in the wavelet domain, in which the optimal weights for scaling and wavelet coefficients can be estimated by solving a proposed weight optimization problem that can robustly eliminate various types of noise, including mixed noise contamination

Summary

Introduction

A. BACKGROUND The objective of high dynamic range (HDR) imaging is to represent the amount of light in a scene with a broad dynamic range, and it is applied to various fields such as computer graphics, medical imaging, in-vehicle cameras, and surveillance systems. Multiple-exposure blending is the most standard approach for HDR imaging because it can generate HDR images by blending a set of multiple-exposure images captured with a consumer camera [1]–[14]. High ISO shooting is a promising approach in taking multiple-exposure images for HDR image generation. This approach enables us to take ghost-free images, but the images are contaminated by heavy noise. Image-blending methods with noise removal have been studied [15]–[22]

Objectives

Results

Conclusion