A Two-stage Deep Network for High Dynamic Range Image Reconstruction

Mapping a single exposure low dynamic range (LDR) image into a high dynamicrange (HDR) is considered among the most strenuous image to image translationtasks due to exposure-related missing information. This study tackles thechallenges of single-shot LDR to HDR mapping by proposing a novel two-stagedeep network. Notably, our proposed method aims to reconstruct an HDR imagewithout knowing hardware information, including camera response function (CRF)and exposure settings. Therefore, we aim to perform image enhancement task likedenoising, exposure correction, etc., in the first stage. Additionally, thesecond stage of our deep network learns tone mapping and bit-expansion from aconvex set of data samples. The qualitative and quantitative comparisonsdemonstrate that the proposed method can outperform the existing LDR to HDRworks with a marginal difference. Apart from that, we collected an LDR imagedataset incorporating different camera systems. The evaluation with ourcollected real-world LDR images illustrates that the proposed method canreconstruct plausible HDR images without presenting any visual artefacts. Codeavailable: https://github. com/sharif-apu/twostageHDR_NTIRE21.