Degrade Is Upgrade: Learning Degradation for Low-Light Image Enhancement
Authors: Kui Jiang, Zhongyuan Wang, Zheng Wang, Chen Chen, Peng Yi, Tao Lu, Chia-Wen Lin1078-1086
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on both the enhancement task and joint detection task have verified the effectiveness and efficiency of our proposed method, surpassing the SOTA by 0.70d B on average and 3.18% in m AP, respectively. |
| Researcher Affiliation | Academia | Kui Jiang1, Zhongyuan Wang1 , Zheng Wang1, Chen Chen2, Peng Yi1, Tao Lu3, Chia-Wen Lin4 1 School of Computer Science, Wuhan University, Wuhan, China 2 University of Central Florida 3 Wuhan Institute of Technology 4 National Tsing Hua University |
| Pseudocode | No | The paper describes the model architecture and procedures but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The code will be available soon. |
| Open Datasets | Yes | Following the setting in Retinex Net (Wei et al. 2018), we use the LOL dataset for training, which contains 500 low/normal-light image pairs (480 for training and 20 for evaluation). ... we also introduce two novel low/normal-light datasets (COCO24700 for training and COCO1000 for evaluation) based on COCO (Caesar, Uijlings, and Ferrari 2018) dataset |
| Dataset Splits | Yes | Following the setting in Retinex Net (Wei et al. 2018), we use the LOL dataset for training, which contains 500 low/normal-light image pairs (480 for training and 20 for evaluation). |
| Hardware Specification | Yes | We use the Adam optimizer with a batch size of 16 for training DRGN on a single NVIDIA Titan Xp GPU. |
| Software Dependencies | No | The paper mentions using the Adam optimizer but does not specify any software names with version numbers (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | In our baseline, the pyramid layer is empirically set to 3, corresponding to the number of RCABs and RCAB depths of [2, 3, 4] and [3, 3, 3], respectively, in the residual group. The training images are cropped to non-overlapping 96 96 patches to obtain sample pairs. Standard augmentation strategies, e.g., scaling and horizontal flipping are applied. We use the Adam optimizer with a batch size of 16 for training DRGN on a single NVIDIA Titan Xp GPU. The learning rate is initialized to 5 10 4 and then attenuated by 0.9 every 6,000 steps. After 60 epochs on training datasets, we obtain the optimal solution with the above settings. Specifically, we train De G for the first 20 epochs and then optimize Re G for 40 epochs. |