Efficient High Dynamic Range Image Reconstruction Network for Dynamic Scenes

Addressing issues related to high complexity and prolonged inference time observed in dynamic scene reconstruction High Dynamic Range (HDR) image models, we propose an efficient network incorporating a spatial-channel attention mechanism and threshold segmentation. The network leverages a spatial-channel attention mechanism for feature alignment to mitigate ghosting effects. Additionally, it utilizes Spatial and Channel Reconstruction Convolution (SCConv) as a replacement for standard convolution to enhance model performance and reduce overall complexity. To filter out areas lacking information in abnormally exposed images, the network introduces a threshold segmentation algorithm. This step aims to extract more pertinent features for HDR image reconstruction. The network achieves a Peak Signal-to-Noise Ratio (PSNR) score of 37.73 dB and a Peak Signal-to-Noise Ratio-μ (PSNR-μ) score of 36.76 dB on the NTIRE 2022 HDR Challenge dataset. While the PSNR score may be suboptimal compared to other state-of-the-art methods, the network excels in the three indicators of Giga Multiply-Add Operations (GMACs), Params, and Runtimes, all of which are optimal.