Learning Semantic Degradation-Aware Guidance for Recognition-Driven Unsupervised Low-Light Image Enhancement
Authors: Naishan Zheng, Jie Huang, Man Zhou, Zizheng Yang, Qi Zhu, Feng Zhao
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate its effectiveness for improving the ULLIE approaches on the downstream recognition tasks while maintaining a competitive visual quality. ... Extensive experiments show the effectiveness of our SDAG in instructing existing ULLIE methods to produce improved enhancement results that elevate the visual recognition performance significantly while preserving the visual quality. ... Experimental Setup Low-Light Image Classification. In the low-light image classification task, due to the lack of a dedicated low-light image classification dataset, we choose the CUB (Wah et al. 2011) dataset to simulate the low-light images for evaluation. ... We investigate the effectiveness and consistency of the proposed SDAG in performance improvement across two recognition models, i.e., VGG16 (Simonyan and Zisserman 2014) and Alex Net (Krizhevsky, Sutskever, and Hinton 2012). ... Based on the Zero-DCE network, we conduct experiments to demonstrate the effectiveness of the proposed SDAG and evaluate the classification performance on the CUB dataset by the VGG recognition model. |
| Researcher Affiliation | Academia | Naishan Zheng*, Jie Huang*, Man Zhou , Zizheng Yang, Qi Zhu, Feng Zhao University of Science and Technology of China, Hefei, China {nszheng, hj0117, manman, yzz6000, zqcrafts}@mail.ustc.edu.cn, fzhao956@ustc.edu.cn |
| Pseudocode | No | The paper describes the algorithm steps in text and flowcharts (e.g., Figures 3 and 4) but does not provide formal pseudocode blocks or sections labeled 'Algorithm'. |
| Open Source Code | Yes | Code will be available at https://github.com/zheng980629/SDAG. |
| Open Datasets | Yes | In the low-light image classification task, due to the lack of a dedicated low-light image classification dataset, we choose the CUB (Wah et al. 2011) dataset to simulate the low-light images for evaluation. ... Dark Face Detection. In dark face detection tasks, we evaluate the effectiveness of our SDAG on the DARK FACK (Yuan et al. 2019) dataset. |
| Dataset Splits | Yes | The training and testing sets in our experiments are identical to the official selection (Wah et al. 2011), which include 5994 images for training and 5794 images for testing. ... It consists of 10,000 dark images, with 6000 images for training/validation and 4000 for testing. Due to the unavailable bounding boxes of the testing set, we randomly select 500 images from the training and validation sets for performance comparison while the remaining 5,500 for training. |
| Hardware Specification | Yes | We implemented our SDAG with Py Torch on a single NVIDIA GTX 2080Ti. |
| Software Dependencies | No | The paper mentions 'Py Torch' as the implementation framework but does not provide specific version numbers for PyTorch or any other software dependencies. |
| Experiment Setup | Yes | We use the Adam optimizer with β1 = 0.9, β2 = 0.99 for a total of 300K iterations. The filter weights are randomly initialized with the standard Gaussian function. The initial learning rate is set to 2 10 4 and reduced by 0.5 every 50k iteration. The batch and patch sizes are set to 16 and 64 64, respectively. |