Graphical languages, such as musical notation, use the 2D arrangement of symbols to convey meaning, using convention, abstraction and the fundamental sign-signified relationship of language.
The relative location of symbols is meaningful, such as in this UML diagram example, as well as their shape or direction, which can alter meaning.
SignWriting is a graphical language used to transcribe the gestures and movements of Sign Languages, using the 2D page to capture the multimodal 3D reality of signing.
Quevedo
Quevedo is a python library and command line application for creating, annotating and managing datasets of graphical languages, with a focus on the training and evaluation of machine learning algorithms for their recognition.
Features
- Dataset management, including hierarchical dataset organization, subset partitioning, and semantically guided data augmentation.
- Structural and visual annotation of source images using a web interface, with support for different users and the live visualization of data processing scripts.
- Deep learning network management, training, configuration and evaluation, using darknet.
Install & Use
- Quevedo requires
python >= 3.7, and can be installed from PyPI:
$ pip install quevedo[web]- Create a dataset, visually annotate it using your own browser, and train and evaluate neural networks:
$ quevedo -D new/dataset create ; cd new/dataset
$ quevedo web
$ quevedo -N network_name prepare train test- More documentation available at https://agarsev.github.io/quevedo.
Deep Learning
Machine learning techniques developed in the field of computer vision are necessary to adequately process graphical languages. While the researcher can use any toolkit and algorithm they prefer, Quevedo includes a module to facilitate the use of deep learning neural networks with Quevedo datasets.
Pipelines
Quevedo allows you to train different neural networks to recognize different objects and features. These networks can then be composed into a pipeline to build an expert system, capable of performing a bigger task than each of the networks by themselves.
For example, a detection network can first locate the graphemes within a logogram, and then specialist networks be used to classify each of the graphemes.