Back to results
US 8,509,537 B2Granted
Learning weights of fonts for typed samples in handwritten keyword spotting
AssigneeXerox Corporation
InventorsFlorent C. Perronnin, Thierry Lehoux, Francois Ragnet
Life of this patent
Estimating the term…
Abstract
A wordspotting system and method are disclosed. The method includes receiving a keyword and, for each of a set of typographical fonts, synthesizing a word image based on the keyword. A keyword model is trained based on the synthesized word images and the respective weights for each of the set of typographical fonts. Using the trained keyword model, handwritten word images of a collection of handwritten word images which match the keyword are identified. The weights allow a large set of fonts to be considered, with the weights indicating the relative relevance of each font for modeling a set of handwritten word images.
Claims28
A method comprising:
- receiving a keyword;
- for each font of a set of typographical fonts, synthesizing a word image based on the keyword;
- training a keyword model based on the synthesized word images and a respective font weight for each font of the set of typographical fonts, the font weights being learned by maximizing a similarity between a weighted distribution of descriptors of synthesized training word images and a distribution of descriptors of handwritten training word images, each synthesized training word image descriptor being based on features extracted from a respective synthesized word image, each handwritten training word image descriptor being based on features extracted from a respective handwritten training word image;
- using the trained keyword model, identifying at least one handwritten word image of a collection of handwritten word images which matches the keyword;
- and
- at least one of the synthesizing, training, and identifying being performed with a computer processor.
3 more elements
A method comprising:
- receiving a keyword;
- for each font of a set of typographical fonts, synthesizing a word image based on the keyword;
- training a keyword model based on the synthesized word images and a respective font weight for each font of the set of typographical fonts, wherein the font weights are learned using a set of handwritten word image descriptors, each of the handwritten word image descriptors being based on features extracted from a respective handwritten word image drawn from a collection of handwritten word images;
- using the trained keyword model, identifying at least one handwritten word image of the collection of handwritten word images which matches the keyword;
- and
- at least one of the synthesizing, training, and identifying being performed with a computer processor.
3 more elements
A method comprising:
- learning font weights, including:
- training a first probabilistic model on descriptors extracted from a set of handwritten training word images;
- for each font of a set of typographical fonts, training a respective font probabilistic model on descriptors extracted from a set of synthesized training word images in the respective font;
- and
- computing a weight for each of the typographical fonts to maximize a similarity between the first probabilistic model and a mixture model of the font probabilistic models;
- receiving a keyword;
- for each font of the set of typographical fonts, synthesizing a word image based on the keyword;
- training a keyword model based on the synthesized word images of the keyword and the respective learned font weight for each font of the set of typographical fonts;
- using the trained keyword model, identifying at least one handwritten word image of a collection of handwritten word images which matches the keyword;
- and
- wherein at least one of the learning of font weights, synthesizing, training the keyword model, and identifying is performed with a computer processor.
8 more elements
A method comprising:
- learning font weights, including:
- for each font of a set of typographical fonts, training a respective font probabilistic model on descriptors extracted from a set of synthesized training word images in the respective font;
- scoring handwritten training word image descriptors extracted from handwritten training word images against each of the font probabilistic models and for each handwritten training word image descriptor, assigning the handwritten training word image descriptor to the font probabilistic model for which a respective score is highest;
- and
- generating font weights based on a number of the handwritten training word image descriptors assigned to each of the font probabilistic models;
- receiving a keyword;
- for each font of a set of typographical fonts, synthesizing a word image based on the keyword;
- training a keyword model based on the synthesized word images and the respective learned font weight for each font of the set of typographical fonts;
- using the trained keyword model, identifying at least one handwritten word image of a collection of handwritten word images which matches the keyword;
- and
- at least one of the learning, synthesizing, training of the keyword model, and identifying being performed with a computer processor.
8 more elements
A computer implemented processing system comprising:
- a synthesizer which synthesizes a word image in each of a set of typographical fonts based on a received keyword;
- a keyword model, stored in memory, which is trained on descriptors extracted from the synthesized word images and a set of generated font weights learned by maximizing a similarity between a weighted distribution of descriptors extracted from synthesized training word images and a distribution of descriptors extracted from handwritten training word images;
- a scoring component which scores handwritten word images of a collection of handwritten word images against the keyword model and, based on the scores, identifies any matching handwritten word images;
- and
- a computer processor for implementing the font weight generator, synthesizer, and scoring component.
2 more elements
A computer implemented method for wordspotting comprising:
- receiving a keyword to be searched for in a collection of handwritten word images extracted from one or more documents;
- for each of a set of fonts, automatically generating a synthesized word image based on the keyword;
- modeling the keyword with a semi-continuous hidden Markov model which takes into account weights assigned to each of the fonts, the weights being learned by maximizing a similarity between a weighted distribution of descriptors of synthesized training word images and a distribution of descriptors of handwritten training word images, each descriptor being based on features extracted from a respective synthesized or handwritten training word image;
- with the model, identifying matching word images in the collection of handwritten word images;
- and
- outputting information based on the identified matching word images.
3 more elements
Dependencies are inferred from the claim text.