Structure-embedded ghosting artifact suppression network for high dynamic range image reconstruction

Abstract

In high-dynamic-range (HDR) image reconstruction, the background offset among multiple multi-exposure low-dynamic-range (LDR) images, wide-range movement of targets, and missing edge structure information in the over-/under-exposure region cause both ghosting and blurring artifacts. This study proposed a structure-embedded ghosting artifact suppression network (SGARN) to achieve detailed preservation and ghosting artifact suppression to address this issue. According to the different image feature maps’ correlation in channels, a channel and multi-head joint attention network (CMAN) was designed to highlight the features conducive to high-quality HDR image reconstruction. A dense multi-scale information transfer network (DMITN) was designed to integrate the characteristics of different combinations of convolution kernels with different receptive fields. In addition, a structure-embedded network was designed to predict the edge structure to be compensated from the reference image. The predicted edge was integrated into the reconstructed HDR image. Compared with state-of-the-art methods, the proposed method can achieve better visual performance and higher objective evaluation results on three public datasets. The source codes of the proposed method are available at https://github.com/lhf12278/SGARN.