Text image processing using word spacing equalization for icr system employing artificial neural network
AssigneeKonica Minolta Laboratory U.S.A., Inc.
InventorSaman Sarraf
Life of this patent
Estimating the term…
Abstract
In an intelligent character recognition (ICR) method for recognizing hand-written text images using a long-short term memory (LSTM) recurrent neural network (RNN), text images are segmented into text line images, and the text lines images are pre-processed to normalize the line height and to equalize the word spacings in each text line. Both training images used to train the RNN network and test images containing text to be recognized by the trained RNN network are pre-processed to have identical heights and identical word spacings between words. This method improves character recognition accuracy.
Claims12
A hand-written text character recognition method using an artificial neural network, comprising:
- obtaining training data containing a plurality of training text line images and corresponding label data, each training text line image being a binary image containing one line of hand-written text, the label data containing digital values representing characters contained in each training text line image;
- normalizing a height of each training text line image to a pre-determined standard line height, the standard line height being identical for all training text line images;
- performing word spacing equalization for each training text line image to generate a corresponding pre-processed training text line image, including:
- identifying a plurality of word spacings in the training text line images;
- and
- modifying the training text line image to change all word spacings to a pre-determined standard length, without changing any areas of the training text line image between adjacent word spacings, to generate the pre-processed training text line image, the standard length being identical for all training text line images;
- and
- training an untrained long-short term memory (LSTM) neural network using the pre-processed training text line images and the label data to obtain a trained LSTM neural network.
A hand-written text character recognition method using an artificial neural network, comprising:
- obtaining a test text line image, the test text line image being a binary image containing one line of hand-written text;
- normalizing a height of the test text line image to a pre-determined standard line height;
- performing word spacing equalization for the test text line image to generate a pre-processed test text line image, including:
- identifying a plurality of word spacings in the test text line images;
- and
- modifying the test text line image to change all word spacings to a pre-determined standard length, without changing any areas of the test text line image between adjacent word spacings, to generate the pre-processed test text line image;
- and
- processing the pre-processed test text line image using a trained long-short term memory (LSTM) neural network to recognize characters contained in the test text line image.
A computer program product comprising a non-transitory computer readable memory device having a computer readable program code embedded therein for controlling a data processing apparatus, the computer readable program code configured to cause the data processing apparatus to execute a text character recognition process, the process comprising:
- receiving training data containing a plurality of training text line images and corresponding label data, each training text line image being a binary image containing one line of hand-written text, the label data containing digital values representing characters contained in each training text line image;
- normalizing a height of each training text line image to a pre-determined standard line height, the standard line height being identical for all training text line images;
- performing word spacing equalization for each training text line image to generate a corresponding pre-processed training text line image, including:
- identifying a plurality of word spacings in the training text line images;
- and
- modifying the training text line image to change all word spacings to a pre-determined standard length, without changing any areas of the training text line image between adjacent word spacings, to generate the pre-processed training text line image, the standard length being identical for all training text line images;
- and
- training an untrained long-short term memory (LSTM) neural network using the pre-processed training text line images and the label data to obtain a trained LSTM neural network.
Dependencies are inferred from the claim text.