LA-HDR:

Light Adaptive HDR Reconstruction Framework for Single LDR Image Considering Varied Light Conditions

Xiangyu Hu, Liquan Shen, Mingxing Jiang, Ran Ma and Ping An School of Communication and Information Engineering, Shanghai University

This is the official repository for LA-HDR: Light Adaptive HDR Reconstruction Framework.

We provide two types of reference implementations for LA-HDR: TensorFLow & PyTorch. Please check the respective repos for details.

Meanwhile, considering the ambiguities in evaluating HDR-VDP2, we provide our HDR-VDP2 evaluation implementation for reference. See ./hdrvdp-2.2.1/metric_vdp2.m for details.

Notes

1. LA-HDR is strongly suggested to be trained seperately by each sub-networks sequentially.

2. As the whole solution is focused on HDR reconstruction, we do not put much efforts on low-light denoising, thus the denoising effect of DeNet is rather poor. The pre-trained model for DeNet is not provided in PyTorch version. Users may evaluate its performance in TensorFlow one.

3. The training parameters in each training script are provided only as the reference setup.

4. Users are suggested to develop a better denoising solution to replace DeNet, or maybe just replace it with solutions like CycleISP, etc. Recent low-light denoising solutions tend to consider the denoising problem in the high bit-depth / RAW domain, therefore, EnhanceNet is designed in LA-HDR for helping such denoising measures. The hist_block module in EnhanceNet can be considered to deal with various ISP strategies. Such improvement are not investigated in this LA-HDR version.

Citing

Plain Text:

X. Hu, L. Shen, M. Jiang, R. Ma and P. An, "LA-HDR: Light Adaptive HDR Reconstruction Framework for Single LDR Image Considering Varied Light Conditions," IEEE Transactions on Multimedia, 2022.

BibTex:

@ARTICLE{lahdr_2022,
  author={Hu, Xiangyu and Shen, Liquan and Jiang, Mingxing and Ma, Ran and An, Ping},
  journal={IEEE Transactions on Multimedia}, 
  title={LA-HDR: Light Adaptive HDR Reconstruction Framework for Single LDR Image Considering Varied Light Conditions}, 
  year={2022},
  pages={1-16},
  doi={10.1109/TMM.2022.3183404}}