JPH02310782A - Fuzzy neuron - Google Patents

Abstract

PURPOSE:To perform pattern recognition corresponding to likelihood by performing the soft matching of a fuzzy reference signal decided with a membership function with a distribution signal fetched from voice or an image. CONSTITUTION:The recognition of the distribution signal is performed by includ ing a passage membership function means 40 which sets a passage membership function of single peak in a specific area, a prohibition membership function means 60 which sets a prohibition membership function, an inversion means 70 which finds the membership value of a position designated by the distribution signal (a) for each membership function and includes maximum value output means 44 and 64 to output the maximum value and also, inverts the maximum value of the membership obtained from the prohibition membership function, and a minimum value output means 20Z which find the minimum value of the whole from the output of the maximum value obtained from the passage membership function and that in which the maximum value of the prohibition membership function is inverted, and performing the collation of the distribution signal in the specific area with the membership function supplied in advance. Thereby, the optimum pattern processing can be performed on even ambiguous information.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a fuzzy neuron, particularly to a fuzzy neuron for effectively performing pattern recognition of audio signals, image signals, etc. using fuzzy inference.

[Prior art] Neural networks, which use networks imitating the human brain for electrical signal processing, are used for pattern recognition, signal processing, knowledge processing, etc., and are made up of multiple units that correspond to neurons in the brain and are connected in a complex manner. By appropriately determining the operation of each unit and the connection form between units, pattern recognition functions and knowledge processing functions can be performed.

These neural networks are classified into pattern transmission type and automatic transmission type based on the network structure.The former forms a network that converts a human cover turn into a certain output pattern, and the latter forms a network that outputs the pattern closest to the input cover turn. Form.

In the pattern transmission neural network, neurons in each unit are layered into an input layer, a middle layer, and an output layer.
Each unit is connected in the direction from the input layer to the output layer, and as a result, the human crab unit and the output crab unit are independent. The arithmetic functions and image information are then stored as the connection form and connection strength between each unit, and the connection strength is arbitrarily changed using the difference between the obtained output and the desired output. The strength of the connections at this time forms a predetermined parameter as the number of connections or the weight of the synapse.

On the other hand, in an automatic transmission neural network, the human and output patterns are common, all units in the network are interconnected, and the strength of the connections (synaptic weights) identifies similar incoming patterns. It is used for.

Pattern recognition is performed by such a neural network, but as is well known, in order to effectively operate this neural network, a learning algorithm known as pack propagation is used. Learning such as pack propagation is extremely effective for making neural networks work in a wide range of application fields such as recognition, speech synthesis, signal processing, and knowledge processing.

Normally, neural net networks have a hierarchical structure consisting of an input layer, a middle layer, and an output layer, and this middle layer is known as a hidden layer. Conversely, when learning data is input manually, the description of synaptic connections is changed (backward) to reduce the resulting error, and the learning algorithm is repeated until it converges.

Therefore, in the past, neural networks to which learning functions such as pack propagation algorithms have been added have been successful in various pattern recognition and other applications.

[Problem to be solved by the invention] However, such conventional neural networks always require a learning function in order to perform optimal recognition, and for this reason, when there are many parameters, the learning itself takes a lot of time. The problem was that the amount of money was wasted.

In addition, in conventional neural networks, it is usually "0" or "
It has been reported that various arithmetic operations are performed by digital signal processing using binary words of ``1'', which makes ambiguous recognition difficult, and that the field of its use is limited.

The present invention has been made in view of the above-mentioned conventional problems.
The purpose is to provide a fuzzy neuron that can perform optimal pattern processing even for ambiguous information that cannot be processed using binary words.

[Means for Solving the Problems] In order to achieve the above object, the present invention is characterized by constructing a fuzzy neuron in which fuzzy inference is introduced into a neural network that performs various pattern recognition.

Conventional strict matching between binary words (
Approximate inference, i.e., fuzzy inference, which obtains not exact but almost valid inference results, is currently being used in the control system of a certain forest. Until now, the concept of using such fuzzy inference for pattern recognition has not yet been realized.

In view of the current situation, the present invention proposes a fuzzy neuron that performs accurate pattern recognition using fuzzy inference. passing membership function means for setting at least one prohibited membership function within the specific area; and prohibited membership function means for setting at least one prohibited membership function within the specific area; maximum value output means for calculating the membership value and outputting the maximum value within the specific area; inverting means for inverting the maximum membership value obtained from the prohibited membership function; a minimum value output means for calculating the overall minimum value from the maximum value output of the prohibited membership function and the output obtained by inverting the maximum value of the prohibited membership function, It is characterized by recognition of distributed signals through matching.

That is, in the present invention, when recognizing a distributed signal in a specific region such as an audio signal or an image signal, a pass membership function and a prohibition membership function are calculated for this signal along a predetermined feature extraction line. Both are taken up as targets for inference, and the degree of certainty of recognition is output as a grade by comparing the distribution signal with both of these membership functions, and depending on the magnitude of this grade, the distribution signal is determined as the most applicable pattern. It becomes possible to recognize it.

Further, the fuzzy neuron according to the present invention includes a pass membership function means for setting at least one unimodal pass membership function within the specific region, and a pass membership function means for setting at least one unimodal pass membership function in the specific area, and a maximum value output means for calculating a membership value of a position and outputting the maximum value within the specific area; prohibited membership function means for setting one prohibited membership function within the specific area; and the prohibited membership function. an inverting means for inverting a function; a minimum value outputting means for determining a membership value at a position specified by the distribution signal for the prohibited membership function and outputting a minimum value within the specific area; total minimum value output means for calculating the overall total minimum value from the maximum value output obtained from the ship function and the minimum value output obtained from the prohibited membership function, The distribution signal is recognized by comparing it with the determined membership function.

That is, according to the present invention, the prohibited membership function indicates a region to which the target pattern should not originally apply, and even if this is inverted after comparison with the distribution signal, or the prohibited membership function itself is inverted. It is also possible to use it as a comparison signal.

[Operation] Therefore, according to the present invention, the distributed signal within a specific region is compared with at least one unimodal pass membership function to determine the maximum value of the distributed signal present within the pass membership function, On the other hand, the distribution signal is compared with the prohibited membership function to take the maximum value of the distributed signals present within the prohibited membership function.

Then, the maximum value of the distribution signal in the prohibited membership function is inverted, and the minimum value between this inverted value and the distribution signal in the pass membership function is determined, thereby forming a predetermined pattern given by the membership function. and the pattern of the distributed signal are matched by ambiguous fuzzy inference.

Therefore, according to the present invention, unlike the conventional hard matching using binary words, it is possible to obtain a pattern recognition solution by soft matching based on the degree of certainty. It becomes possible to express the degree as a degree of certainty.

Therefore, according to the pattern recognition of the present invention, handwritten characters,
It becomes possible to extract the most probable pattern from image information with other ambiguous expressions, audio information mixed with noise, etc., and it can be used in a wide range of application fields as a field of artificial intelligence.

Further, in the fuzzy neuron according to the present invention, both the passing membership function and the prohibited membership function are extracted along the feature cutting line within a specific range, and it is possible to determine whether the distribution signal passes through a predetermined passing area, and Pattern matching is performed to determine whether the distributed signal avoids a predetermined prohibited area, and when comparing the prohibited membership function and the distributed signal, this comparison result is inverted to prevent the distributed signal from entering the prohibited area. This is done by detecting or comparing the forbidden membership function itself with the distribution signal in an inverted state.

Regardless of the method, according to the present invention, a fuzzy reference signal defined by a membership function and a distributed signal captured from audio or images are soft matched, thereby making it possible to recognize patterns according to their certainty. do.

[Embodiments] Hereinafter, preferred embodiments of the present invention will be described based on the drawings.

In the following embodiments, distribution signals within a specific area will be explained by taking as an example the identification of handwritten characters, particularly handwritten numerals from "0" to "9".

Of course, the present invention is not limited to such handwritten digits, and can be widely used for other character recognition, graphic recognition, voice recognition, etc.

Overview of Pattern Recognition FIG. 2 shows an overall schematic configuration in which fuzzy neurons according to the present invention are used for pattern recognition of handwritten digits.

In the figure, the number "3" is displayed as a distributed signal within the specific area 100, and therefore, in the present invention, this number can be stylized as general image information, and can be used as an arbitrary character. This shows that the present invention can similarly handle symbols, figures, etc.

This distribution signal is captured by any image detection means, which will be explained in detail later, but for example, it is captured as a coordinate signal from an image input tablet, or numbers written on paper etc. are captured as an electrical signal by a scanner. be able to.

In FIG. 2, these detection means are indicated by the reference numeral 10, and output the distribution signal a as a serial signal at a predetermined scanning timing.

In the figure, a fuzzy neuron in which a pass membership function and a prohibited membership function determined by each feature extraction line, which will be described later, are set is provided individually for each number, and the distribution signal a is transmitted to each feature extraction line 101. ~10
7 and is supplied to all the fuzzy neurons 20-1 to 20-22 in parallel as a serial signal. In the figure, the feature cutting line is a specific area 10.
The line is set as a line that passes 0 horizontally and 3 times vertically, and a signal passing area E and a prohibited area (2), which will be described later, are determined along these characteristic cut-out lines.

In Figure 2, the fuzzy neurons range from "1" to "0".
This is because in order to correctly recognize each digit, it is necessary to divide the same digits into multiple groups due to variations in the characteristics of handwritten digits.

For example, for "3", two sets of fuzzy neuron groups are given, and the distribution signal is compared with both of these fuzzy neuron groups, and the one with a high degree of matching is selected and output as its grade, that is, its probability. Ru.

In the figure, reference numeral 22 indicates a MAX circuit, which outputs the maximum grade value for each group of "3".

As is clear from the figure, "2", "5", "6", "7"
and “9” are similarly matched with two grouped patterns, and the number “4” is matched with 3 patterns for each by separating the faces in order to more reliably distinguish it from other numbers. fuzzy neurons divided into groups 2
The maximum values of 0-6 to 20-8 and 20-9 to 20-12 are respectively outputted by the MAX circuit 24.26, and both of these maximum values are selected as the minimum value by the MIN circuit 28, and this is the number "
It is output as a grade of 4.

As described above, the 10 fuzzy neuron outputs for the handwritten digits in Figure 2 indicate the grade of certainty as described above, and in Figure 3, this grade μ ranges from "1" to rOJ as shown. is obtained as a set of singletons of unordered sequences.

Therefore, from this example of the Mengleton set, the probability of the distribution information can be pattern recognized as the number "3".

Of course, if a predetermined threshold value is given to these grades, the recognized numbers can be output as they are.

Feature Clipping Line and Membership Function According to the present invention, each fuzzy neuron is characterized by having a pass membership function and a prohibited membership function, and the procedure for setting these membership functions for each predetermined feature clipping line is described below. This will be explained based on FIGS. 4 to 7.

FIG. 4A shows the process of collecting a plurality of handwritten digits, for example, obtaining desired parameters from 30 handwritten digits,
Based on this, the member sub-function is determined.

The handwritten characters shown in FIG. 4A include variations in the writer's personality and size, and FIG. 4B shows a size normalization process in which the characters are standardized to predetermined dimensions.

Figure 4C shows a state in which all the handwritten characters collected in the above manner are overwritten in a single specified frame, and if this variation is small, the membership function can be found from this overwriting, In addition, as mentioned above, when the variation is large, grouping is performed, and for the number "3", it is divided into two types of groups as shown in the fourth diagram,
For each, fuzzy neurons 20-4 and 20-5 as shown in FIG. 1 are created.

FIG. 5 shows seven characteristic extraction lines 101 to 107 for a specific area 100 similar to those shown in FIG. 2, and pattern recognition for all pixels in the specific area is performed for number pattern recognition. It is preferable to perform pattern recognition using a small number of feature cutout lines selected in this way without any problems.

Of course, the setting of such a feature extraction line can be arbitrarily selected depending on the characteristics of the distribution signal. For example, FIG. 6A shows another feature extraction line suitable for the number "0",
It is preferable to cut out the specific chain acid 100 radially.

Similarly, FIG. 6B shows concentric feature cutting lines for the English letter "X", and in this way, the optimum feature cutting line is selected depending on the type of target image or audio information.

Of course, it is preferable that such feature cutting lines are usually lines along the vertical and horizontal base lines of the specific area, and should be selected so that the number of feature cutting lines can be minimized and can be commonly used for all images.

Returning to FIG. 5 mentioned above, the passage member subfunction and prohibition member subfunction for fuzzy inference, which are the features of the present invention, will be explained using the feature extraction line 106 as an example.

As is clear from FIG. 5, the double diagonally shaded areas El and E2 in the feature extraction line 106 are defined as the Excitation 6A area through which the number "3" must pass.

On the other hand, area 11. I2 is a prohibited area in which the number "3" must not enter, and is defined as an inhibition area.

Furthermore, region N is a region in which it is unclear whether or not it will pass due to variations in the number "3", and is defined as a neutral region.

Therefore, on this feature extraction line 106, there is a passage area El with respect to the feature extraction line direction X.

Passage membership functions μEl'μE2 corresponding to E2 are set separately, and similarly, a prohibited membership function μ μ is set for the prohibited region I2.12.

■2° ■2 Of course, at this time, no membership function is set for the neutral region N.

As is well known, the membership function is a function that has a grade (degree) from θ to 1, and is also called a membership degree function.
The passing membership function for the passing region E is shown in Fig. 7A,
As shown by B, it is given as a membership function of the part where all handwriting is likely to pass.

Therefore, by comparing this passage membership function μEl'μE2 with the actual distribution signal and determining the grade of the position where the handwriting passed, it is determined with what probability the handwriting passed through the passing area. be able to. Therefore, regarding this pass membership function, it is necessary to independently compare a plurality of pass regions with the distribution signal as separate pass membership functions even if they are on the same feature extraction line. Otherwise, even if the handwriting passes through only one of the passing regions, it will be determined that all the passing regions on the feature extraction line have been cleared.

The X marks on the passing membership functions shown in FIGS. 7A and B each indicate the locus position of the number "3" shown in FIG. The maximum value at that time is indicated as YμE1(x) depending on which position the distribution signal to be compared with the function passes through. Similarly, ¥μE2(x) is obtained for the passage area E2, and both of these indicate the grade at the actual handwriting position on the passage membership function.

In other words, if the handwriting completely passes through the passage area, the grade will be "1", while if it is in the slope part of the membership function, the grade will be determined by the slope of the membership function from "1" to "0". .

Therefore, for all passing areas E of the feature cutting line, the minimum value of the grade of each passing area El, E2 is determined, and the maximum passing value is determined as follows.

(vμ (X)) △ (Vμ (X))X
El x E2 to one (Vμ (X)) X Ei where i indicates the passband number (Ei).

Therefore, for the passing region E, the grade whose distribution signal is the farthest within the set passing region E is output.

On the other hand, the prohibited membership functions through which all handwriting is prohibited are shown in Figure 7C and D, and μI
It is expressed as I'μI2. This prohibited area I is originally an area where handwriting is not allowed to pass, and the maximum value of the prohibited membership functions and distribution signals in Figure 7 C and D, that is, Yμrt (X) Yμl2 (X) This grade indicates the maximum value of the distributed signal that has entered the prohibited area, and the value decreases as the pattern of the distributed signal moves away from the set pattern.

Therefore, in this embodiment, as an index indicating the possibility of not passing through the prohibited area, this grade is inverted and the grades of all prohibited areas within one feature extraction line, in this example, the feature extraction line 106, are calculated. The minimum value for the inverse value of is taken, which is given by the following equation:

Then, if j is the number Ij of the forbidden area, then the above formula and
Furthermore, using De Morgan's law, it can be expanded as follows.

1ΩμI(X) Then, the above equation can be further processed by a new composite prohibited membership function that combines all prohibited regions into one, regardless of the number of prohibited regions, as shown below, and this composite prohibited membership function is shown in FIG. 7E.

As described above, the sum of the grades that the distribution signal must pass through and the grades that have entered the prohibited area for the passing area E and the prohibited area I, respectively, are the possibility of passing through the passing area and the possibility of not passing through the prohibited area. By taking the minimum value of the characteristic, the grade μ for the th feature extraction line can be obtained, which is expressed by the following formula.

μ − [to (vμ − (x))] k 1 x El △ [8Vμ, (x)] Jx ■” − [to (Vμ (x) ) ] △ (Ωμt(x))
XEI And the final number recognition grade μ is μ″″Ωμk
Since the generation of the membership function described above is shown in detail in Japanese Patent Application No. 1983-206009 (Fuzzy Processor and MAX Circuit Assembly Apparatus) related to the invention by the present inventor, it will be described here. A detailed description of this will be omitted.

Comparison of Membership Function and Distribution Signal Figure 8 shows the principle of comparison between membership function and distribution signal.For membership function μ on the feature extraction line (x), distribution signal a1 is ! It will be "0" in areas other than the handwriting, and it will be "0" in the handwriting area.
1”.

Therefore, for the distribution signal a1, the minimum value of the membership function μ and the distribution signal a1 is taken for each X-chain saw on one feature extraction line, and this is applied to the entire area of the feature extraction line (x). The maximum value on the scanning line is output as grade μ.

Therefore, the grade μ is "1" for the distributed signal a1.

On the other hand, if the position in the passing region shifts as shown in the distribution signal a2, then the minimum value with the passing membership function μ becomes rO,75J, and this becomes the maximum value over the entire feature extraction line. Therefore, the grade μ is rO,75
It is concluded that J.

Detection of Distribution Signal Next, an apparatus for detecting the distribution signal a of the specific area 100 used in this embodiment will be explained based on FIGS. 9 to 11.

FIG. 9 shows a detector in which a specific area 100 is divided into nXm pixels, and has a structure in which, for example, line sensors such as COD are arranged in a plurality of rows. Therefore, this sensor 30 makes it possible to capture arbitrary pixel information from the specific area 100.

Therefore, by setting an arbitrarily selected feature cutting line from this sensor 30, it becomes possible to extract the distribution signal a for comparison that corresponds to the feature of the image information.

FIG. 10 shows a signal detection means for extracting a desired distribution signal as serial information from a predetermined feature extraction line of the sensor 30 by scanning control, and from the sensor 30 of the selected feature extraction line, a shift register 32 is output. A passing signal of handwriting is written into each element. In the example, the information on the position where the handwriting passed is "1" or rHJ, and the others are indicated as "0" or rLJ.

Transfer of pixel information from the sensor 30 to the readout layer, which is the shift register 32, is performed using L supplied to the shift register 32.
This is done at one time by the OAD signal, and then the shift register 32 is sequentially shifted by the CLK (clock) signal, and the distribution information a is read out as a serial signal.

The shift register 32 is provided with one dummy stage at both ends thereof,
The rLJ signal is stored in advance in this dummy first stage, and this causes the start of reading pixel data using the CLK signal (ST
ART) and end (S TOP).

In FIG. 11, the sensor 30 and the shift register 32 are shown integrally, and the pixel information and dummy information of the sensor 30 at this time are transferred to the shift register 32 by the rHJ input of LOAD, and then C' A state in which the LK signal is sequentially read out as a serial signal a is shown.

Specific Example of Fuzzy Neuron (First Example) When the distribution signal a from a specific area is taken in as described above, this signal a is assigned a predetermined membership in a fuzzy neuron predetermined for each number. The grade .mu. of each number is determined by comparison with the function, and the pattern of the input distribution signal a is determined starting from the one with the highest grade .mu..

FIG. 12 shows typical forms of membership functions for passing and inhibiting given to each number for a specific feature extraction line, and FIG. 13 shows timing signals for the fuzzy neuron. The output circuit is shown.

Here, a method for determining the grade of the extent to which the passing area and prohibited area are affected given the distribution information on a single feature cutting line will be described with reference to the embodiment shown in FIG.

Each fuzzy neuron 20 in FIG. 2 described above has a passing fuzzy synapse 20a1 in FIG.
(minimum value) output means 202, and the number of each pass or prohibit fuzzy synapse is selected depending on the target image, and in the case of the numbers in the example, seven feature extraction lines and the selected number of pass areas and prohibit areas. Determined by the combination of

For a single feature extraction line, the distribution signal a is output from the detection means shown in FIG. 11 described above, and in FIG.
This distribution signal a is commonly input to the passing fuzzy synapse 20a and the inhibiting fuzzy synapse 20b.

First, the passage fuzzy synapse 20a includes passage membership function means 40 for setting at least one unimodal passage membership function μE in the specific region, and for example, the passage membership function μE1 shown in FIG. 7A is targeted. It is set for each image information. The membership function at this time is either the S function shown in Figure 12A, the two functions shown in Figure 12B, or the π function shown in Figure 12C, which is a combination of both, alone or in combination. These are set in the above-mentioned patent application 1986-2Q60.
No. 09.

Unimodal pass membership function μ output from pass membership function means 40. is compared with the distribution signal a by the MIN circuit 42, and further
During line scanning, the MAX (maximum value) output means 44 outputs the maximum value among the membership values output from the MEN circuit 42. This maximum value output operation is
As described in the figure, in FIG. 1, the output of the MAX output means 44 is always fed back to the MAX output means 44 by the hold circuit 46 in order to obtain the maximum value in one line. The contents of this hold circuit 46 are reset every scan, and for this purpose an RE (reset) pulse is supplied to the hold circuit 46.

Among fuzzy synapses, there are cases where a membership subfunction is not defined for the feature extraction line, and for such fuzzy synapses, matching the distribution signal with the membership function becomes meaningless, and such undefined If a defined membership function exists, a “1” is sent to terminal 48.
This "1" signal and the maximum value output are supplied to the MAX circuit 50, and the output thereof is supplied to the MIN circuit 52 of the minimum value output means 20z.

Therefore, in the normal case where the member subfunction is defined, the NA multiplied signal at the terminal 48 is "0", and the M
The AX circuit 50 simply passes the maximum value of the MAX output means 44.

On the other hand, the prohibition fuzzy synapse 20b has almost the same configuration as the pass fuzzy synapse 20a, but has the feature that the maximum value output is inverted.

At least one prohibited membership function within the specific region, this function can be formed as a multimodal function along the feature cutting line as shown in FIG. The ship function means 60 outputs the prohibited membership function μm to the MIN circuit 62.

Similar to the passing fuzzy synapse 20a, the MIN circuit 62 is supplied with the distribution signal a, and the membership value at the position specified by the distribution signal a is output to the membership function μ, which is It is supplied to the MAX output means 64.

The maximum value of the MAX output means 64 is always held by a hold circuit 66, and this hold circuit 66 is reset for each scan by an RE pulse.

The maximum value is inverted by the inverting means 70 in the inhibit fuzzy synapse 20b, and the inverted output is compared by the NA multiplied signal MAX circuit 72 supplied from the terminal 68, and the inhibit membership function is defined in the inhibit fuzzy synapse 20b. If not, "1" is input as NA, and this verification is prohibited.

Similarly, the output of the MAX circuit 72 is output to the minimum value output means 20.
z is supplied to the MIN circuit 52, and this MIN circuit 52 also constantly obtains the minimum value for each line by a hold circuit 74, and the certainty of matching with the corresponding number is output from the sample hold circuit 76 as a grade μ. Ru.

Timing circuit As described above, the passing fuzzy synapse 20a and the inhibiting fuzzy synapse 20b generate the distribution signal a and the membership function μE,
The timing of each section for this purpose is controlled by the timing circuit shown in FIG. 13.

In FIG. 13, the original clock pulse from clock pulse generator 76 is applied to AND gate 78, while
The gate 78 is opened by a signal from JKFF8Q controlled by the start and stop signals, which are control signals, and the original clock pulse at this time is outputted from the gate 78 as a CLK signal to the shift register of the detection means.

Further, this CLK pulse is counted by a counter 82, and its initial output is used as a reset signal RE to reset each hold circuit.

The counter 82 has a counter up cycle designated in advance, and outputs a sample pulse SMP when the counter goes up, which is used as a hold signal for the sample hold circuit 76 of the minimum value output means 20z mentioned above, and when one line of scanning is completed, the MIN pulse SMP is output. The output of the circuit 52 is output as grade μ.

On the other hand, the sequentially increasing output of the counter 82 is converted into an analog signal by the D/A converter 84, and the output S is supplied to the read terminal of each membership function means 40, 60 in FIG. 1 as a staircase waveform as shown in the figure. and the distribution signal a
The membership functions μE and μ■ are output to each MIN circuit 42, 62 as signals synchronized with the serial signal of.

FIG. 14 shows a timing chart of the circuit shown in FIG. 13, in which clock signals are sequentially output from the start signal, and the output clock from the sensor 30 of the detection means to the shift register 32 is output in sequence from the start signal. Image information is loaded, and at the same time various reset actions are performed.
For this reason, the first stage of the shift register is used as a dummy stage.

The readout signal S of each membership function means is an analog output and has a stepped waveform as shown in the figure, and a sample pulse SMP is output at the end of one line scan, and the collated result is output as a grade μ.

As is clear from the above description, according to the embodiment shown in FIG.
The distribution signal a in a specific region is compared with a unimodal pass membership function to obtain a pass membership function μ. , the membership value of the position specified by the distribution signal a is obtained, and at the same time, the distribution signal a is checked against the prohibited membership function μI to determine the membership value of the position specified by the distribution signal a for the prohibited membership function μm. The maximum membership values within one line along the feature cutting line are output. Then, in the prohibition fuzzy synapse 20b, the maximum membership value obtained from the prohibition membership function μI is inverted, and the overall minimum value of this inverted output and the maximum value output obtained from the passing membership function μE is the grade μ This will be required as follows.

Of course, when setting multiple passbands on the same feature cutout line, the pass fuzzy synapses 20a are connected in parallel as many as the number of passbands required, and all their outputs are
It is sufficient to supply the signal to the MIN circuit 52.

Therefore, pattern recognition of the distributed signal can be performed by comparing the distributed signal a within the specific area 100 with the membership functions μE and μI given in advance.

Second Embodiment FIG. 15 shows a second embodiment of the fuzzy neuron according to the present invention. In this second embodiment, one membership function means 200 is provided for each number, and the signal detection means One shift register is also provided as shown by the reference numeral 202, and these membership functions μE/μ■ and the distribution signal a
are compared in one MIN circuit 204.

In the second embodiment, label data for obtaining a predetermined membership function is sequentially supplied from the label data memory 206 to the single membership function means 200, and the membership function μE/μI is calculated each time. It is possible to change.

On the other hand, data of a predetermined feature cutout line is sequentially selected and supplied from the pixel data memory 208 to the shift register 202, and thereby both memories 206.
By sequentially switching the contents of 208, it becomes possible to sequentially perform verification using a single verification device.

In order to switch the contents of both the memories 206 and 208,
An address decoder 210 is provided to read each label and pixel data from the memories 206, 208 in an order preset by a program counter 212.

Then, the maximum value of the output of the MIN circuit 204 for each line scan is determined by the MAX output means 214, and this M
A hold circuit 216 is connected to the AX output means 214, and the maximum value for each line scan indicated by each period T is held.

In the embodiment, the label data memory 206 supplies the NA multiplied signal MAX output means 214 for an undefined feature extraction line, and for such an undefined feature extraction line, a IJ, making the distribution signal a meaningless.

The output of the MAX output means 214 is connected to the M
IN output means 220, this switch 218 selects the membership function μE that the membership function means 200 passes through.
, the output of the MAX output means 214 is directly supplied to the MIN output means 220, while the membership function means 200 outputs the prohibited membership function μI.
When outputting , the output of the MAX output means 214 is supplied to the MIN output means 220 through the inverting means 222 .

Similarly to the first embodiment, a hold circuit 224 is connected to the MIN output means 220, and the output of the MIN output means 220 is fed to the grade μ
Output.

Therefore, according to the second embodiment, it is possible to simplify the circuit configuration and sequentially repeat the necessary collation while switching the label data and pixel data for membership creation stored in the data memory in a predetermined procedure. It becomes possible.

Third Embodiment In the above-mentioned embodiment, after the prohibited membership function and the distribution signal are compared and matched, the result of this comparison is inverted by the inverting means, but in the present invention, the prohibited membership function itself is inverted. is reversed by a reversing means,
The membership value at the position specified by the distribution signal is determined by this inverted prohibited membership function, and the minimum value within the specified area is output, while the maximum value output from the pass membership function is It is also possible to obtain the total minimum value from the minimum value and maximum value obtained from both the prohibited fuzzy synapse and the pass fuzzy synapse in the same manner as in the above embodiment.

Fourth Embodiment In all of the above embodiments, the specific passing membership function means and prohibition membership function means are -human power-output membership function circuits (patent application No. 63-206008), a staircase wave sweep signal is used as an input, and the distribution signal obtained from the feature extraction line is also passed through the shift register 32 and sequentially transferred to the MI
The signal is sent to the N circuit 42.

This embodiment is a time sweep mode type system, and although the system is simple, it has the disadvantage that it takes a long processing time.

Therefore, fuzzy neurons that enable parallel processing are effective.

FIG. 16 shows an example thereof. The distribution signal on the feature cutout line taken in from the sensor 300 is supplied to the MIN array or switch array 310 (for passing) and 312 (for prohibition), and a predetermined peak distributed on the output line of each membership function generation circuit 320 is output. ON10FFLM for gender pass membership function or prohibited membership function
Send to AX circuits 330 (for passing) and 332 (for M circuits).

General.

The output of the MAX circuit 330 corresponding to the overmembership function is supplied as is to the MIN circuit 340. On the other hand, the output of the MAX circuit 332 corresponding to the prohibited membership function is supplied to the MIN circuit 340 via the inverting means 350.

Although FIG. 16 only depicts blocks related to a single feature extraction line, in reality, similar blocks are prepared in parallel as many as the number of feature extraction lines, and their output (M
The output of the AX circuit 330 and the output of the inverting means 350) are
If supplied to the rN circuit 340, the output of this MIN circuit 340 will give the grade of a predetermined pattern of numbers, letters, etc.

Regarding the membership function generation circuit, please refer to Japanese Patent Application No. 61-268568 and Japanese Patent Application No. 6
1-268569, detailed description thereof will be omitted here.

[Effects of the Invention] As explained above, according to the present invention, when pattern recognition is performed on distribution information such as images or audio, the probability of passing areas and prohibited areas is determined by comparing the membership function and the distribution signal. This method has the advantage that pattern recognition can be performed while searching for , and that recognition results can be reliably obtained even for ambiguous pattern recognition.

[Brief explanation of drawings]

Figure 1 is a block circuit diagram showing an example using a pair of passing fuzzy synapses and inhibiting fuzzy synapses when the fuzzy neuron according to the present invention is used for numeric pattern recognition, and Figure 2 is for digit recognition. An overall configuration diagram of the fuzzy neuron according to the present invention, FIG. 3 is an explanatory diagram of a singleton showing the grade output result for identifying the most probable number from the grades obtained in FIG. 2, and FIG. An explanatory diagram showing the data collection process for creating membership functions in numeric pattern recognition. Figure 5 is an explanatory diagram showing feature extraction lines and passing and prohibited areas in digit pattern recognition. Figure 6 is another example of feature extraction lines. FIG. 7 is an explanatory diagram showing the passing membership function and prohibited membership function for a specific feature extraction line. FIG. 8 is an explanatory diagram showing M IN comparison between the membership function and the distribution signal. Fig. 9 is an explanatory diagram showing an example of a sensor for image recognition, and Fig. 10 is an explanatory diagram showing a combination of a sensor and a shift register for converting a sensor signal at a feature extraction line into a distribution signal that is a serial signal. Figure 11 is a simplified configuration diagram of the detection means in Figure 10, Figure 12 is a waveform diagram of the principle of the membership function, and Figure 13 is a simplified diagram of the detection means in Figure 10.
The figure is a timing circuit diagram of fuzzy inference in Figure 1, Figure 14 is a timing chart diagram of fuzzy inference, Figure 15 is a block circuit diagram showing another embodiment of the fuzzy neuron according to the present invention, □ Figure 16 is a timing circuit diagram of fuzzy inference. FIG. 7 is a block diagram showing still another embodiment of the parallel fuzzy neuron according to the present invention. 20 ... Fuzzy neuron 20a ... Passing fuzzy synapse 20b ・
... Prohibition fuzzy synapse 20z ... Minimum value output means 30 ... Sensor 32 ... Shift register 40 ... Passage membership function means 44 ...
Maximum value output means 60... Forbidden membership function means 64...
Maximum value output means 70 ... Inversion means 100 ... Specific areas 101 to 107 ... Feature cutting line a ...
Distribution signal μ... Grade μE... Passing membership function μ■... Prohibited membership function pattern ↓ Nin Wari Tan, Sobo 3 Meshiba Mi, Me7a Kaitsu 5ξ, Tsununo Azusagi Dai 4 Figure % <'Jz91 human soil) to be dumb and brave Figure 5 A B , / 已-) 。もMinistry〉nosodashi layer stupid - Figure 6 0! Menhashi 77″F4 kettle (tsu4乍j＼ Figure 7 10000s 1 Membership Sekiyu door and now 7+ta Otsu's MIN clay flight Figure 8 name 9 Figure 1o Figure 1) Figure 11 Member 岑Bu PA's house (Fig. 12, Fig. 15')

Claims (2)

[Claims] (1) A fuzzy neuron that recognizes a distributed signal within a specific region, comprising: pass membership function means for setting at least one unimodal pass membership function within the specific region; and at least one pass membership function within the specific region. prohibited membership function means for setting prohibited membership functions; a value output means; an inversion means for inverting the maximum membership value obtained from the prohibited membership function; and an output obtained by inverting the maximum value output obtained from the passing membership function and the maximum value of the prohibited membership function. a minimum value output means for determining the overall minimum value from and, and a fuzzy neuron, characterized in that the distribution signal is recognized by comparing the distribution signal in a specific region with the membership function given in advance. (2) a fuzzy neuron that recognizes a distributed signal within a specific region, a pass membership function means for setting at least one unimodal pass membership function within the specific region; and for the pass membership function. maximum value output means for determining a membership value at a position specified by the distribution signal and outputting the maximum value within the specific area; and prohibited membership for setting one prohibited membership function within the specific area. function means; inversion means for inverting the prohibited membership function; a total minimum value output means for calculating the overall minimum value from the maximum value output obtained from the pass membership function and the minimum value output obtained from the prohibited membership function; A fuzzy neuron that recognizes a distribution signal by comparing the distribution signal within a region with the membership function given in advance.