Systems and methods for robust industrial optical character recognition

Patent metadata

Field	Value
Publication number	US-10810465-B2
Application number	US-201816024910-A
Country	US
Kind code	B2
Filing date	Jul 1, 2018
Priority date	Jun 30, 2017
Publication date	Oct 20, 2020
Grant date	Oct 20, 2020

Abstract

An auto-encoder is configured to verify character detection and/or classification results generated by an automated optical character recognition system. The auto-encoder may be trained to reconstruct visual representations of the detected character, and a determination of whether the character detection result comprises a true positive or false positive may be based on a reconstruction error between the image data in which the character was detected and a reconstructed image generated by the auto-encoder.

First claim

We claim:

1. A method of operation in an image processing system for recognizing one or more characters in an image, the method comprising:
receiving a classifier result indicating detection of a character at specified coordinates within the image;
retrieving input image data corresponding to the specified coordinates within the image;
generating a reconstruction of the input image data by use of an auto-encoder trained to replicate the detected character; and
determining whether the classifier result is a true positive based on a reconstruction error between the input image data and the reconstruction of the input image data, wherein generating the reconstruction of the input image data comprises:
processing the input image data through a plurality of encoder layers of the auto-encoder to produce an inner layer representation of the input image data; and
decoding the inner layer representation by use of a plurality of decoder layers of the auto-encoder to generate the reconstruction of the input image data from the inner layer representation.

2. The method of claim 1 , further comprising calculating the reconstruction error between the input image data and reconstructed image data generated by the auto-encoder.

3. The method of claim 2 , wherein calculating the reconstruction error comprises computing a root-mean-square error between the input image data and the reconstructed image data.

4. The method of claim 1 , further comprising invalidating the classifier result in response to the reconstruction error exceeding a determined verification threshold.

5. A method of operation in an image processing system for recognizing one or more characters in an image, the method comprising:
receiving a classifier result indicating detection of a character at specified coordinates within the image;
retrieving input image data corresponding to the specified coordinates within the image;
generating a reconstruction of the input image data by use of an auto-encoder trained to replicate the detected character;
determining whether the classifier result is a true positive based on a reconstruction error between the input image data and the reconstruction of the input image;
acquiring a plurality of different sets of input image data corresponding to the classifier result;
generating a plurality of reconstruction images by use of the auto-encoder, each reconstruction image corresponding to a respective one of the different sets of input image data;
calculating a plurality of reconstruction errors, each reconstruction error between a respective one of the different sets of the input image data and corresponding reconstruction image; and
determining whether the classification result comprises a true positive based on a minimum reconstruction error of the plurality of reconstruction errors.

6. The method of claim 5 , wherein generating the plurality of different sets of input image data comprises retrieving respective sets of input image data from the image within an enlarged area around the specified coordinates, the enlarged area corresponding to a size of one or more ground-truth images used to train the auto-encoder.

7. A method of operation in an image processing system for recognizing one or more characters in an image, the method comprising:
receiving a classifier result indicating detection of a character at specified coordinates within the image;
retrieving input image data corresponding to the specified coordinates within the image;
generating a reconstruction of the input image data by use of an auto-encoder trained to replicate the detected character;
determining whether the classifier result is a true positive based on a reconstruction error between the input image data and the reconstruction of the input image;
invalidating the classifier result in response to the reconstruction error exceeding a determined verification threshold; and
determining the reconstruction error based on reconstruction errors of one or more ground-truth images used to train the auto-encoder.

8. An image processing system, comprising:
an auto-encoder processor, comprising:
a plurality of encode layers configured to generate a compressed representation of input image data corresponding to a character classification result,
a plurality of decode layers configured to generate a reconstructed image from the compressed representation of the input image data, and
a difference calculator configured to determine a reconstruction error between the input image data and the reconstructed image,

wherein the auto-encoder processor is configured to determine whether the character classification result is one of a true positive and a false positive based on the reconstruction error.

9. The system of claim 8 , wherein:
each encode layer of the plurality of encode layers comprises a respective neural network comprising a plurality of neurons; and
each decode layer of the plurality of decode layers comprises a respective neural network comprising a plurality of neurons.

10. The system of claim 9 , wherein:
each neuron of a first encode layer of the plurality of encode layers comprises a two-dimensional convolutional filter, and
each neuron of the encode layers other than the first encode layer comprises a one-dimensional convolutional filter.

11. The system of claim 9 , wherein:
the plurality of encode layers comprises a first encode layer comprising a first number of neurons, and second encode layer comprising a second number of neurons, and a third encode layer comprising a third number of neurons;
the first number is larger than the second number; and
the second number is larger than the third number.

12. The system of claim 9 , wherein:
the plurality of decode layers comprises a first decode layer comprising a first number of neurons, and second decode layer comprising a second number of neurons, and a third decode layer comprising a third number of neurons;
the third number is larger than the second number; and
the second number is larger than the first number.

13. The system of claim 8 , wherein:
the auto-encoder processor is configured to invalidate the character classification result in response to determining that the character classification result is a false positive;
invalidating the character classification result comprises one or more of: removing the character classification result from result data, marking the character classification result as a false positive, and recording that the character classification result is a false positive on a non-transitory storage medium.

14. The system of claim 8 , further comprising an auto-encoder network configuration engine adapted to train the encode layers and the decode layers to replicate a target character, the target character corresponding to a character of the character classification result.

15. The image processing system of claim 8 , further comprising an image capture device operably coupled with the auto-encoder processor and configured to capture the input image data.

16. A non-transitory computer-readable storage medium comprising instructions configured to cause a computing device to perform operations for improving accuracy of a character recognition system, the operations comprising:
receiving a detection result indicating detection of a character at a designated location within a captured image;
acquiring input image data corresponding to the designated location within the captured image;
configuring an auto-encoder network to replicate a target character corresponding to the character of the detection result;
reconstr

Assignees

• Datalogic Usa Inc

Inventors

• Goncalves Luis
• Belgiovine Mauro

Classifications

• G06V30/1916

  Validation; Performance evaluation · CPC title

• G06V30/12

  Detection or correction of errors, e.g. by rescanning the pattern · CPC title

• G06V30/19173

  Classification techniques · CPC title

• G06V10/82

  using neural networks · CPC title

• G06F18/217

  Validation; Performance evaluation; Active pattern learning techniques · CPC title