A dense optical flow map-guided convolutional neural network for multi-exposure image fusion in dynamic scenes

Abstract

In recent years, multi-exposure image fusion-based high dynamic range (HDR) imaging technology has attracted widespread attention in optical instrumentation and measurement-based fields such as 3D measurement, industrial welding, and biological testing. However, existing methods are prone to producing ghosting artifacts when handling dynamic scenes, which results in significant limitations in their applicability. To address this issue, we propose a dense optical flow map-guided convolutional neural network for multi-exposure image fusion method in dynamic scenes, called DOFM-HDRNet, in this paper. This network leverages an attention module guided by dense optical flow maps between non-reference and reference images, which helps to highlight and retain complementary features of the reference image during feature extraction, while suppressing motion and overexposed regions and enhancing image details. Additionally, the fusion module employs dilated residual dense blocks (DRDBs) to expand the receptive field, enabling more accurate estimation of regions with missing details caused by motion or overexposure. Qualitative and quantitative experiments demonstrate that, compared with existing methods, the proposed approach achieves improvements of 0.40 dB in PSNR-u and 0.048 % in HDR-VDP-2 on the Kalantari dataset, and exhibits strong suppression capability of ghosting artifacts in HDR imaging for dynamic scenes.