HDR-NFlow: High Dynamic Range imaging with normalizing flow

Shuaikang Shang, Xuejing Kang, Anlong Ming

Abstract: High Dynamic Range (HDR) imaging aims to generate a high-quality HDR image by fusing multi-exposure Low Dynamic Range (LDR) images. When input LDR images have large object motion and severe saturation, previous methods suffer from ghosting artifacts, which results in unpleasant HDR images and hinders real-world applications. To address this critical issue to reconstruct high-quality HDR images, we propose a novel HDR imaging framework based on the normalizing flow (called HDR-NFlow), which regards HDR imaging as a conditional generation task and consists of a conditional encoder and an invertible flow network. Specifically, the conditional encoder utilizes the proposed Composite Attention Merge Module (CAMM) to capture long-range context and fusion dependency of multi-exposed frames to align the large object motions and an Asymmetric Selective Kernel Detail (ASKD) module to capture texture information via locally stripy extraction. With the extracted features as reasonable conditions, the invertible flow network hallucinates realistic content for saturated regions and generates an HDR image by realizing the conversion of Gaussian distribution to HDR image distribution. We conduct extensive experiments on commonly used benchmark datasets to demonstrate that our method achieves state-of-the-art performance both quantitatively and qualitatively.

External IDs:dblp:journals/ijon/ShangKM25