Open source library to work with the CASIA Chinese Handwriting dataset.
PyCasia is on the Python Package Index, so installation is as easy as pip install pycasia. Requires Python 3.5 or newer.
The pycasia.CASIA object is the interface for all the data. You can use it to explore the dataset, or use it as a base class for a more complicated use.
Datasets are directories full of data files from a given distribution. They come in isolated character (GNT) or handwritten text (DGR) files. Four are automatically downloaded by the library, but you can add more.
These datasets are downloaded from the publicly available data hosted on the project webpage. You should expect a long download during your first run.
HWDB1.1trn_gnt_P1 and HWDB1.1trn_gnt_P2 are two parts of the publicly available set for training applications. They were split for easy downloading.
HWDB1.1tst_gnt is the training portion of that set.
competition-gnt is the dataset from some Chinese handwriting competitions.
To add other datasets, add a new dictionary in the datasets variable of the CASIA object. You will need to include the download URL and the dataset type, either GNT or DGR. If you have data that isn't publicly available, make sure there is a folder named after the dataset in the base dataset directory, and the download code won't be called.
Example:
CASIA.datasets["competition-gnt"] = {
"url": "http://www.nlpr.ia.ac.cn/databases/Download/competition/competition-gnt.zip",
"type": "GNT"
}
You can download all datasets using the get_all_datasets() method, or just specific datasets using the get_dataset(dataset) method.
On OS X and Linux, datasets are stored in ~/CASIA_data. On Windows, they're saved in the CASIA_data in your home directory. If you want to save the data in a diffent location, specify a path when you create the CASIA object. Eg: dataset = CASIA(path="/CASIA_data")
You can load all of the character image (GNT) data using the load_character_images() method, or a particular dataset using the load_dataset(dataset) method. If you want to read the data on a file by file basis, just use the static CASIA.load_gnt_file() method to get the data.
These are generators yielding data as (image, label) pairs. The images are Pillow.Image.Image objects.
You may want to explore the data by yourself. You can get the data as JPEGs by calling the get_raw() function. You can then inspect the data to your leisure.
You can build your own class to implement more complicated usage of the dataset. Just inherit from CASIA.
Early release. Features may change. Can open individual character images (GNT files) but not sentences. So far, no plans to develop readers to use DGR files or online datasets. Pull requests welcome.
The datasets are hosted in mainland China, and are often difficult to download from other countries, as the connection gets reset. get_dataset attempts the download five times, but sometimes that doesn't work. You can try again, or download the data manually. WGET has been effective for manual downloads.
While useful for many applications, the publicly available data is only a fraction of the total set. If you need more, you should fill out an application form from the projects maintainers to get the full set.
The datasets are only licensed for research use, and certainly no commercial use. If you want to publish your data, you should fill out an application form from the projects maintainers. You should not host the data in any form, including in your repository.