JPH05101028A - Integral decision method for plural feature quantity - Google Patents

Abstract

(57) [Abstract] [Purpose] To optimize or semi-optimize the integrated judgment of a plurality of feature quantities for one recognition target. [Structure] Two feature vectors A and B extracted from a certain recognition target are respectively input to neural networks 11 and 12.
To the integrating neural network 13, and the class corresponding to the neuron having the maximum value in the output is used as the recognition result. Like the other neural networks, the integration neural network 13 also obtains a high recognition rate because the connection between the elements of the outputs of the neural networks 11 and 12 is optimized by back propagation learning.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an integrated determination method for inputting a plurality of feature quantities extracted from a certain recognition target to different neural networks and integrating the outputs thereof to determine the recognition target.

[0002]

2. Description of the Related Art A conventional technique relating to an integrated determination method in which a plurality of feature quantities extracted from a certain recognition target are individually input to a plurality of neural networks and the output values of the plurality of neural networks are integrated to determine the recognition target. Will be explained. Here, a case where two feature quantities are input to two separate neural networks will be described as an example.

Further, as the neural network, a multilayer structure type neural network including an input layer, an intermediate layer and an output layer is adopted, and this is used for back propagation.
Algorithm [Reference Rumelhart, D. et al.
E. et al. : Parallel Distrib
eded Processing, Vol. 1, MIT
Press (1986)] will be described as an example.

FIG. 4A shows an example of the structure of a multilayer structure type neural network, where 1 is a neuron in the input layer, 2 is a neuron in the intermediate layer, and 3 is a neuron in the output layer. The neurons are connected by a connecting line 4, and the neurons in the middle layer and the output layer are connected by a connecting line 5. FIG. 4B shows a configuration example of the integrated determination method adopted in the conventional technique. The first feature amount is input to the first neural network 6 and the second feature amount is input to the second neural network 7. Then, based on the output values of the two neural networks 6 and 7, the recognition target is integratedly determined by the integration determination unit 8.

Conventionally, when the output values of a plurality of neural networks are integrated to determine a recognition target, the following two methods have been adopted. The first is a method in which the output values of a plurality of neural networks are added or multiplied, and the class corresponding to the neuron with the maximum value is used as the recognition result.

The second is a method of calculating a basic probability based on the output value of each neural network, and using this basic probability, a recognition result is obtained by a Dempster coupling algorithm [Reference: Shafer, G. et al. :
A Mechanical Theory of
Evidence, Princeton Univer
city Press (1976)].

The above two methods will be specifically described below by taking handwritten numeral recognition as an example. Now, two feature quantities (particularly 32 dimensions) of horizontal grayscale frequency distribution and vertical grayscale frequency distribution (distribution of the number of black pixels when the handwritten digit is viewed in the horizontal direction or the vertical direction) are calculated from the handwritten numbers. These two feature quantities are extracted and input to two separate neural networks.

The scales of the neural networks are 32-10-10 (32 neurons in the input layer,
10 neurons in the middle layer, 1 neurons in the output layer
0). Next, after each neural network is independently trained by the backpropagation algorithm, the output values of the two neural networks are used to make an integrated decision.

For example, as the teacher signal, 0.90 is given to the neurons corresponding to the correct answer class, and 0.10 is given to the other neurons to perform learning. After learning with 750 patterns, an unknown test was performed with other 750 patterns. As a result, the recognition rate was 83.20% in the lateral distribution input and 51.33% in the vertical distribution input. As a first method for making an integrated determination using the outputs of two neural networks, a case will be described in which a class corresponding to a neuron having a maximum value obtained by adding the output values of the two is used as a recognition result.

The result corresponding to an unknown handwritten numeral "0" is shown in FIG. Looking at the output value of the neural network using the lateral distribution as an input, the output value of the neuron corresponding to “8” was the maximum, and the output value of the neuron corresponding to the correct answer “0” was the second highest. On the other hand, when looking at the output value of the neural network using the vertical distribution as the input, the output value of the neuron corresponding to the correct answer "0" was the maximum. As a result of adding both output values, the value corresponding to the correct answer “0” was the maximum, and the correct recognition result was obtained. The recognition rate for all unknown data improved to 86.00%.

Next, as a second method for making an integrated decision, in the case where the basic probabilities are calculated from the output values of the neural network and the two obtained basic probabilities are integrated by the Dempster's joint algorithm to obtain the recognition result. State. The result corresponding to a certain unknown handwritten numeral "0" is shown in FIG. The basic probability calculated based on the output value of the neural network with the lateral distribution as an input is 0.365, which is the maximum value corresponding to “8”, and the value of the neuron corresponding to “0” of the correct answer is The second largest was 0.305. On the other hand, in the basic probability calculated based on the output value of the neural network with the vertical distribution as an input, the value corresponding to the correct answer “0” was 0.447 at the maximum. The probability value obtained by the Dempster's combining algorithm using both basic probabilities has the maximum value corresponding to the correct answer "0", and the correct recognition result was obtained. The recognition rate of all unknown data improved to 86.53%.

[0012]

In the integrated determination method in the prior art, when integrating the output values of a plurality of neural networks in which a plurality of feature quantities are used as individual direct inputs,
It is common to treat each output value equally. In other words, for example, in the addition method, it is difficult to obtain an engineering guideline that regards which feature value is important and which weight should be added with a large weight. Therefore, multiple output values are added with equivalent weights. The method has been adopted.

In the case of feature quantities that are orthogonal to each other in the feature space, such as the horizontal distribution and the vertical distribution described in the section of the prior art, the two feature quantities are considered to be almost equally effective. Be done. However, when a plurality of different types of feature amounts generated by one recognition target (for example, voice and image) are determined as inputs, it is not necessarily said that all feature amounts contribute equally in the recognition process. Not limited. Rather, they are considered to be classified into a particularly important feature amount and an auxiliary necessary feature amount according to the recognition target or the nature of the recognition problem.

In order to realize such weighting between feature quantities, in the conventional technique, a human being decides the weighting parameter by trial and error. In this case, there is no guarantee that the obtained parameter value is optimum. As described above, in the conventional integrated determination method, when a plurality of feature quantities are individually input to a plurality of neural networks and a plurality of output values are combined to perform integrated pattern recognition, which feature Since it is difficult to obtain an engineering guideline for whether weights should be weighted more importantly, it is common to treat them equally. In such a conventional technique, there is a problem that sufficient performance cannot be realized by a pattern recognition problem in which determination is performed in an integrated manner based on a plurality of different types of feature values.

An object of the present invention is to combine a plurality of different types of feature amounts generated by a certain recognition target and perform pattern recognition in an integrated manner, by weighting the respective feature amounts by an engineering method. It is to establish an integrated judgment method that can be determined optimally or suboptimally.

[0016]

In order to achieve the above-mentioned object, the integrated determination method of the present invention provides a neural network for integration in addition to a neural network group for directly inputting a plurality of feature quantities extracted from a certain recognition target. A network is provided, the output of the neural network group is used as an input to the integrating neural network, and the recognition target is determined by the output value of the integrating neural network.

[0017]

In the integrated determination method of the present invention, an integrating neural network is provided in addition to the neural network group for directly inputting a plurality of feature quantities extracted from the recognition object, and the output of the neural network group is integrated into the integrated network. As an input to the neural network for learning, the judgment mechanism is learned by the backpropagation algorithm. By this learning, in the integrated decision method of the present invention, the neural network and the given learning set can be used to realize an integrated decision unit having an optimal or suboptimal function with respect to the integrated decision, and further,
It is possible to realize an integrated determination method having a determination capability superior to that of the conventional integrated determination method.

[0018]

DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described by taking as an example a case where two feature quantities (feature vector A and feature vector B) extracted from a certain recognition target are input to two separate neural networks. The number of classes to be classified is 10. Further, the feature vector A that is the first feature amount is 32 dimensions, and the feature vector B that is the second feature amount is 16 dimensions.

FIG. 1 shows an example of the configuration of a recognition device to which the integrated determination method of the present invention is applied. The feature vector A that is the first feature amount is input to the first neural network 11, and the feature vector B that is the second feature amount is input to the second neural network 12, and the first and second neural networks are input. The outputs of 11 and 12 are input to the integrating neural network 13. The scale of the neural network is 32-10-10 for the neural network 11 that inputs the feature vector A, and the feature vector B is
The neural network 12 for inputting is 16-10-
Set to 10. The integration neural network 13 is 20-10-10.

FIG. 2 shows an embodiment of a learning method of the integrating neural network 13 in the integrated judging method of the present invention. That is, first, the first feature amount (feature vector A) extracted from the learning data is input to the first neural network 11, and the classification mechanism is learned by the back propagation algorithm (step 110). As a teacher signal, 0.90 is given to the neurons corresponding to the correct class, and 0.10 is given to the other neurons.

Secondly, the second feature quantity (feature vector B) extracted from the learning data is input to the second neural network 12, and the backpropagation algorithm is used to learn the classification mechanism (step 120). .. As a teacher signal, 0.90 is assigned to neurons corresponding to the correct answer class, and 0.1 is assigned to other neurons.
Give 0.

Thirdly, each output of the first neural network 11 and the second neural network 12 after learning is used as an input to the integrating neural network, and a backpropagation algorithm is used to provide an integrated decision mechanism. Learning is performed (step 130). As a teacher signal, 0.90 is given to the neurons corresponding to the correct class, and 0.10 is given to the other neurons.

The above learning operation (steps 110 to 13)
0), the connection between the neurons of the outputs of the neural networks 11 and 12 is the integration neural network 1
It is clear that the integrated decision unit having the optimum or sub-optimal function for integrated decision, which is optimized by the elements in 3, can be realized engineeringly by learning the neural network.

FIG. 3 shows an embodiment of the integrated judgment method of the present invention. First, the first feature quantity (feature vector A) extracted from the input data is used as the first neural network 1
1 is input (step 210). Secondly, the second feature amount (feature vector B) extracted from the input data is input to the second neural network 12 (step 22).
0).

Third, the first neural network 1
The output of 1 and the output of the second neural network 12 are input to the integrating neural network 13 (step 230). Fourthly, the class corresponding to the neuron having the maximum value in the output of the integrating neural network 13 is set as the recognition result (step 240).

The above operation (steps 210 to 240)
Then, it is obvious that the integrated decision processing using the outputs of the plurality of neural networks when the plurality of feature quantities are input can be optimally or sub-optimally realized. In the above description, as a teacher signal to the integrating neural network 13, 0.9 is assigned to the neuron corresponding to the correct class.
Although the example has been described in which the assertive value is given to 0 and other neurons such as 0.10. However, when the maximum and minimum real values are determined within a predetermined range, the maximum value is given to the correct class neuron. It is clear that the real value of can be applied to the case where the other neurons are given a value that is greater than or equal to the minimum value and less than the maximum value (may be different for each neuron). In other words, it may be better to give the neuron other than the correct class an intermediate value between the maximum value and the minimum value.

[0027]

As described above, according to the present invention, in the integrated determination method using a plurality of neural networks, in addition to a neural network group which directly inputs a plurality of feature quantities extracted from a certain recognition target, A neural network for use is provided, the output of the neural network group is used as an input to the integrating neural network, and the judgment mechanism is learned by the backpropagation algorithm. Therefore, according to the integrated determination method of the present invention, an integrated determination unit having an optimum or sub-optimal function for integrated determination by learning can be realized by engineering using a neural network and a given learning set. It is possible to realize an integrated determination method having a determination capability superior to that of

[Brief description of drawings]

FIG. 1 is a block diagram showing a recognition device to which the method of the present invention is applied.

FIG. 2 is a flowchart showing an embodiment of a learning method for the neural network shown in FIG.

FIG. 3 is a flowchart showing an embodiment of the integrated determination method of FIG.

FIG. 4A is a diagram showing a multilayer structure type neural network, and FIG. 4B is a block diagram showing a conventional integrated determination method of a plurality of feature amounts.

FIG. 5 is a diagram showing a specific example of conventional integration determination processing.

FIG. 6 is a diagram showing another specific example of conventional integration determination processing.

Claims (1)

[Claims] 1. An integrated determination method, wherein a plurality of feature quantities extracted from a certain recognition target are individually input to a plurality of neural networks, and output values of the plurality of neural networks are integrated to determine a recognition target. An integrated determination method for a plurality of features, wherein outputs of a plurality of neural networks are input to an integration neural network, and a recognition target is determined based on output values of the integration neural network.