US 11,810,374 B2 · Training text recognition systems

Abstract

In implementations of recognizing text in images, text recognition systems are trained using noisy images that have nuisance factors applied, and corresponding clean images (e.g., without nuisance factors). Clean images serve as supervision at both feature and pixel levels, so that text recognition systems are trained to be feature invariant (e.g., by requiring features extracted from a noisy image to match features extracted from a clean image), and feature complete (e.g., by requiring that features extracted from a noisy image be sufficient to generate a clean image). Accordingly, text recognition systems generalize to text not included in training images, and are robust to nuisance factors. Furthermore, since clean images are provided as supervision at feature and pixel levels, training requires fewer training images than text recognition systems that are not trained with a supervisory clean image, thus saving time and resources.

Claims20

Dependent claims

In a digital medium environment to train a text recognition system, a method implemented by a computing device, the method comprising:

receiving a noisy image including text having at least one distortion artifact that alters a visual appearance of the text;
receiving an ideal text image having clean text corresponding to the text of the noisy image;
determining a reconstructed clean image by removing one or more nuisance factors from the noisy image that are a basis of the at least one distortion artifact of the text in the noisy image;
generating a text prediction of the text in the noisy image;
and
training the text recognition system based on the text prediction, a difference between the reconstructed clean image without the one or more nuisance factors and the ideal text image, and a difference between the noisy image and the ideal text image.

3 more elements

A text recognition system implemented at least partially in computer hardware in a digital medium environment, the text recognition system comprising:

a noisy image module configured to receive a noisy image including text having at least one distortion artifact that alters a visual appearance of the text;
a clean image module configured to determine a reconstructed clean image by removing one or more nuisance factors from the noisy image that are a basis of the at least one distortion artifact of the text in the noisy image;
a text prediction module configured to generate a text prediction of the text in the noisy image;
and
a training module configured to train the text prediction module to recognize image text based on the text prediction, a difference between the reconstructed clean image without the one or more nuisance factors and an ideal text image, and a difference between the noisy image and the ideal text image.

2 more elements

In a digital medium environment to train a text recognition system, a method implemented by a computing device, the method comprising:

receiving a noisy image including text having at least one distortion artifact that alters a visual appearance of the text;
determining a reconstructed clean image by removing one or more nuisance factors from the noisy image that are a basis of the at least one distortion artifact of the text in the noisy image;
extracting features from the noisy image and additional features from the reconstructed clean image;
and
training the text recognition system based on a text prediction of the text in the noisy image, and based on a feature-matching loss term determined from a difference between the extracted features from the noisy image and the additional extracted features from the reconstructed clean image without the one or more nuisance factors.

2 more elements

Dependencies are inferred from the claim text.

Training text recognition systems

Life of this patent

Training text recognition systems

Life of this patent

Abstract

Claims20