Process images containing calligraphy samples for pix2pix training.
400 characters from Zhao Mengfu's Thousand Character Classic in Regular and Cursive Script (《赵孟頫真草千字文》) (download here) were labeled and ~200 were used in training to produce the results in the demo. However, any Chinese calligraphy or calligraphic imagery can potentially be labeled using this tool.
To annotate an image, first cd into ./tools and
python annotate.py image.png
Use -s option to specify width of a single character (in pixels). Default is 220.
First press s to mark the boundaries of a new character, then press a to start a new polyline. Move your cursor to draw, and press a again to end the polyline. Pan the window with arrow keys. Press z to undo and o to save your progress.
Information about each annotated character, including polyline coordinates, source image path, and stroke order, will be stored in out.json
Once all the charaters are annotated, use
python data.py
to convert all the data into pix2pix-tensorflow-ready, side-by-side, annotation-on-the-right, 256p, images. The output will be stored in dataset folder.
Use pix2pix-tensorflow for training. Detailed instructions can be found in the said project's README.md.