Time delay neural network for printed and cursive handwritten character recognition

• US 5,105,468 A
• Filed: 04/03/1991
• Issued: 04/14/1992
• Est. Priority Date: 04/03/1991
• Status: Expired due to Term

• Current Assignee: American Telephone & Telegraph Company (AT&T, Inc.)
• Original Assignee: AT&T, Inc.
• Inventors: Guyon, Isabelle, LeCun, Yann, Denker, John S.
• Primary Examiner(s): BRINICH, STEPHEN M
• Application Number:
  US07/679,988
• Time in Patent Office: 377 Days
• Field of Search: 382/10, 382/14-15, 382/37-39, 382/13
• US Class Current: 382/158
• CPC Class Codes: G06V 30/10 Character recognition
• CPC Class Codes: G06V 30/18057 Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
• CPC Class Codes: G06V 30/347 Sampling; Contour coding; Stroke extraction

Abstract

A time delay neural network is defined having feature detection layers which are constrained for extracting features and subsampling a sequence of feature vectors input to the particular feature detection layer. Output from the network for both digit and uppercase letters is provided by an output classification layer which is fully connected to the final feature detection layer. Each feature vector relates to coordinate information about the original character preserved in a temporal order together with additional information related to the original character at the particular coordinate point. Such additional information may include local geometric information, local pen information, and phantom stroke coordinate information relating to connecting segments between the end point of one stroke and the beginning point of another stroke.

The network is also defined to increase the number of feature elements in each feature vector from one feature detection layer to the next. That is, as the network is reducing its dependence on temporally related features, it is increasing its dependence on more features and more complex features.

Claims

1. Apparatus for outputting one of a plurality of predetermined symbols in response to an input written character wherein the character is represented as a temporally related sequence of frames, each frame having a plurality of feature elements, the apparatus comprising,

means responsive to the sequence of frames for generating a second sequence of frames having a length N and including frames of phantom information selected from the group consisting of pen-up strokes and cyclic closure strokes;

time-delay neural network means including hidden layer means for subsampling and detecting constrained features from the second sequence of frames to create a complex frame sequence having a length M, where M is at most substantially equal to N/2, and output layer means fully connected to the hidden layer means for identifying the one of the plurality of predetermined symbols which matches most closely the input written character.

2. The apparatus as defined in claim 1 wherein the hidden layer means comprises,

first hidden sublayer means for subsampling and detecting constrained features from the second sequence of frames to generate a third frame sequence having length K, where K is at most substantially equal to N/2;

second hidden sublayer means for subsampling and detecting constrained features from the third frame sequence to generate a fourth frame sequence having length J, where J is at most substantially equal to K/2; and

third hidden sublayer means for subsampling and detecting constrained features from the fourth frame sequence to generate the complex frame sequence.

3. The apparatus as defined in claim 1 wherein the complex frame sequence includes R features per frame and wherein the second sequence of frames includes S features per frame, R being greater than or equal to S.

4. A time delay neural network for outputting one of a plurality of predetermined symbols in response to an input written character wherein the character is represented as a temporally related sequence of frames, each frame having a plurality of feature elements, the network comprising hidden layer means for subsampling and detecting constrained features from the temporally related sequence of frames to create a complex frame sequence having a length M, where M is at most substantially equal to N/2, and output layer means responsive to the complex frame sequence and being fully connected to the hidden layer means for identifying the one of the plurality of predetermined symbols which matches most closely the input written character.

5. The apparatus as defined in claim 4 wherein the hidden layer means comprises,

first hidden sublayer means for subsampling and detecting constrained features from the temporally related sequence of frames to generate a second frame sequence having length K, where K is at most substantially equal to N/2;

second hidden sublayer means for subsampling and detecting constrained features from the second frame sequence to generate a third frame sequence having length J, where J is at most substantially equal to K/2; and

third hidden sublayer means for subsampling and detecting constrained features from the third frame sequence to generate the complex frame sequence.

6. The apparatus as defined in claim 4 wherein the complex frame sequence includes R features per frame and wherein the second sequence of frames includes S features per frame, R being greater than or equal to S.

7. The apparatus as defined in claim 5 wherein the complex frame sequence includes R features per frame and wherein the second sequence of frames includes S features per frame, R being greater than or equal to S.