Usage Instruction For ReBlurSR Dataset

July 12, 2024 · View on GitHub

Dataset Using

Generate ReBlurSR-Train and ReBlurSR-Test

Download the dataset files from the Google Drive
Unzip the files to the corresponding folders. The structure tree of the unzipped folders is as follows:
```
.
├── ALL_HR
|── ALL_mask
|── valid
|   ├── defocus
|   |   ├── LR
|   |   |   └── X4
|   └── motion
|       ├── LR
|       |   └── X4
|── area_class.npy
|── degree_class.npy
|── defocus_motion_class.npy
|── train_validation.npy
└── train_validation_split.py
```
File description for the unzipped folder:
- README.md: This file.
- ALL_HR: The high-resolution images of the ReBlurSR dataset, including the ReBlurSR-Train and ReBlurSR-Test subsets.
- ALL_mask: The blur map for the HR images in ALL_HR. For each sample, the blur map value is 0(0) if the pixel is blurred, and 1(255) if the pixel is non-blurred.
- valid: contains the LR versions of the ReBlurSR-Test subset. defocus and motion subfolders contain the defocus and motion subsubsets of the ReBlurSR-Test subset, respectively. Its structure tree is as follows:
```
valid
├── defocus
│   ├── LR
│   │   └── X4
└── motion
    └── LR
       └── X4
```
- area_class.npy: The category of the area of the blur region in the ReBlurSR-Test subset. Samples are divided into 3 classes: small, medium, and large.
```
    0:"large",      1:"medium",     2:"small
```
- degree_class.npy: The category of the degree of the blur region in the ReBlurSR-Test subset. Samples are divided into 3 classes: heavy, little, and middle.
```
    0:"heavy",      1:"little",     2:"middle"
```
- defocus_motion_class.npy: The category of the blur type in the ReBlurSR-Test subset. Samples are divided into 2 classes: defocus and motion.
```
    0:"defocus",    1:"motion"
```
- train_validation.npy: The category of the samples in the ReBlurSR-Train and ReBlurSR-Test subsets. Samples are divided into 2 classes: train and validation.
```
    0:"train",      1:"validation"
```
- train_validation_split.py: The script to generate the ReBlurSR-Train and ReBlurSR-Test subsets from ALL_HR and ALL_mask folders according to the defocus_motion_class.npy and train_validation.npy files.
run the train_validation_split.py script to generate the ReBlurSR-Train and ReBlurSR-Test subsets.
```
python train_validation_split.py # required packages: numpy, tqdm
```

The generated ReBlurSR-Train and ReBlurSR-Test subsets are saved in the train and valid folders, respectively. The structure tree of the complete train folder and valid folder are as follows:

train
├── motion
│   ├── HR
│   └── mask
└── defocus
    ├── HR
    └── mask

valid
├── motion
│   ├── HR
│   ├── LR
|   |   └── X4
│   └── mask
└── defocus
    ├── HR
    ├── LR
    |   └── X4
    └── mask

Generate the subsets of the ReBlurSR-Test

You can generate the subsets of the ReBlurSR-Test according to the area_class.npy, degree_class.npy, and defocus_motion_class.npy files.