Ultra-High-Definition Low-Light Image Enhancement: A Benchmark and Transformer-Based Method

Authors: Tao Wang, Kaihao Zhang, Tianrun Shen, Wenhan Luo, Bjorn Stenger, Tong Lu

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct systematic benchmarking studies and provide a comparison of current LLIE algorithms. As a second contribution, we introduce LLFormer, a transformer-based low-light enhancement method. Extensive experiments on the new dataset and existing public datasets show that LLFormer outperforms state-of-the-art methods.
Researcher Affiliation Collaboration Tao Wang1, Kaihao Zhang2, Tianrun Shen1, Wenhan Luo3*, Bjorn Stenger4, Tong Lu1* 1State Key Lab for Novel Software Technology, Nanjing University, China 2 Australian National University, Australia 3 Shenzhen Campus of Sun Yat-sen University, China 4 Rakuten Institute of Technology, Japan
Pseudocode No The paper includes architectural diagrams and mathematical formulations, but no explicit pseudocode or algorithm blocks.
Open Source Code Yes The source code and pre-trained models are available at https://github.com/Tao Wangzj/LLFormer.
Open Datasets Yes To build this dataset of image pairs, we use normal-light 4K and 8K images from public data (Zhang et al. 2021). These UHD images were crawled from the web and captured by various devices.
Dataset Splits Yes UHD-LOL includes two subsets, UHD-LOL4K and UHD-LOL8K, containing 4K and 8K-resolution images, respectively. The UHD-LOL4K subset contains 8, 099 image pairs, 5, 999 for training and 2, 100 for testing. The subset of UHD-LOL8K includes 2, 966 image pairs, 2, 029 for training and 937 for testing.
Hardware Specification No The paper does not provide specific details on the hardware used for running experiments, such as GPU or CPU models.
Software Dependencies No The paper mentions using the Adam optimizer but does not specify software dependencies (e.g., programming languages, libraries, frameworks) with version numbers.
Experiment Setup Yes The LLFormer is trained on 128 128 patches with a batch size of 12. For data augmentation, we adopt horizontal and vertical flips. We use the Adam optimizer with an initial learning rate of 10 4 and decrease it to 10 6 using cosine annealing. The numbers of encoder blocks in the LLFormer from stage 1 to stage 4 are {2, 4, 8, 16}, and the number of attention heads in A-MSA are {1, 2, 4, 8}. The numbers corresponding to decoders from stage 1 to 3 are {2, 4, 8} and {1, 2, 4}.