Multi-Scale Dilated Deformable Convolutional Network for High Dynamic Range Imaging

Existing methods for high dynamic range (HDR) imaging from multi-exposure low dynamic range (LDR) images usually align LDR images and gamma-corrected images uniformly, after which the reference and aligned non-reference features are directly concatenated and fed into the fusion network. However, the LDR domain is not suitable for detecting small misalignments. Additionally, direct concatenation will bring information redundancy and weaken the complementary information from non-reference images. To address these problems, we propose an end-to-end network based on multi-scale dilated deformable convolution. Firstly, a multi-scale dilated deformable (MDD) alignment module is designed to generate offsets from multi-scale HDR domain features, which are then used to align non-reference LDR and HDR domain features to reference features. To further suppress the misalignments in non-reference images while highlighting useful features, a dual attention module (DAM) is adopted. Finally, we employ a progressive fusion strategy to fully utilize the information under different exposures for continuously modulating the features from the reference image. Our method achieves superior evaluation results and outperforms the HDR-GAN by 0.2298dB, 0.0377dB, 0.0007dB, and 0.0001dB in terms of PSNR-l, PSNR-μ, SSIM-l and SSIM-μ, respectively, effectively mitigating ghosting artifacts in HDR imaging and retaining fine details.