System, method and application for the recognition, verification and similarity ranking of facial or other object patterns

• US 6,128,398 A
• Filed: 02/09/1998
• Issued: 10/03/2000
• Est. Priority Date: 01/31/1995
• Status: Expired due to Fees

• Current Assignee: Viisage Technology, Inc. (Advent International Corporation)
• Original Assignee: Miros, Inc.
• Inventors: Kuperstein, Michael, Kottas, James A.
• Primary Examiner(s): COUSO, YON JUNG
• Application Number:
  US09/020,443
• Time in Patent Office: 967 Days
• Field of Search: 382/116, 382/117, 382/118
• US Class Current: 382/118
• CPC Class Codes: G06V 40/172 Classification, e.g. identification
• CPC Class Codes: G07C 9/253 visually

Abstract

A system and method is disclosed for determining the likelihood that two object patterns arise from the same object source. It is able to do this without having previously been exposed to either of the object patterns. This system utilizes an adaptive processor trained to make this determination using a large number of example patterns in different views and orientations. It accommodates large verifications of how an object source is presented in an image. The system also does not require the storage of any information about any particular pattern, which greatly minimizes the storage requirements and improves the throughput of the system. Further, there is no need for accessing a database of previously stored features from a given pattern source. The system is particularly useful for object patterns which consist of facial images. The system employs a new technique for locating an object of interest within a pattern using an adaptive processor to determine a region of interest. The technique is resistant to irrelevant overlapping patterns. The system and method of the present invention employ a technique for performing authentication of the validity of a card or a user'"'"'s image on a computer network system which has a user'"'"'s face image stored within. Finally, this invention provides a new technique for naturally aligning the face. This technique, which is convenient to use and yields candid shots, is useful in facial verification.

Claims

1. A system for determining the likelihood that two object patterns arise from the dame object source, said system comprising:

a. means for receiving first and second object patterns each including a plurality of feature components having a plurality of values;

b. Means for assigning a predetermined weight to each said feature component in each of said object patterns, wherein the assigned weights are one of a plurality of weights in corresponding first and second weight sets in a neural network, wherein each weight set includes a plurality of weight subsets, each weight subset corresponding to the location of one of said feature components, wherein particular weights in each subset are assigned according to the value of the component; and

c. means for determining an output of said neural network by calculating a comparison function of the assigned weights in the two respective weight sets, wherein the values of the weights are predetermined such that the output is a measure of the likelihood that two object patterns arise from the same object source, wherein the value of said predetermined weights are derived from a training procedure and are independent of the value of the feature component.

2. The system of claim 1 further comprising a learning unit which comprises:

means for generating sets of training pairs of first and second training object patterns derived from known object sources, the training pairs comprising first and second training object patterns that vary from each other, some of which derive from the same object source;

means for adjusting the value of said assigned weights of the first and second weight sets to have a greater difference for said training pairs arising from two different object sources that generate an output incorrectly indicating the same source; and

means for adjusting the value of said assigned wights of the first and second sets to have a smaller difference measure for said pairs arising from the same object source that generate an output incorrectly indicating different sources.

3. The system of claim 1 further comprising:

means for defining a bounding region around said object pattern within a larger pattern, the region wherein portions outside of the bounding region are discarded.

4. The system of claim 3 further comprising a neural network for deriving the location of said bounding region adaptively.

5. The system of claim 3 wherein said object patterns each comprise human facial images and further comprising

means for locating the position of the eyes in said human face;

and

means for rotating the facial image about a fixed point until the eyes are level horizontally, and registering the two object patterns based on said eye positions.

6. The system of claim 1 further comprising:

means for discretizing each feature value in discrete bins as a function of the value'"'"'s proportion of the mean of the feature values; and

means for assigning each feature value to one of a set of possible weight values in said weight set based on said discretizing.

7. A method for determining the likelihood that two object patterns arise from the same object source, said method comprising:

a. receiving first and second object patterns each including a plurality of feature components having different values;

c. assigning a predetermined weight to each said feature component in each of said object patterns, wherein the assigned weights are one of a plurality of weights in corresponding first or second weight sets in a neural network, wherein each weight set includes a plurality of weight subsets, each weight subset corresponding to the location of one of said feature components, wherein particular weights in each subset are assigned according to the value of the component; and

d. determining an output of said neural network by calculating a comparison function of the assigned weights in the two respective weight sets, wherein the values of the weights are predetermined such that the output is a measure of the likelihood that two object patterns arise from the same object source, and wherein the value of said predetermined weights are derived from a training procedure and are independent of the value of the feature component.

8. The method of claim 7 further comprising performing a training procedure to determine said weight values, the training procedure comprising the steps of:

generating sets of training pairs of first and second training object patterns derived from known object sources, the training pairs comprising first and second training object patterns that vary from each other, some of which derive from the same object source;

adjusting the value of said assigned weights of the first and second weight sets to have a greater difference measure for said training pairs arising from two different objects sources that generate an output incorrectly indicating the same source; and

adjusting the value of said assigned weights of the first and second weight sets to have a smaller difference measure for said training pairs arising from the same object source that generate an output incorrectly indicating different sources.

9. The method of claim 7 further comprising:

defining a bounding region around said object pattern within a larger pattern, the region wherein portions outside of the bounding region are discarded.

10. The method of claim 9 further comprising a neural network for deriving the location of said bounding region adaptively.

11. The method of claim 9 further comprising:

locating the position of the eyes, by determining the average eye position; and

rotating the facial image about the average eye position until the eyes are level horizontally, and registering the two object patterns based on said eye positions.

12. The method of claim 7 further comprising:

ranking each feature value as a function of the value'"'"'s proportion of the mean of the feature values; and

assigning each feature value to one of a set of possible weight values in said weight set based on said ranking.