Marnerides Demetris Bashford-Rogers Thomas Hatchett Jonathan Debattista Kurt

Abstract

High dynamic range (HDR) imaging provides the capability of handling realworld lighting as opposed to the traditional low dynamic range (LDR) whichstruggles to accurately represent images with higher dynamic range. However,most imaging content is still available only in LDR. This paper presents amethod for generating HDR content from LDR content based on deep ConvolutionalNeural Networks (CNNs) termed ExpandNet. ExpandNet accepts LDR images as inputand generates images with an expanded range in an end-to-end fashion. The modelattempts to reconstruct missing information that was lost from the originalsignal due to quantization, clipping, tone mapping or gamma correction. Theadded information is reconstructed from learned features, as the network istrained in a supervised fashion using a dataset of HDR images. The approach isfully automatic and data driven; it does not require any heuristics or humanexpertise. ExpandNet uses a multiscale architecture which avoids the use ofupsampling layers to improve image quality. The method performs well comparedto expansion/inverse tone mapping operators quantitatively on multiple metrics,even for badly exposed inputs.