High Dynamic Range Image Reconstruction of Dynamic Scenes via Hybrid CNN and Flatten-Transformer

Aiming to address the ghost removal issue in reconstructing high dynamic range (HDR) images within dynamic scenes, we propose HDR-CFT, a network designed for HDR image reconstruction. HDR-CFT effectively captures both local and global features from multi-exposure images, leveraging the strengths of both CNN and Flatten-Transformer. Initially, a CNN-based feature extraction network is employed to extract features and align input multi-exposure images. Subsequently, a Context-Aware Flatten-Transformer Block is constructed using Flatten-Transformer and deformable convolution for HDR image fusion. Additionally, SCConv is integrated into the network to eliminate model redundancy and enhance overall representation capability. Extensive experiments conducted on datasets provided by Kalantari et al. and the NTIRE 2021 HDR Challenge demonstrate the superior performance of our proposed method compared to current state-of-the-art techniques in terms of reconstruction quality, yielding high-fidelity HDR images.