Innovative OCR systems and methods that combine a template based generative model with a discriminative model
View Patent ↗A method employing a hybrid classification model is used to perform optical character recognition operations for an image. Image data from the image is provided to a generative classification model of the hybrid model, and generative image classifications operations are performed, generating a feature data set which is outputted from the generative classification model. This feature data set is then provided to the discriminative classification model, and discriminative classification operations are performed to generate a classification of the image.
1. A method of performing optical character recognition operations on an image:
providing image data of the image to a generative classification model;
using a computer processor, performing generative image classification operations on the image data by the generative classification model;
outputting a feature data set of the image data;
providing the feature data set to a discriminative classification model;
performing discriminative classification operations on the provided feature data set; and
generating a classification of the image;
wherein the feature data set is a feature vector defined as:
ϕ
c
,
i
,
j
(
I
g
)
=
-
I
g
(
i
,
j
)
-
ϑ
c
(
i
,
j
)
ϑ
c
(
i
,
j
)
(
1
-
ϑ
c
(
i
,
j
)
)
(
3
)
where I g (i,j) is the pixel value at location (i,j) of the observed character image, and θ C (i,j) is an estimated value derived from the character template of character c at Location (i,j).
2. The method of claim 1 , wherein the generative classification model is based on character templates.
3. The method of claim 2 , wherein the generative classification model employs a likelihood-based font template.
4. The method of claim 1 , wherein the generative classification model is a parametric generative classification model.
5. The method of claim 4 , wherein the discriminative classification model employs a kernel, and the kernel of the discriminative classification model is a function of outputs matching font templates with an observed character image of the image.
6. The method of claim 4 , wherein a kernel of the discriminative classification model is a function of a Fisher score of the generative classification model, given as:
K ( I s ,I t )=exp(γ∥Φ( I s )−Φ( I t )∥) (7),
where K denotes the function, I s is one image patch, I t is a second image patch, γ is an adjustable parameter, and Φ is a feature vector.
7. The method of claim 1 , wherein the discriminative classification model employs a kernel, and the kernel of the discriminative classification model is a function of the output of the generative classification model.
8. A hybrid classification model for optical character recognition, the hybrid classification model comprising:
a generative classification model stored in computer memory; and
a discriminative classification model, combined with the generative classification model and configured to use a function data set from the generative classification model;
wherein the feature data set is a feature vector defined as:
ϕ
c
,
i
,
j
(
I
g
)
=
-
I
g
(
i
,
j
)
-
ϑ
c
(
i
,
j
)
ϑ
c
(
i
,
j
)
(
1
-
ϑ
c
(
i
,
j
)
)
,
where I g (i,j) is the pixel value at location (i,j) of the observed character image, and θ(l) is an estimated value derived from the character templates.
9. The hybrid classification model of claim 8 , wherein the generative classification model is a parametric generative classification model.
10. The hybrid classification model of claim 8 , wherein the generative classification model is based on character templates.
11. The hybrid classification model of claim 10 , wherein a kernel of the discriminative classification model is a function of a Fisher score of the generative classification model, given as:
K ( I s ,I t )=exp(γ∥Φ( I s )−Φ( I t )∥) (7),
where K denotes the function, I s is one image patcn, I t is a second image patch, γ is an adjustable parameter, and Φ is a feature vector.
12. The hybrid classification model of claim 8 , wherein the discriminative classification model employs a kernel, and the kernel of the discriminative classification model is a function of the output of the generative model.
13. The hybrid classification model of claim 8 , wherein the generative classification model employs a likelihood-based font template.
14. The hybrid classification model of claim 13 , wherein the discriminative classification model employs a kernel, and the kernel of the discriminative classification model is a function of outputs matching font templates with an observed character image.
15. A computer program product, operative in a data processing system and implemented on a computer readable medium for classifying an image comprising:
providing image data of the image to a generative classification model;
performing generative image classification operations on the image data by the generative classification model;
outputting a feature data set of the image data;
providing the feature data set to a discriminative classification model;
performing discriminative classification operations on the provided feature data set; and
generating a classification of the image;
wherein the feature data set is a feature vector defined as:
ϕ
c
,
i
,
j
(
I
g
)
=
-
I
g
(
i
,
j
)
-
ϑ
c
(
i
,
j
)
ϑ
c
(
i
,
j
)
(
1
-
ϑ
c
(
i
,
j
)
)
(
3
)
where I g (i,j) is the pixel value at location (i,j) of the observed character image, and θ c (i,j) is an estimated value derived from the character template of character c at Location (i, j).