Back to results
US 12,651,384 B2Granted
Generating handwriting via decoupled style descriptors
AssigneeBrown University
InventorsAtsunobu Kotani, Stefanie Tellex, James Tompkin
Life of this patent
Estimating the term…
Abstract
A method of representing a space of handwriting stroke styles includes representing writer-, character- and writer-character-level style variations within a recurrent neural network (RNN) model using decoupled style descriptors (DSD) that model the style variations such that character style variations depend on writer style.
Claims20
A method of representing a space of handwriting stroke styles of a writer, the method comprising:
- representing writer-, character- and writer-character-level style variations within a recurrent neural network (RNN) model using decoupled style descriptors (DSD) that model the style variations such that character style variations depend on writer style by:
- receiving one or more stroke sequences comprising a plurality of timestamped points;
- generating, for the writer, a stroke dataset by assigning each of the one or more stroke sequences to a character of a plurality of characters;
- and
- learning, based on the stroke dataset, a representation that encodes writer-independent character representations, writer-dependent character-string style descriptors, and writer-dependent global style descriptors.
2 more elements
A method comprising:
- receiving one or more stroke sequences comprising, for each of a plurality of timesteps, a representation of a difference in spatial location, wherein each of the one or more stroke sequences correspond to written input of a writer;
- generating, for a writer, a stroke dataset by assigning each of the one or more stroke sequences to a character of a first plurality of characters from the written input;
- training, using the stroke dataset, a recurrent neural network for an encoder function to learn:
- writer-independent character representations,
- writer-dependent character-string style descriptors, and
- writer-dependent global style descriptors;
- and
- using the trained recurrent neural network to generate a second plurality of characters, wherein at least one character of the second plurality of characters is different from all characters of the first plurality of characters.
5 more elements
A method comprising:
- receiving one or more stroke sequences comprising, for each of a plurality of timesteps, a representation of a difference in spatial location, wherein each of the one or more stroke sequences correspond to written input of a writer;
- generating, for a first writer, a stroke dataset by assigning each of the one or more stroke sequences to a character of a first plurality of characters from the written input;
- training, using the stroke dataset, a recurrent neural network for an encoder function to learn:
- writer-independent character representations,
- writer-dependent character-string style descriptors, and
- writer-dependent global style descriptors;
- and
- using the trained recurrent neural network to generate a second plurality of characters emulating writing of a second writer different from the first writer.
5 more elements
Dependencies are inferred from the claim text.