Automatic biometric identification based on face recognition and support vector machines

• US 20090074259A1
• Filed: 07/29/2005
• Published: 03/19/2009
• Est. Priority Date: 07/29/2005
• Status: Active Grant

• Current Assignee: Telecom Italia S.p.A.
• Original Assignee: Telecom Italia S.p.A.
• Inventors: Grosso, Enrico, Bicego, Manuele, Tistarelli, Massimo, D'Amico, Roberta, Baltatu, Madalina, D'Alessandro, Rosalia
• Granted Patent: US 8,275,175 B2
• Time in Patent Office: Days
• US Class Current: 382/118
• CPC Class Codes: G06F 18/2433 Single-class perspective, e.g. one-against-all classification; Novelty detection; Outlier detection
• CPC Class Codes: G06V 10/764 using classification, e.g. of video objects
• CPC Class Codes: G06V 40/172 Classification, e.g. identification

Abstract

An automatic biometric identification method based on face recognition and support vector machines, includes enrolling a user to generate a user'"'"'s reference template; and identifying the user based on the user'"'"'s reference template, wherein generating a user'"'"'s reference template includes acquiring a number of user'"'"'s face images, and training a one-class support vector machine based on the user'"'"'s face images only.

Claims

1-29. -29. (canceled)

30. An automatic biometric identification method based on face recognition and support vector machines, comprises:

generating a user'"'"'s reference template; and

identifying the user based on the user'"'"'s reference template;

wherein generating a user'"'"'s reference template comprises;

acquiring a number of user'"'"'s face images; and

training a support vector machine based on the user'"'"'s face images;

the support vector machine being a one-class support vector machine trained based on user'"'"'s face images only.

31. The method of claim 30, wherein training the one-class support vector machine comprises:

extracting a feature vector from each user'"'"'s face image; and

training the one-class support vector machine based on the extracted feature vectors.

32. The method of claim 31 wherein extracting a feature vector from each user'"'"'s face image comprises:

computing a transform in the frequency domain of each user'"'"'s face image to generate a respective spectrum; and

forming a feature vector for each user'"'"'s face image based on coefficients of frequencies of the respective spectrum.

33. The method of claim 32, wherein the transform is a Fourier-based transform.

34. The method of claim 33, wherein each feature vector comprises coefficients of significant frequencies of the respective spectrum.

35. The method of claim 33, wherein the transform is a Fourier-Mellin transform.

36. The method of claim 33, wherein the transform is a Fourier transform.

37. The method of claim 36, wherein each feature vector comprises coefficients of low frequencies of the respective spectrum.

38. The method of claim 37, wherein each feature vector comprises the coefficients of the lowest 27 up to 30 frequencies of the spectrum.

39. The method of claim 37, wherein each feature vector comprises a coefficient of a continuous component, the real parts of the coefficients of the lowest 27 frequencies, and imaginary parts of the coefficients of the lowest 27 frequencies of the respective spectrum.

40. The method of claim 31, wherein extracting a feature vector from each user'"'"'s face image comprises:

dividing each user'"'"'s face image into a number of regions; and

extracting a feature vector from at least some regions in the user'"'"'s face image.

41. The method of claim 40, wherein extracting a feature vector from at least some of the regions in the user'"'"'s face image comprises:

computing a number of local binary pattern histograms for each region; and

forming a feature vector for each region based on the local binary pattern histograms computed for the region.

42. The method of claim 41, further comprising:

weighting the feature vectors based on the significance of the face portions in the respective regions.

43. The method of claim 41, further comprising:

training a number of one-class support vector machines equal to the number of regions in the user'"'"'s face images, each one-class support vector machine being trained based on the feature vectors extracted from corresponding region in the user'"'"'s face images.

44. The method of claim 31, wherein extracting a feature vector from each user'"'"'s face image comprises:

forming the feature vectors with the intensity value of at least some pixels in the user'"'"'s face image.

45. The method of claim 31, wherein the one-class support vector machine is based on a kernel function with a variance which is set substantially equal to the average Euclidean distance between the feature vectors.

46. The method of claim 31, wherein the one-class support vector machine operates by mapping the feature vectors into a high-dimensional feature space and by finding a hypersphere that encompasses most of the feature vectors extracted from user'"'"'s face images, the hypersphere having a given center and a given radius defining a nominal acceptance threshold for user identification;

further comprising;

adjusting the radius of the hypersphere on a per user basis to reduce false rejections during the user identification.

47. The method of claim 46, wherein adjusting the radius of the hypersphere comprises:

extracting feature vectors from a new training set on a per user basis;

computing the average value of the distances from the feature vectors extracted from the new user training set to the boundary of the user'"'"'s hypersphere and the corresponding standard deviation; and

computing an effective acceptance threshold based on the radius of the hypersphere and the computed average distance and standard deviation.

48. The method of claim 47, wherein the average distance and the standard deviation are computed during the user enrollment phase.

49. The method of claim 46, further comprising:

adjusting the center of the hypersphere to reduce the false acceptance rate computed for the user.

50. The method of claim 30, wherein identifying the user based on the user'"'"'s reference template comprises:

acquiring a number of user'"'"'s face images;

extracting feature vectors from the user'"'"'s face images; and

identifying the user based on extracted feature vectors, the user'"'"'s reference template, and a one-class support vector machine trained for the user.

51. The method of claim 49, wherein identifying the user comprises:

inputting the extracted feature vectors and the user'"'"'s reference template into the one-class support vector machine trained for the user, the one-class support vector machine computing the distance between each feature vector and the center of the hypersphere; and

identifying the user based on the computed distances.

52. The method of claim 50, wherein identifying the user based on the computed distances comprises:

comparing each computed distance with the threshold acceptance; and

identifying the user based on the outcomes of comparisons.

54. A computer program product capable, when loaded and run in a processing system, of implementing the method according to claim 30.

53. An automatic biometric identification system based on face recognition and support vector machines, comprising:

a module for generating a user'"'"'s reference template, the module comprising a support vector machine;

a module for identifying the user based on the user'"'"'s reference template;

wherein said module for generating a user'"'"'s reference template comprises;

a module for acquiring a number of user'"'"'s face images; and

a module for training the support vector machine based on the user'"'"'s face images;

the support vector machine being a one-class support vector machine trained based on user'"'"'s face images only.