Single Image HDR Reconstruction via A Feature Modulation Network

Reconstructing HDR images from a single LDR image with noise and saturated regions is a highly challenging problem. Recent approaches have utilized cascaded network structures to address this challenge in image space. However, this approach increases the difficulty of network training and may lead to cumulative errors. To this end, we propose a feature modulation network (FMN) for image denoising and HDR generation in feature space. Our FMN introduces a simple yet effective Feature Modulation Block (FMB) that combines an Enhanced Attention Module (EAM) and a Dilated Residual Dense Block (DRDB). The EAM effectively eliminates image noise using multiple residual blocks, while the DRDB learns the details lost in saturated regions of the input image with a large receptive field. Moreover, we employ a Space Feature Transform (SFT) layer to perform spatial-wise feature modulation. Extensive experiments are conducted on publicly available datasets, and the results demonstrate that our method achieves competitive performance compared to state-of-the-art methods.

I. Introduction

High dynamic range (HDR) imaging is a technique that aims to capture and reproduce a wider range of brightness than traditional digital imaging methods. HDR images contain more information about the luminance of a scene, resulting in a more visually immersive experience for viewers. As a result, over the past few decades, HDR imaging has garnered significant attention in academic research and has found widespread application in the entertainment industry, including film, television, games, and virtual reality (VR) [1]. The most common approach to generating HDR images is by merging multiple low dynamic range (LDR) images captured at different exposures [2]. However, this way often introduces artifacts due to object motion across the different LDR images [3], [4], and it is not always feasible to capture multiple images of the same scene. Therefore, the objective of this study is to reconstruct an HDR image from a single LDR image.