Learning Degradation Uncertainty for Unsupervised Real-world Image Super-resolution
Authors: Qian Ning, Jingzhu Tang, Fangfang Wu, Weisheng Dong, Xin Li, Guangming Shi
IJCAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on popular real-world datasets show that our proposed method has performed better than other unsupervised approaches. |
| Researcher Affiliation | Academia | Qian Ning1 , Jingzhu Tang1 , Fangfang Wu1 , Weisheng Dong1 , Xin Li2 and Guangming Shi1 1Xidian University 2West Virginia University ningqian@stu.xidian.edu.cn, tangjingzhu@stu.xidian.edu.cn, ffwu xd@163.com, wsdong@mail.xidian.edu.cn, xin.li@mail.wvu.edu, gmshi@xidian.edu.cn |
| Pseudocode | No | The paper provides architectural diagrams (Figure 1, Figure 3) but does not include structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide any explicit statement or link regarding the availability of open-source code. |
| Open Datasets | Yes | In our experiments, only 200 degraded LR images collected by the Canon camera are used as real LR images, while the DIV2K dataset provides HR images to train the DSN. Additionally, the Real SR validation set is used for evaluation. NTIRE 2020 RWSR [Lugmayr et al., 2020] challenge offers two tracks for unsupervised SR training. The HR images from DIV2K are used as HR images in both tracks. In Track1, the synthetic degraded Flickr2K dataset is treated as real LR images. |
| Dataset Splits | Yes | Additionally, the Real SR validation set is used for evaluation. Furthermore, Track1 provides a validation dataset for quantitative comparison, which contains 100 images with the same degradation as the training LR images. |
| Hardware Specification | No | The paper does not provide any specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments. |
| Software Dependencies | No | The paper does not provide specific software dependencies with version numbers (e.g., Python 3.x, PyTorch 1.x, CUDA x.x). |
| Experiment Setup | Yes | In our implementation, we calculated the perceptual loss on VGG19 features of the conv5 4 convolutional layer. ... LDSN = α1Lkl + α2Lkl per + α3Ladv, where α1 = 1, α2 = 0.01, α3 = 0.01. We randomly select 16 RGB HR patches sized by 256 256 as batch input. The initial learning rate is 0.0001 and decreases by half for every 100 epochs. We train the model for 500 epochs. ... LSRN = β1LSR con + β2LSR per + β3LSR adv, where β1 = 0.01, β2 = 1, β3 = 0.005. We randomly select 16 RGB LR patches sized by 64 64 as batch input. The initial learning rate is 0.0001. We adopt an exponential moving average (EMA) for more stable training and better performance. We train the model for 1000 epochs. |