Network-based system and method for detection of faces and the like

• US 5,642,431 A
• Filed: 06/07/1995
• Issued: 06/24/1997
• Est. Priority Date: 06/07/1995
• Status: Expired due to Fees

• Current Assignee: Massachusetts Institute of Technology
• Original Assignee: Massachusetts Institute of Technology
• Inventors: Sung, Kah-Kay, Poggio, Tomaso
• Primary Examiner(s): Couso, Jose L.
• Assistant Examiner(s): Do, Anh Hong
• Application Number:
  US08/478,395
• Time in Patent Office: 748 Days
• Field of Search: 382/118, 382/227, 382/317, 382/115, 235/380, 348/77
• US Class Current: 382/118
• CPC Class Codes: G06F 16/5854 using shape and object relationship
• CPC Class Codes: G06F 18/214 Generating training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06F 18/24137 Distances to cluster centroïds
• CPC Class Codes: G06V 40/164 using holistic features

Abstract

A face detection system (100) includes an imaging device, a computer having a pattern prototype synthesizer and an image classifier, and an output display device. Face and non-face pattern prototypes are synthesized by a network training process using a number of example images. A new applied image is analyzed by taking a distance from the applied image to each of the prototypes, and a decision is made as to whether the applied image contains a face based on the distances thus obtained. The system and method may be applied to various other detection tasks.

Claims

1. A method of detecting the presence of a predetermined pattern type in an applied image, the method comprising:

(a) synthesizing a network by;

(i) generating a first set of prototypes representative of the predetermined pattern type;

(ii) generating a second set of prototypes not representative of the predetermined pattern type by operating the network with the first set of prototypes, identifying a set of problem images yielding false positives when applied to the network, and using the problem images to form the second set of prototypes; and

(iii) training the network by application of training images, for which it is known whether each such training image is representative of the predetermined pattern type; and

(b) operating the network by;

(i) preprocessing the applied image to obtain a preprocessed applied image;

(ii) determining a distance metric between the preprocessed applied image and each of the first and second sets of prototypes to obtain a plurality of distance metrics;

(iii) applying said plurality of distance metrics to said network; and

(iv) producing a detection signal indicative of whether the applied image is representative of the predetermined pattern type.

2. A method as in claim 1, wherein the training further comprises determining a distance metric between each of the training images and each of the first and second sets of prototypes to obtain a plurality of training distance metrics, and applying said plurality of training distance metrics to said network.

3. A method as in claim 1, wherein the distance metric includes a first component based on a Mahalanobis-based distance measure.

4. A method as in claim 3, wherein the distance metric includes a second component based on a Euclidean-based distance measure.

5. A method as in claim 1, wherein training the network further includes operating the network, identifying additional problem images yielding false positives when applied to the network, and using the additional problem images to augment the second set of prototypes.

6. A method as in claim 1, wherein the preprocessing includes masking, image resizing, brightness correction, and histogram equalization.

7. A method as in claim 1, further comprising scaling and scanning the applied image into a plurality of uniformly sized windows for application to the network.

8. A method as in claim 1, wherein said applied image is one of a plurality of overlapping portions of an overall image, wherein the image classifier further determines a strength of detection based on how many of said overlapping portions when applied to said image classifier result in said detection signal being indicative of the presence of said predetermined pattern type.

9. A method as in claim 1, wherein said first set of prototypes is generated from a set of example images by collecting the set of example images and by collecting a second set of images obtained by taking a mirror image of each of the set of example images.

10. A method as in claim 1, wherein said first set of prototypes is generated from a set of example images by collecting the set of example images and by collecting a third set of images obtained by taking a rotated version of each of the set of example images.

11. A method as in claim 1, wherein said image is one of a plurality of overlapping portions of an overall image, wherein said method further includes operating said network with respect to each of said overlapping portions, and wherein said method further includes determining a strength of detection based on how many of said overlapping portions result in said detection signal being indicative of said predetermined pattern type.

12. A system for detecting the presence of a predetermined pattern type in an applied image, the system comprising:

a pattern prototype synthesizer including a preprocessor, an image database, and a prototype generator, the preprocessor normalizing and filtering a plurality of example patterns, the image database storing the plurality of example patterns and having a first section and a second section, the first section containing first example patterns representative of the predetermined pattern type and the second section containing second example patterns not representative of the predetermined pattern type, the second example patterns being generated from images applied to the system for which the predetermined pattern type is falsely detected, and the prototype generator forming prototype patterns based on the example patterns; and

an image classifier operatively coupled to the pattern prototype synthesizer, the image classifier including a distance metric processor producing a distance measure representative of a distance between the applied image and each of the prototype patterns, and a network operatively coupled to the distance metric processor, the network accepting as input the distance measure and producing as output a detection signal representative of the presence of the predetermined pattern type in the applied image.

13. A system as in claim 12, wherein said preprocessor further normalizes and filters the applied image.

14. A system as in claim 12, wherein the distance measure produced by the distance metric processor includes a first Mahalanobis-based component.

15. A system as in claim 14, wherein the distance measure produced by the distance metric processor includes a second Euclidean-based component.

16. A system as in claim 12, wherein the preprocessor includes a window resizer, a masker, a brightness corrector, and a histogram equalizer.

17. A system as in claim 12, wherein said preprocessor is adapted to scale and scan said applied image to produce a plurality of uniformly sized pattern windows, and wherein said image classifier is adapted to accept as input each of said pattern windows.

18. A system as in claim 12, wherein said prototype generator forms prototype patterns from mirror images of said example patterns.

19. A system as in claim 12, wherein said prototype generator forms prototype patterns from rotated versions of said example patterns.