JPH02242392A - Pattern recognition processing system - Google Patents

Abstract

PURPOSE:To reduce the learning cost by using specifically learning neural cells as the input stage and using neural cells, which learn the combination of first outputs of a prescribed number of cells in the preceding stage as the standard signal, as the intermediate stage and judging the truth at the time when second outputs are larger than a prescribed value. CONSTITUTION:Neural cells 3 are allowed to learn the signal indicating the feature quantity of each divided area of a picture pattern as the standard signal, and cells 3 are so connected that signals indicating feature quantities are inputted, thereby constituting an input stage 1. Neural cells 4 are allowed to learn the required combination of first outputs of a prescribed number of cells 3 in the hierarchy of the preceding stage as the standard signal, and cells 4 are so connected that first outputs are inputted, thereby constituting an intermediate stage 2a.... The logical value is judged to be the truth when second outputs of cells 3 and 4 are larger than the prescribed value, and the picture pattern is discriminated when AND of second outputs of all cells 3 and 4 are the truth in AND circuits 5a, 5b.

Description

[Detailed Description of the Invention] [Summary] Regarding a pattern recognition system configured with a neural network, the present invention aims at a pattern recognition processing method that can obtain a high recognition rate with a relatively small learning cost and an economical configuration. Each standard signal corresponding to one or more points set in a required spatial region of the dimensional input signal space is used as an input signal, the first output is a value determined for each standard signal, and the value of the second output is 1.0゛. Learning was performed using each teacher signal, and learning was performed using a teacher signal in which the second output value was O, with each input signal being a signal corresponding to two or more points located at a predetermined distance from each standard signal. The input stage and the required number of intermediate stages consisting of neural net cells are connected in a hierarchical manner so that the final stage has one cell, and is provided for each category of the required image pattern, and the image pattern is divided. The cell, which has been trained with the signal indicating the required feature quantity for each area as the standard signal, is connected to input the signal indicating the special quantity as the input stage, and the predetermined number of cells in the previous layer are The cell which has learned the required combination of the first outputs of the cell as the standard signal is connected to the intermediate stage so that the first output is input, and the second output is lower than the predetermined darkness value. If it is larger, the logical value is considered to be true, and if the AND of the second outputs of all the cells is true, it is configured to be identified as an image pattern of the category.

[Industrial Application Field] The present invention relates to a pattern recognition processing method for performing pattern recognition using a neural network having a built-in learning function that allows a required output signal specified by a teacher signal to be obtained from an input signal.

[Conventional technology]

As is well known, a neural network consists of, for example, an input layer, a required number of intermediate layers, and an output layer, as shown in FIG. It is configured so that input is made to all the swords of the layer, and a predetermined function of the sum of these is input to the current node. Generally, the weights are different for each pair of nodes.

When a neural network is used to recognize a pattern such as an image, a signal of a feature amount extracted from a character, for example, is input to an input terminal connected to the input layer of the neural network 2, and the recognition result is obtained at an output terminal.

To this end, each input signal is, for example, a real value in the range of 0 to 1 representing a predetermined feature detected by scanning a character, and this value is input to each node of the input layer of the neural network.
For example, when the pattern "A" is recognized, the output is 1.0, and in other cases, 0 is output.

As the feature quantity input in the case of recognizing an image pattern such as a character, for example, a quantity related to the inclination of a line forming a character pattern is used, as is known in the art. That is, as illustrated in FIG. 6(a), for example, a slit 42 of an appropriate length and width is formed.
A detector that scans the character 45 by moving the a, etc., and outputs the ratio of the area occupied by the black line of the character that falls within the slit frame as a value such as O1 when all are mills and 1.0 when all are black. will be established.

For example, set the slope to 0° (vertical line), -4
Slits 42a with angles of 5° (slanted line upward to the right), +45° (slanted line upward to the left), and 90" (horizontal line)
By taking the maximum value of each output detected by 42d, we can calculate the amount for the line that has the largest amount of both the proximity of the slope and the length extending in that direction among the lines with slopes close to those lines. The four-element character feature amount is obtained.

If more detailed identification is required, for example, Fig. 6(b)
Divide the area of the pattern into four parts as shown in the figure below, and find the feature amount as described above for each area. In the case of the above example, 16
It can be made into a feature amount of an element, or further divided into four to make a feature amount of, for example, 64 elements.

As is well known, learning is performed by inputting an input signal such as a feature value and a teacher signal indicating a desired output signal from the input into a neural network, and then calculating the weights between each node using a predetermined algorithm. , by repeating the operation of changing based on the difference between the actual output signal and the teacher signal, the difference between the two signals gradually decreases, and the difference converges to within a predetermined value. You can learn about one of the patterns.

However, when you want to recognize a certain character, you need to learn to obtain recognition output by sequentially inputting the features of each of the standard patterns as well as a number of patterns that are distorted to the required range from that pattern. Therefore, in order to increase the recognition rate, it is necessary to collect and learn the feature amounts for various distortion patterns of the categories to be recognized. A learning processing method capable of reducing this human labor is described in a patent application entitled "Neural Network Learning Processing Method" filed by the applicant on February 28, 1989.

According to this method, learning is performed in which the output is set to 1.0 for each standard signal input determined at appropriate intervals in a required subspace on the input signal space, and several sets of inputs are provided around each standard signal. When learning is performed with the output set to 0 for a signal, due to the learning characteristics of the neural network that can interpolate and extrapolate function values, the output in the standard signal for all inputs in the small island-shaped space surrounded by the neural network. 1.0 to 0 around
An output value that gradually decreases to is interpolated, and an output value outside the island is extrapolated to O. Therefore, by filling the subspace with such islands, there is no need to fill the subspace with points,
The amount of data and learning processing required for learning can be reduced.

[Problem to be solved by the invention] However, when the feature values are detailed as described above and the number of input feature value elements is increased, the number of nodes in each layer of the neural network increases, and accordingly, the number of nodes in each node increases. As scale increases, it is generally necessary to increase the number of middle classes;
At the same time, the convergence time of learning processing increases rapidly, and the number of cases in which learning does not converge due to initial conditions and other factors also increases.

Therefore, for example, it is possible to divide the feature values into a small number of sets of elements for each subregion of the pattern, simply recognize each set using an independent neural network, and use the logical product of the results as the overall result. If this method is adopted, there is a problem in that the relationship between partial areas is ignored, and erroneous recognition is likely to occur.

An object of the present invention is to provide a pattern recognition processing method that can obtain a high recognition rate with a relatively small learning cost and an economical configuration.

[Means to solve the problem]

FIG. 1 is a block diagram showing the configuration of the present invention.

The figure shows the configuration of a pattern recognition processing method, in which each standard signal corresponding to one or more points set in a required spatial region of a multidimensional input signal space is used as an input signal, and the first output is a value determined for each standard signal. , learning using each teacher signal with the value of the second output as 1.0, and learning with the value of the second output as 0 with each input signal being a signal corresponding to each point of two or more located at a predetermined distance from each standard signal. An input stage 1 consisting of neural network cells (hereinafter referred to as neural cells) and a required number of intermediate stages 2a and 2b that perform learning using a teacher signal as follows.
, -. are connected in a hierarchical manner so that the cell in the final stage becomes one, and is provided for each category of the required image pattern, and a signal indicating the required feature amount for each region into which the image pattern is divided. Neural cell 3 that has been trained to use as the standard signal
is connected to input a signal indicating the feature amount as input stage 1, and a neural cell is trained to use a required combination of first outputs of a predetermined number of cells in the previous layer as the standard signal. 4 are connected to input the first output to form intermediate stages 2a, 2b, -.
,- for all neural cells 3.4.
The configuration is such that when the logical AND of the outputs is true, it is identified as an image pattern of the category.

[For production]

With this processing method, the individual neural cells 3 and 4 only need to be relatively small-scale neural networks, making learning processing easy, and neural cells with the same configuration can be connected to support any number of feature elements. It is highly economical because it allows you to configure a system that

〔Example〕

FIG. 2 is a block diagram showing an embodiment of the present invention. In the case of this example, the image pattern is divided into four regions, and for each region, four elemental feature quantities related to lines in the four directions as described above are collected and input to each neural cell 10 of the input stage. Each neural cell 10 has two outputs ID and CF, and the 10 outputs of the four neural cells 10 are connected to the input of the neural cell 11 in the next stage.

Neural cell 11 has the same configuration as neural cell 10, and has output ID and CF, but uses only CF ratio output,
This is inputted to the AND circuit 12 together with the four CF ratio outputs of the input stage, and the AND circuit 12 sets the logic value to 1 if the input signal value is greater than a predetermined value (for example, 0.8), and if it is less than that, the logic The circuit is configured to output the logical product of 5 human forces as the value O, and this output is set to 1 when it is recognized as a predetermined category of the image pattern (for example, a specific character). A circuit block 13 is provided for each necessary category (for example, each letter of the alphabet).

For this purpose, each neural cell 10 has a 3rd output of 4 human power and 2 outputs.
It has a conceptual configuration as shown in FIG. 3(a), and is configured to execute the calculation shown in FIG. 3(b).

In other words, let the power of four people be X(0) to X(3), and
(4) With =1 as input, a set of 5 weights each set as 11° to 1A11-1 is set for each node pair of the input layer and the intermediate layer, and the weight of each set and the input A predetermined function calculation is performed on the result of adding the five products of
0) ~Z(I-1) and Z(I)=1 as input, 2
The output ID and cp are obtained by the same calculation as above using the weights W2' and W2' of m. Here, the number of nodes in the middle layer is experimentally determined, and the weight u1° ~ 411-1
, -2°, and -21 are separately learned as described later, and the required values are set.

The learning of each neural cell 10 in the input stage is based on four element features for each area in charge, which are collected as described above for image patterns such as a standard glyph pattern and some typical distorted glyph patterns. Learning is performed by giving each teacher signal in which the quantity signal is a standard signal, each standard signal is an input signal, the ID output is a value determined for each standard signal, the CF ratio output is 1.0, and the ID output is a predetermined value, Next, each standard signal is learned to be increased or decreased by a small fixed value (for example, 0.05) as a peripheral signal, and the peripheral signal is input to obtain a CF ratio output O. The value of the 10 outputs is set to be a different value for each standard signal so that each standard signal can be identified in the range of 0 to 1.0.

As a result of the above learning, due to the learning characteristics of the neural network, C
From the standard signal indicated by the black dot 18 to the peripheral signal indicated by the circumference 19, the output gradually decreases from 1.0 to O. Therefore, if the value above a certain dark value is taken as a logical value 1, then A state is reached in which it can be identified that the signal of the above-mentioned ball within the circumference 19 (signal corresponding to all the distorted character shapes within that range) has been input. Regarding ID output, specified values as shown in parentheses are output for input signals in each area divided so that each area contains one black dot of the standard signal as shown by the broken line in the figure. becomes.

Therefore, from the above, when the CF ratio output value is equal to or higher than the dark value, it can be determined from the IF ratio output value that the input is a signal of a certain specific distortion pattern.

Learning of the neural cell 11 is carried out by using the four neural cells 10 in the input stage to learn the IDs of standard signals in four areas of the same pattern.
The same learning as in the case of the neural cell 10 is performed using the four values designated as output values as the four elements of each standard signal.

Note that, like the neural cell 11 in this example, the ID output is not used in the case of the final stage, so learning is unnecessary, but especially when the number of elements of the input stage is large, Therefore, in neural cells other than the final stage when the intermediate stage has two or more layers, it is also necessary to learn about ID output, and the output value in that case can identify each standard signal in the same way as above. It is determined as follows.

By learning in this way, the neural cell 11 outputs an output corresponding to a CF ratio output logic value of 1 when the 10 outputs from the neural cell 10 in the input stage are the outputs for a specific pattern or close to it. .

That is, for example, if each neural cell 10 in the input stage is made to learn four islands as described above by learning based on the feature quantities of four patterns, in the case of recognition, each 221 Urcell 10-10 and C
There can be 4'=256 cases in which both the F outputs have a predetermined positive value, and the intermediate stage neural cell 11 functions to select four combinations corresponding to the pattern from among them.

As mentioned above, the 10 output is the specified value even if the input signal is not in the island, so it is necessary to confirm it with the CF output.For this purpose, the CF outputs of the previous neural cells 10 and 11 are combined by the AND circuit 12. and take the logical product,
The output is set to logical value 1 only when all CF outputs are 1,
Indicates that the pattern has been recognized.

〔Effect of the invention〕

As is clear from the above description, according to the present invention, in a pattern recognition system configured with a neural network,
This has a significant industrial effect in that it is possible to construct a system that can obtain a high recognition rate with a relatively small learning cost and an economical configuration.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the configuration of the present invention, FIG. 2 is a block diagram showing an embodiment of the present invention, FIG. 3 is a diagram showing an example of the configuration of a neural cell, and FIG. 4 is an explanatory diagram of learning results. Fig. 5 is an explanatory diagram of the neural network, and Fig. 6 is an explanatory diagram of feature quantities. In the figure, 1 is an input stage, 2a and 2b are intermediate stages, 3.4.
10. 11 is a neural cell, 5a, 5b, 12 is an AND circuit, and 13 is a circuit block 58z. Block diagram showing the configuration of the present invention. Diagram showing an example of the configuration of a neural cell Diagram explanatory diagram of a neural network (a) Φ) Diagram explanatory diagram of feature quantities Fig.

Claims (1)

[Claims] Using each standard signal corresponding to one or more points set in a required spatial region of a multidimensional input signal space as an input signal, the first output is a value determined for each standard signal, and the second output is a value determined for each standard signal. Learning using each teacher signal with a value of 1.0, and using a teacher signal with a second output value of 0 using each input signal as a signal corresponding to each point of 2 or more located at a predetermined distance from each standard signal. The input stage (1) consisting of cells (3, 4) of the neural network that underwent learning and the required number of intermediate stages (2a, 2b) are arranged in a hierarchical manner so that the final stage has one cell. The cell (3), which is connected and provided for each category of the required image pattern, and which has been trained as the standard signal with a signal indicating the required feature amount for each region into which the image pattern is divided, is used to indicate the feature amount. The input stage (1) is connected to input a signal, and the required combination of the first outputs of a predetermined number of cells (3, 4) in the previous layer is learned as the standard signal. A cell (4) is connected to input the first output to form the intermediate stage (2a, 2b), and when the second output is larger than a predetermined threshold value, the logical value is true, and all of the cells are connected. If the logical product of the second output is true, it is identified as an image pattern of the category (5a, 5b
) A pattern recognition processing method characterized by being configured as follows.