Pattern recognition with hierarchical networks

• US 20030002731A1
• Filed: 05/24/2002
• Published: 01/02/2003
• Est. Priority Date: 05/28/2001
• Status: Active Grant

• Current Assignee: Honda Research Institute Europe GmbH (Honda Motor Company)
• Original Assignee: Honda Research Institute Europe GmbH (Honda Motor Company)
• Inventors: Korner, Edgar, Wersing, Heiko
• Granted Patent: US 7,308,134 B2
• Time in Patent Office: Days
• US Class Current: 382/161
• CPC Class Codes: G06V 10/454 Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]

Abstract

Within the frameworks of hierarchical neural feed-forward architectures for performing real-world 3D invariant object recognition a technique is proposed that shares components like weight-sharing (2), and pooling stages (3, 5) with earlier approaches, but focuses on new methods for determining optimal feature-detecting units in intermediate stages (4) of the hierarchical network. A new approach for training the hierarchical network is proposed which uses statistical means for (incrementally) learning new feature detection stages and significantly reduces the training effort for complex pattern recognition tasks, compared to the prior art. The incremental learning is based on detecting increasingly statistically independent features in higher stages of the processing hierarchy. Since this learning is unsupervised, no teacher signal is necessary and the recognition architecture can be pre-structured for a certain recognition scenario. Only a final classification step must be trained with supervised learning, which reduces significantly the effort for the adaptation to a recognition task.

Due to the improved learning efficiency, not only two dimensionally objects, but also three dimensional objects with variations of three dimensional rotation, size and lightning conditions can be recognized. As another advantage this learning method is viable for arbitrary nonlinearities between stages in the hierarchical convolutional networks, like e.g. non-differentiable Winner-Take-All nonlinearities. In contrast thereto the technology according to the abovementioned prior art can only perform backpropagation learning for differentiable nonlinearities which poses certain restrictions on the network design.

Claims

1. A method for recognizing a pattern having a set of features, the method comprising the following steps:

convolving a plurality of fixed feature detectors with a local window scanned over a representation of the pattern to generate a plurality of feature maps;

applying a nonlinearity function to each feature map separately;

sensing local combinations of at least one simple feature of the feature maps; and

recognizing the pattern by classifying it on the basis of the sensed local combinations, wherein essentially statistically independent features are pre-set for the local combination of features.

2. The method of claim 1, wherein the statistically independent features are pre-determined by means of an independent component analysis of convolutions of training patterns.

3. The method of claim 1, wherein the statistically independent features are pre-determined by means of a principal component analysis of convolutions of training patterns.

4. The method of claim 1, wherein a winner-takes-all strategy is applied on the result of the convolution to generate the feature maps.

5. The method of claim 1, wherein said nonlinearity function applied to each feature map is an arbitrary, a non-differentiable nonlinearity function.

6. The method of claim 1, farther comprising the step of:

pooling the feature maps wherein the feature maps of a preceding step are locally averaged and subsampled.

7. The method of claim 1, wherein the recognizing step uses a 1-layered sigmoidal function trained with a gradient descent technique to classify the pattern.

8. The method of claim 1, wherein classifying the pattern in the recognizing step uses one of a radial basis function network, a nearest neighbour matching algorithm, and a multi-layer-perceptron network.

9. The method of claim 1, wherein said convolving step and said generating step are repeated.

12. A computer software program, capable of operating on a computing device that implements a method according to claim 1 when executed on the computing device.

10. A method for recognizing a pattern having a set of features, the method comprising the following steps:

convolving a plurality of fixed feature detectors with a local window scanned over a representation of the pattern to generate a plurality of feature maps;

applying a nonlinearity function to each feature map separately;

sensing local combinations of the simple features of the feature maps; and

recognizing the pattern by classifying it on the basis of the sensed local combinations;

wherein a winner-takes-all strategy is applied on the result of the convolution to generate the feature maps.

13. A computer software program, capable of operating on a computing device that implements a method according to claim 10 when executed on the computing device.

11. A system for training a hierarchical network, the hierarchical network comprising:

means for convolving a plurality of fixed feature detectors with a local window scanned a representation of the pattern to generate a plurality of feature maps;

means for applying a nonlinearity function to each feature map separately;

intermediate means for sensing local combinations of simple features of the feature maps; and

means for recognizing the pattern by classifying it on the basis of the sensed local combinations;

wherein said intermediate means are incrementally trained to enhance the statistical independence of the local combinations of features.

15. The apparatus of claim 14 wherein the statistically independent features are pre-set by means of an independent component analysis of training patterns.

16. The apparatus of claim 14, wherein the statistically independent features are pre-set by a principal component analysis of training patterns.

17. The apparatus of claim 14, wherein the means for convolving generates the feature maps applies a winner-takes-all strategy to the result of the convolution.

18. The apparatus of claim 14, the means for applying a nonlinearity function applies an arbitrary, non-differentiable nonlinearity function to each feature map.

19. The apparatus of claim 14, further comprising:

pooling means for locally averaging and subsampling feature maps generated by the convolution means.

20. The apparatus of claim 14, wherein the recognizing means are designed to use a sigmoidal function trained with a gradient descent technique to classify the pattern.

21. The apparatus of claim 14, wherein the recognizing means are designed to use one of a radial basis function network, a nearest neighbor matching algorithm, and a multi-layer-perceptron network to classify the pattern.

22. The apparatus of claim 14, wherein the hierarchical network is implemented by means of a parallel computation network.

23. The apparatus of claim 14, for use in optically recognizing at least one of characters and objects in digital representations.

24. The apparatus of claim 14, for use in optically recognizing at least one of handwritten digits and characters.

26. The apparatus of claim 25, wherein the recognizing are tuned to a particular whole view of the pattern.

27. The apparatus of claim 25, further comprising:

pooling means for locally averaging and subsampling feature maps generated by the convolution means.

28. The apparatus of claim 25, wherein the recognizing means are designed to use a sigmoidal function trained with a gradient descent technique to classify the pattern.

29. The apparatus of claim 25, wherein the recognizing means are designed to use one of a radial basis function network, a nearest neighbour matching algorithm, and a multi-layer-perceptron network to classify the pattern.

30. The apparatus of claim 25, wherein the hierarchical network is implemented by means of a parallel computation network.

31. The apparatus of claim 25, for use in optically recognizing at least one of characters and objects in digital representations.

32. The apparatus of claim 25, for use in optically recognizing at least one of handwritten digits and characters.

14. A pattern recognition apparatus with a hierarchical network, the hierarchical network comprising:

means for inputting a representation of a pattern;

means for convolving a plurality of fixed feature detectors with a local window scanned over the representation of the pattern to generate a plurality of feature maps;

means for applying a nonlinearity function to each feature map separately;

intermediate means for sensing local combinations of simple features of the feature maps; and

means for recognizing the pattern by classifying it on the basis of the sensed local combinations;

wherein said intermediate means utilize pre-set essentially statistically independent features.

25. A pattern recognition apparatus with a hierarchical network, the hierarchical network comprising:

means for inputting a representation of a pattern;

means for convolving a plurality of fixed feature detectors with a local window scanned over the representation of the pattern to generate a plurality of feature maps;

means for applying a nonlinearity function to each feature map separately;

intermediate means for sensing local combinations of simple features of the feature maps; and

means for recognizing the pattern by classifying it on the basis of the sensed local combinations;

wherein the convolving means use a use of a winner-takes-all strategy to generate the feature maps.