DRHDR: A Dual branch Residual Network for Multi-Bracket High Dynamic Range Imaging

TL;DR

This paper presents DRHDR, a dual-branch neural network that effectively fuses multiple exposure brackets for high dynamic range imaging in dynamic scenes, balancing quality and computational efficiency.

Contribution

The introduction of a dual-branch residual network with deformable convolution and spatial attention for improved multi-bracket HDR imaging in dynamic scenes.

Findings

Achieves high-quality HDR results with reduced artifacts.

Operates efficiently by using two resolution-specific branches.

Effectively aligns features and suppresses ghosting artifacts.

Abstract

We introduce DRHDR, a Dual branch Residual Convolutional Neural Network for Multi-Bracket HDR Imaging. To address the challenges of fusing multiple brackets from dynamic scenes, we propose an efficient dual branch network that operates on two different resolutions. The full resolution branch uses a Deformable Convolutional Block to align features and retain high-frequency details. A low resolution branch with a Spatial Attention Block aims to attend wanted areas from the non-reference brackets, and suppress displaced features that could incur on ghosting artifacts. By using a dual branch approach we are able to achieve high quality results while constraining the computational resources required to estimate the HDR results.