Single-Image HDR Reconstruction Assisted Ghost Suppression and Detail Preservation Network for Multi-Exposure HDR Imaging
Huafeng Li, Zhenmei Yang, Yafei Zhang, Dapeng Tao, Zhengtao Yu
Abstract
The reconstruction of high dynamic range (HDR) images from multi-exposure low dynamic range (LDR) images in dynamic scenes presents significant challenges, especially in preserving and restoring information in oversaturated regions and avoiding ghosting artifacts. While current methods often struggle to address these challenges, our work aims to bridge this gap by developing a multi-exposure HDR image reconstruction network for dynamic scenes, complemented by single-frame HDR image reconstruction. This network, comprising single-frame HDR reconstruction with enhanced stop image (SHDR-ESI) and SHDR-ESI-assisted multi-exposure HDR reconstruction (SHDR-A-MHDR), effectively leverages the ghost-free characteristic of single-frame HDR reconstruction and the detail-enhancing capability of ESI in oversaturated areas. Specifically, SHDR-ESI innovatively integrates single-frame HDR reconstruction with the utilization of ESI. This integration not only optimizes the single image HDR reconstruction process but also effectively guides the synthesis of multi-exposure HDR images in SHDR-A-MHDR. In this method, the single-frame HDR reconstruction is specifically applied to reduce potential ghosting effects in multi-exposure HDR synthesis, while the use of ESI images assists in enhancing the detail information in the HDR synthesis process. Technically, SHDR-ESI incorporates a detail enhancement mechanism, which includes a self-representation module and a mutual-representation module, designed to aggregate crucial information from both reference image and ESI. To fully leverage the complementary information from non-reference images, a feature interaction fusion module is integrated within SHDR-A-MHDR. Additionally, a ghost suppression module, guided by the ghost-free results of SHDR-ESI, is employed to suppress the ghosting artifacts. Experimental results on four public datasets demonstrate the efficacy and superiority of the proposed method. The code is available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/lhf12278/SAMHDR</uri> .