The Ghost-like artifact caused by motion and ill-exposed areas is one of the most challenging problems in high dynamic range (HDR) image reconstruction. Previous deep learning methods produce HDR results with high-quality in challenging scenes where large foreground motions exist via complicated network design. However, they are too resource-consuming when the resolution of the input image is large. To this end, an attention-guided network with inverse tone-mapping guided up-sampling is proposed for improving the efficiency. The proposed network consists of two main modules: fusion module (SK-AHDRNet) and inverse tone-mapping guided up-sampling module. In the fusion module, low-resolution low dynamic range (LDR) images are used to obtain low-resolution HDR via an attention guided network. In the up-sampling module, inverse tone-mapping is adopted to generate the High-resolution HDR from the reference image that guides the generation of the final HDR result. The proposed fusion module scores 43.17 with PSNR metric and 61.02 with HDR-VDP-2 metric on test which outperforms all conventional works. Compared with the plain fusion module, the running time of up-sampling network is reduced by 81.5% while the PSNR value is decreased from 43.17 to 38.18 with of the original resolution.
ASJC Scopus subject areas
- Media Technology
- Hardware and Architecture
- Computer Networks and Communications