Personal identity verification process and system
AssigneeOmniperception Limited
InventorsJosef Kittler, Terry Windeatt
Life of this patent
Estimating the term…
Abstract
A personal identity verification process and system adopts an error correcting output code (ECOC) classifier approach. Image data representing an input facial images is processed to generate a first (ECOC) classifier output vector and image data representing a class of different training images of the face of an individual are processed to generate a set of second ECOC classifier output vectors. Authenticity of the input facial image against the class of training images is evaluated by measuring distance between the first ECOC classifier output vector and the set of second ECOC classifier output vectors using a first order Minkowski metric.
Claims8
A personal identity verification process for evaluating authenticity of an input facial image against a class of different training images representing the face of an individual, the process including
- processing image data representing the input facial image to generate a first error correcting output code (ECOC) classifier output vector,
- processing image data representing the different training images of the class to generate a set of second ECOC classifier output vectors,
- comparing the first ECOC classifier output vector with the set of second ECOC classifier output vectors, wherein the comparing includes measuring a distance d(y) between the first ECOC classifier output vector and the set of second ECOC classifier output vectors using a first order Minkowski metric, the distance d(y) being given by the expression
- d
-
- (
- y
- _
- )
- =
- 1
- N
-
- ∑
- l
- =
- 1
- N
-
- ∑
- j
- =
- 1
- b
-
-
- y
- j
- l
- -
- y
- j
-
- ,
- where yj is the jth element of the first ECOC classifier output vector,
- yjl is the jth element of the lth member of the set of second ECOC classifier output vectors.
- N is the number of members in the set, and
- b is the number of classifiers, and
- the input facial image is evaluated to be authentic if the distance d(y)≦t, where t is a predetermined threshold value, and
- evaluating authenticity of the input facial image with respect to the class of training images in dependence on the comparing.
A personal identification verification process for evaluating authenticity of an input facial image against the class of different training images representing the face of an individual, the process including:
- processing image data representing the input facial image to generate a first error correcting output code (ECOC) classifier output vector,
- processing image data representing the different training images of the class to generate a set of second ECOC classifier output vectors, including processing image data representing a plurality of the classes of different training images to generate a respective set of second ECOC classifier output vectors for each class, each class of training images representing the face of a different individual,
- comparing the first ECOC classifier output vector with the set of second ECOC classifier output vectors by measuring a distance between the first ECOC classifier output vector and each set of second ECOC classifier output vectors, and
- evaluating authenticity of the input facial image with respect to the class of training images in dependence on the comparing by combining distances measured to evaluate authenticity of input facial image with respect to one of classes of training images. including combining distances measured as a combination that is a summation of exponential kernel functions, each kernel function having an exponent related to a different respective distance, wherein the combination is given by the expression:
- k
-
- (
- y
- _
- )
- =
- ∑
- α
-
-
- w
- α
-
- exp
- -
- {
- d
- α
-
- (
- y
- _
- )
- σ
- 2
- }
- where α is an index signifying class,
- dα(y) is the measured distance between the first ECOC classifier output vector and a respective set of second ECOC classifier output vectors and
- Wαis a respective predetermined weighing function, and
- the input facial image is evaluated to be authentic if k(y)≦tl, where tl is a predetermined threshold value.
Dependencies are inferred from the claim text.