FATN: Flow-Guided Alignment-Based Transformer Network for High Dynamic Range Imaging

High dynamic range (HDR) imaging presents a formidable challenge, aiming to produce ghost-free HDR images featuring lifelike details. The alignment-merger pipeline is a popular choice for this task. However, training deformable alignment is difficult, and existing methods struggle with ghost removal and detail recovery. In this work, a new flow-guided alignment-based transformer network (FATN) is proposed to cope with HDR imaging in dynamic scenes. To improve the training stability of deformable alignment, we propose a flow-guided deformable convolution module (FDCM). To jointly extract both global and local dependencies, we propose a dual-branch transformer merger layer (DTML) as the basic component of the feature merger network. DTML combines a global branch to model remote motions and a local branch for neighborhood content similarity. Extensive experiments showcase the superior performance of the proposed approach over other competing methods both quantitatively and qualitatively.