TRAINING METHOD AND APPARATUS FOR NEUTRAL NETWORK FOR IMAGE RECOGNITION

• US 20170061246A1
• Filed: 09/01/2016
• Published: 03/02/2017
• Est. Priority Date: 09/02/2015
• Status: Active Grant

• Current Assignee: Fujitsu Limited
• Original Assignee: Fujitsu Limited
• Inventors: SATOSHI, Naoi, CHEN, Li, WANG, Song, FAN, Wei, SUN, Jun
• Granted Patent: US 10,296,813 B2
• Time in Patent Office: Days
• US Class Current: 1/1
• CPC Class Codes: G06F 18/213 Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
• CPC Class Codes: G06F 18/214 Generating training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06F 18/28 Determining representative reference patterns, e.g. by averaging or distorting; Generating dictionaries
• CPC Class Codes: G06N 3/08 Learning methods
• CPC Class Codes: G06T 2207/10004 Still image; Photographic image
• CPC Class Codes: G06T 2207/20084 Artificial neural networks [ANN]
• CPC Class Codes: G06T 5/20 using local operators
• CPC Class Codes: G06V 10/772 Determining representative reference patterns, e.g. averaging or distorting patterns; Generating dictionaries

Abstract

A training method and a training apparatus for a neutral network for image recognition are provided. The method includes: representing a sample image as a point set in a high-dimensional space, a size of the high-dimensional space being a size of space domain of the sample image multiplied by a size of intensity domain of the sample image; generating a first random perturbation matrix having a same size as the high-dimensional space; smoothing the first random perturbation matrix; perturbing the point set in the high-dimensional space using the smoothed first random perturbation matrix to obtain a perturbed point set; and training the neutral network using the perturbed point set as a new sample. With the training method and the training apparatus, classification performance of a conventional convolutional neural network is improved, thereby generating more training samples, reducing influence of overfitting, and enhancing generalization performance of the convolutional neural network.

Claims

1. A training method for a neutral network for image recognition, comprising:

representing a sample image as a point set in a high-dimensional space, a space size of the high-dimensional space being a domain size of a space domain of the sample image multiplied by an intensity size of an intensity domain of the sample image;

generating a first random perturbation matrix having a same size as the high-dimensional space;

smoothing the first random perturbation matrix;

perturbing the point set in the high-dimensional space using a smoothed first random perturbation matrix to obtain a perturbed point set; and

training the neutral network using the perturbed point set as a new sample.

2. The training method according to claim 1, wherein the domain size of the space domain is a width of the sample image multiplied by a height of the sample image, and the intensity size of the intensity domain is D1×

D2×

. . . ×

Dn, where n is a dimension of the intensity domain, D1, D2, . . . , Dn are sizes of respective dimensions.

3. The training method according to claim 1, wherein the training the neutral network using the perturbed point set as a new sample comprises:

projecting the perturbed point set into one of a two-dimensional and three-dimensional sub-space of the high-dimensional space to obtain a projected image, and training the neutral network using the projected image as a new sample.

4. The training method according to claim 1, wherein the first random perturbation matrix is generated using a given probability distribution, the given probability distribution being one of a Gaussian distribution and a uniform distribution.

5. The training method according to claim 1, wherein the first random perturbation matrix is smoothed using a low-pass filter having a same dimension as the first random perturbation matrix, the low-pass filter being one of a Gaussian filter, a mean filter and a middle value filter.

6. The training method according to claim 1, wherein the perturbing the point set in the high-dimensional space using the smoothed first random perturbation matrix to obtain a perturbed point set comprises:

adding a value of a corresponding element in the smoothed first random perturbation matrix to each point in the point set in the high-dimensional space.

7. The training method according to claim 1, further comprising:

generating a second random perturbation matrix;

smoothing the second random perturbation matrix; and

perturbing, using the smoothed second random perturbation matrix, feature maps output by a convolution layer in training the neutral network,

wherein a matrix size of the second random perturbation matrix is a length of a feature map multiplied by a width of the feature map multiplied by a number of the feature maps.

8. The training method according to claim 7, wherein the perturbing, using the smoothed second random perturbation matrix, feature maps outputted by a convolution layer in training the neutral network comprises:

adding a corresponding weighted value of the second random perturbation matrix to a value of the feature map, wherein a weighting coefficient is larger if an absolute value of the feature map is larger.

19. A non-transitory computer readable medium storing a training method for a neutral network for image recognition as recited in claim 1.

9. A training method for a neutral network for image recognition, comprising:

transforming a sample image to obtain a transform domain representation of the sample image;

representing a transformed sample image as a point set in a high-dimensional space;

generating a random perturbation matrix having a same size as the high-dimensional space;

smoothing the random perturbation matrix;

perturbing the point set in the high-dimensional space using the smoothed random perturbation matrix to obtain a perturbed point set;

projecting the perturbed point set into a transform domain sub-space of the high-dimensional space to obtain a transform domain representation of a perturbed sample image;

inversely transforming the transform domain representation of the perturbed sample image to obtain a perturbed image; and

training the neutral network using the perturbed image as a new sample.

10. The training method according to claim 9, wherein the transform is one of a Discrete Flourier Transform and a Discrete Cosine Transform.

11. A training apparatus for a neutral network for image recognition, comprising:

a high-dimensional space representing unit configured to represent a sample image as a point set in a high-dimensional space, a space size of the high-dimensional space being a domain size of a space domain of the sample image multiplied by an intensity size of an intensity domain of the sample image;

a first random perturbation matrix generating unit configured to generate a first random perturbation matrix having a same size as the high-dimensional space;

a first smoothing unit configured to smooth the first random perturbation matrix;

a first perturbing unit configured to perturb the point set in the high-dimensional space using the smoothed first random perturbation matrix to obtain a perturbed point set; and

a training sample determining unit configured to train the neutral network using the perturbed point set as a new sample.

12. The training apparatus according to claim 11, wherein the domain size of the space domain is a width of the sample image multiplied by a height of the sample image, and the intensity size of the intensity domain is D1×

D2×

. . . ×

Dn, where n is a dimension of the intensity domain, D1, D2, . . . , Dn are sizes of respective dimensions.

13. The training apparatus according to claim 11, wherein the training sample determining unit is further configured to:

project the perturbed point set into one of a two-dimensional and three-dimensional sub-space of the high-dimensional space to obtain a projected image, and train the neutral network using the projected image as a new sample.

14. The training apparatus according to claim 11, wherein the first random perturbation matrix is generated using a given probability distribution, the given probability distribution being one of a Gaussian distribution and a uniform distribution.

15. The training apparatus according to claim 11, wherein the first smoothing unit is further configured to smooth the first random perturbation matrix using a low-pass filter having a same dimension as the first random perturbation matrix, the low-pass filter being one of a Gaussian filter, a mean filter and a middle value filter.

16. The training apparatus according to claim 11, wherein the first perturbing unit is further configured to:

add a value of a corresponding element in the smoothed first random perturbation matrix to each point in the point set in the high-dimensional space.

17. The training apparatus according to claim 11, further comprising:

a second random perturbation matrix generating unit configured to generate a second random perturbation matrix;

a second smoothing unit configured to smooth the second random perturbation matrix; and

a second perturbing unit configured to perturb, using the smoothed second random perturbation matrix, feature maps output by a convolution layer in training the neutral network,

wherein a matrix size of the second random perturbation matrix is a length of a feature map multiplied by a width of the feature map multiplied by a number of the feature maps.

18. The training apparatus according to claim 17, wherein the second perturbing unit is further configured to:

add a corresponding weighted value of the second random perturbation matrix to a value of the feature map, wherein a weighting coefficient is larger if an absolute value of the feature map is larger.

20. A training apparatus for a neutral network for image recognition, comprising:

a computer, comprising;

a high-dimensional space representing unit configured to represent a sample image as a point set in a high-dimensional space, a space size of the high-dimensional space being a domain size of a space domain of the sample image multiplied by an intensity size of an intensity domain of the sample image;

a first random perturbation matrix generating unit configured to generate a first random perturbation matrix having a same size as the high-dimensional space;

a first smoothing unit configured to smooth the first random perturbation matrix;

a first perturbing unit configured to perturb the point set in the high-dimensional space using the smoothed first random perturbation matrix to obtain a perturbed point set; and

a training sample determining unit configured to train the neutral network using the perturbed point set as a new sample.