A Lightweight Network for High Dynamic Range Imaging

Multi-frame high dynamic range (HDR) reconstruction methods try to expand the range of illuminance with differently exposed images. They suffer from ghost artifacts when camera jittering or object moving. Several methods can generate high-quality HDR images with high computational complexity, but the inference process is too slow. However, the network with small parameters will produce unsatisfactory results. To balance the quality and computational complexity, we propose a lightweight network for HDR imaging that has small parameters and fast speed. Specifically, following AHDRNet, we employ a spatial attention module to detect the misaligned regions to avoid ghost artifacts. Considering the missing details in over-/under-exposure regions, we propose a dual attention module for selectively retaining information to force the fusion network to learn more details for degenerated regions. Furthermore, we employ an encoder-decoder structure with a lightweight block to achieve the fusion process. As a result, the high-quality content and features can be reconstructed after the attention module. Finally, we fuse high-resolution features and the encoder-decoder features into the HDR imaging results. Experimental results demonstrate that the proposed method performs favorably against the state-of-the-art methods, achieving a PSNR of 39.05 and a PSNR-μ of 37.27 with 156.12 GMAcs in NTIRE 2022 HDR Challenge (Track 2 Fidelity).