CONVOLUTIONAL-NEURAL-NETWORK-BASED CLASSIFIER AND CLASSIFYING METHOD AND TRAINING METHODS FOR THE SAME

• US 20150036920A1
• Filed: 07/31/2014
• Published: 02/05/2015
• Est. Priority Date: 07/31/2013
• Status: Active Grant

• Current Assignee: Fujitsu Limited
• Original Assignee: Fujitsu Limited
• Inventors: Fan, Wei, He, Yuan, Sun, Jun, WU, Chunpeng
• Granted Patent: US 9,697,444 B2
• Time in Patent Office: Days
• US Class Current: 382/156
• CPC Class Codes: G06F 18/24 Classification techniques
• CPC Class Codes: G06F 18/24137 Distances to cluster centroïds
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06N 3/048 Activation functions
• CPC Class Codes: G06V 10/454 Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
• CPC Class Codes: G06V 10/764 using classification, e.g. of video objects

Abstract

The present invention relates to a convolutional-neural-network-based classifier, a classifying method by using a convolutional-neural-network-based classifier and a method for training the convolutional-neural-network-based classifier. The convolutional-neural-network-based classifier comprises: a plurality of feature map layers, at least one feature map in at least one of the plurality of feature map layers being divided into a plurality of regions; and a plurality of convolutional templates corresponding to the plurality of regions respectively, each of the convolutional templates being used for obtaining a response value of a neuron in the corresponding region.

Claims

1. A convolutional-neural-network-based classifier, comprising:

a plurality of feature map layers, at least one feature map in at least one of feature maps being divided into a plurality of regions; and

a plurality of convolutional templates corresponding to the plurality of regions respectively, each of the convolutional templates being used for obtaining a response value of a neuron in the corresponding region.

2. The convolutional-neural-network-based classifier of claim 1, wherein the convolutional-neural-network-based classifier is configured for classifying an image, and the image is divided into a plurality of image regions, which correspond to the plurality of regions respectively, according to a fixed scheme.

3. The convolutional-neural-network-based classifier of claim 1, wherein the convolutional-neural-network-based classifier is configured for classifying an image, and the convolutional-neural-network-based classifier further comprises a image dividing unit which is configured for dividing the image into a plurality of image regions corresponding to the plurality of regions respectively.

4. The convolutional-neural-network-based classifier of claim 3, wherein the plurality of regions is divided based on a pixel distribution of the image.

5. The convolutional-neural-network-based classifier of claim 4, wherein the plurality of regions is divided based on a quantity distribution or a density distribution of foreground pixels in a binary image of the image.

6. The convolutional-neural-network-based classifier of claim 3, wherein the plurality of regions is divided based on a structure of the image.

7. The convolutional-neural-network-based classifier of claim 1, wherein for different feature maps in the same feature map layer, or for feature maps in the different feature map layers, the different dividing schemes are used for dividing the plurality of regions.

8. The convolutional-neural-network-based classifier of claim 1, wherein an activation function used for calculating the response value is configured for obtaining a positive response value.

9. The convolutional-neural-network-based classifier of claim 8, wherein the activation function is configured to take an absolute value for a negative convolutional result or to convert a negative convolutional result to a positive small value.

10. A classifying method by using a convolutional-neural-network-based classifier which comprises a plurality of feature map layers, comprising:

dividing at least one feature map in at least one of the plurality of the feature map layers into a plurality of regions;

performing forward propagation by inputting an object to be classified into the convolutional-neural-network-based classifier to obtain an output result, during the forward propagation, each of the convolutional templates, which correspond to the plurality of regions respectively, is used for obtaining a response value of a neuron in the corresponding region; and

classifying the object to be classified according to the output result.

11. The method of claim 10, wherein the object to be classified is an image, the step of dividing at least one feature map in at least one of the plurality of the feature map layers into a plurality of regions comprising:

dividing the image into a plurality of image regions, which correspond to the plurality of regions respectively.

12. The method of claim 11, wherein the object to be classified is an image, the step of dividing at least one feature map in the at least one of the plurality of the feature map layers into a plurality of regions comprising:

dividing the image into a plurality of image regions, which correspond to the plurality of regions respectively, according to a fixed scheme.

13. The method of claim 11, wherein the step of dividing at least one feature map in at least one of the plurality of the feature map layers into a plurality of regions comprises:

dividing the at least one feature map into the plurality of regions based on a pixel distribution of the image.

14. The method of claim 11, wherein the step of dividing at least one feature map in at least one of the plurality of the feature map layers into a plurality of regions comprises:

dividing the at least one feature map into the plurality of regions based on the structure of the image.

15. The method of claim 10, wherein the step of dividing at least one feature map in at least one of the plurality of the feature map layers into a plurality of regions comprises:

for different feature maps in the same layer of feature maps, or for feature maps in the different layers of feature maps, using different dividing schemes to dividing the plurality of regions.

16. The method of claim 10, wherein the step that each of the convolutional templates, which correspond to the plurality of regions respectively, is used for obtaining a response value of a neuron in the corresponding region comprises:

calculating the response value using an activation value configured to acquire a positive response value.

17. A method for training the convolutional-neural-network-based classifier, which comprises a plurality of feature map layers, comprising:

dividing at least one feature map in at least one of the plurality of the feature map layers into a plurality of regions;

performing forward propagation by inputting a training sample with a known flag into the convolutional-neural-network-based classifier to obtain an output result, during the forward propagation, each of the convolutional templates, which correspond to the plurality of regions respectively, is used for obtaining a response value of a neuron in the corresponding region;

performing back propagation according to the difference between the output result and the known flag to correct parameters of the convolutional-neural-network-based classifier comprising the weights in the convolutional template; and

repeating the above steps until a predetermined condition is met.

18. The method according to claim 17, wherein the training sample is an image, the step of dividing at least one feature map in at least one of the plurality of the feature map layers into a plurality of regions comprising:

dividing the image into a plurality of images regions, which correspond to the plurality of regions respectively, according to a fixed scheme respectively, wherein the fixed scheme is determined based on general features of the training sample.

19. The method according to claim 17, wherein the training sample is an image, the step of dividing at least one feature map in at least one of the plurality of the feature map layers into a plurality of regions comprising:

dividing each image into a plurality of image regions, which correspond to the plurality of regions respectively, based on the personal features of each the image respectively.

20. The method according to claim 19, wherein the step of dividing at least one feature map in at least one of the plurality of the feature map layers into a plurality of regions comprises:

dividing the at least one feature map into the plurality of regions based on a pixel distribution of the image.