Spaces:
Sleeping
Sleeping
File size: 4,443 Bytes
c509e76 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 |
# Dataset Preparation
The data files tree should be look like:
```
data/
eval/
dir300/
1_in.png
1_gt.png
...
kligler/
jung/
osr/
realdae/
docunet_docaligner/
dibco18/
train/
dewarping/
doc3d/
deshadowing/
fsdsrd/
tdd/
appearance/
clean_pdfs/
realdae/
deblurring/
tdd/
binarization/
bickly/
dibco/
noise_office/
phibd/
msi/
```
## Evaluation Dataset
You can find the links for downloading the dataset we used for evaluation (Tables 1 and 2) in [this](https://github.com/ZZZHANG-jx/Recommendations-Document-Image-Processing/tree/master) repository, including DIR300 (300 samples), Kligler (300 samples), Jung (87 samples), OSR (237 samples), RealDAE (150 samples), DocUNet_DocAligner (150 samples), TDD (16000 samples) and DIBCO18 (10 samples). After downloading, add the suffix of `_in` and `_gt` to the input image and gt image respectively, and place them in the folder of the corresponding dataset
## Training Dataset
You can find the links for downloading the dataset we used for training in [this](https://github.com/ZZZHANG-jx/Recommendations-Document-Image-Processing/tree/master) repository.
### Dewarping
- Doc3D
- Mask extraction: you should extract the mask for each image from the uv data in Doc3D
- Background preparation: you can download the background data from [here](https://www.robots.ox.ac.uk/~vgg/data/dtd/) and specify it for self.background_paths in `loaders/docres_loader.py`
- JSON preparation:
```
[
## you need to specify the paths of 'in_path', 'mask_path and 'gt_path':
{
"in_path": "dewarping/doc3d/img/1/102_1-pp_Page_048-xov0001.png",
"mask_path": "dewarping/doc3d/mask/1/102_1-pp_Page_048-xov0001.png",
"gt_path": "dewarping/doc3d/bm/1/102_1-pp_Page_048-xov0001.npy"
}
]
```
### Deshadowing
- RDD
- FSDSRD
- JSON preparation
```
[ ## you need to specify the paths of 'in_path' and 'gt_path', for example:
{
"in_path": "deshadowing/fsdsrd/im/00004.png",
"gt_path": "deshadowing/fsdsrd/gt/00004.png"
},
{
"in_path": "deshadowing/rdd/im/00004.png",
"gt_path": "deshadowing/rdd/gt/00004.png"
}
]
```
### Appearance enhancement
- Doc3DShade
- Clean PDFs collection: You should collection PDFs files from the internet and convert them as images to serve as the source for synthesis.
- Extract shadows from Doc3DShade by using `data/preprocess/shadow_extract.py` and dewarp the obtained shadows by using `data/MBD/infer.py`. Then you should specify self.shadow_paths in `loaders/docres_loader.py`
- RealDAE
- JSON preparation:
```
[
## for Doc3DShade dataset, you only need to specify the path of image from PDF, for example:
{
'gt_path':'appearance/clean_pdfs/1.jpg'
},
## for RealDAE dataset, you need to specify the paths of both input and gt, for example:
{
'in_path': 'appearance/realdae/1_in.jpg',
'gt_path': 'appearance/realdae/1_gt.jpg'
}
]
```
### Debluring
- TDD
- JSON preparation
```
[ ## you need to specify the paths of 'in_path' and 'gt_path', for example:
{
"in_path": "debluring/tdd/im/00004.png",
"gt_path": "debluring/tdd/gt/00004.png"
},
]
```
### Binarization
- Bickly
- DTPrompt preparation: Since the DTPrompt for binarization is time-expensive, we obtain it offline before training. Use `data/preprocess/sauvola_binarize.py`
- DIBCO
- DTPrompt preparation: the same as Bickly
- Noise Office
- DTPrompt preparation: the same as Bickly
- PHIDB
- DTPrompt preparation: the same as Bickly
- MSI
- DTPrompt preparation: the same as Bickly
- JSON preparation
```
[
## you need to specify the paths of 'in_path', 'gt_path', 'bin_path', 'thr_path' and 'gradient_path', for example:
{
"in_path": "binarization/noise_office/imgs/1.png",
"gt_path": "binarization/noise_office/gt_imgs/1.png",
"bin_path": "binarization/noise_office/imgs/1_bin.png",
"thr_path": "binarization/noise_office/imgs/1_thr.png",
"gradient_path": "binarization/noise_office/imgs/1_gradient.png"
},
]
```
After all the data are prepared, you should specify the dataset_setting in `train.py`.
|