JPS5853790B2 - Pattern recognition method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a method for recognizing patterns such as characters and figures, and in particular, to a pattern recognition method that enables easy and accurate recognition of patterns with severe shape distortions and positional deviations. It is related to the method.

Generally, in a pattern recognition device, a human cover pattern is classified by comparing the input cover pattern with a plurality of pre-stored reference patterns by some method and determining which reference pattern it is most similar to.

Each recognition method has its own way of comparing these two patterns, but in general, in conventional pattern recognition methods, if the presentation position or size of the input pattern differs from the reference pattern, both patterns are compared. I couldn't make a proper comparison.

To this end, before comparing both patterns, the positions and sizes of the input patterns are "normalized" according to an appropriate law, so that the positions and sizes of both patterns are made the same, and then the two are compared.

For example, to normalize the position, you can move the shape horizontally and vertically to align the center of gravity of the shape, or to normalize the size, you can stretch the character pattern to increase the vertical width and horizontal width of the character. They used methods such as aligning the

However, such conventional comparison methods are easily affected by noise contained in the input cover pattern, that is, smudged or blurred patterns, and it is difficult to recognize multiple printed characters, that is, multiple fonts, and it is difficult to recognize handwritten characters. It was like being powerless.

An object of the present invention is to provide a pattern recognition method that eliminates the above-mentioned conventional drawbacks and allows patterns to be accurately and easily recognized even if the input cover pattern has severe shape distortion or positional deviation. In other words, the pattern recognition method of the present invention divides the input cover pattern and the reference pattern into small fields and performs the comparison in stages, and at each stage of the comparison, the position of the reference pattern itself to be compared with the input cover pattern is determined. Normalization is performed in parallel with pattern comparison, thus eliminating the need to perform position and size normalization independently in advance, and comparing multiple parts that make up an information pattern. A plurality of first feature extraction boards capable of discriminating each of the plurality of types corresponding to the plurality of types formed by classifying the patterns are provided, and patterns representing input information are partially superimposed on each other. divide into a plurality of sections and supply each to a plurality of first feature extraction boards to determine the type and position of the similar type corresponding to a partial pattern included in the pattern; The first discrimination output information respectively extracted from the first feature extraction board is supplied to a plurality of first feature alignment plates, and the first discrimination output information is parallelized while allowing positional deviation within a certain range. the first spatial pattern information extracted from the plurality of first feature arranging boards respectively is supplied to the plurality of second feature extraction boards to form the pattern. The types and positions of at least two combinations of the similar types corresponding to the included partial patterns are determined, and a plurality of pieces of second discrimination output information are respectively extracted from the plurality of second feature extraction boards. A recognition operation is performed by supplying the second discrimination output information to a second feature arranging board and arranging the second discrimination output information in parallel as second spatial pattern information while allowing positional deviation within a certain range. The present invention is characterized in that a pattern representing the input information is recognized by repeating a recognition operation at least twice.

In the following, the invention will be described in detail by way of example embodiments with reference to the drawings.

According to the pattern recognition method of the present invention, even if the position or size of the presented pattern is shifted, or even if there is some distortion in the shape of the presented pattern, the correct pattern can be recognized without being affected by it. However, the basic concept of the pattern recognition method of the present invention will be explained first.

In the method of the present invention, in order to perform pattern recognition without being affected by misalignment, shape, or size, when comparing the input cover pattern and the reference pattern, both patterns are recognized at various positions. When observed in a small field of view, if the difference between both patterns does not exceed a certain amount in any small field of view, pattern recognition is performed by treating both patterns as the same and generating a recognition output that displays the same. Do it like this.

Pattern recognition by comparison of a small field of view as described above is not performed by only one stage of operation, but the comparison output of one stage is used as the input of the next stage, and the same operation is performed in multiple stages while expanding the field of view. Try to do it repeatedly in series.

Furthermore, in the comparison at each stage, a certain degree of tolerance is allowed little by little for pattern positional deviation.

Furthermore, as described above, the size of the field of view for observing the pattern is increased toward the later stages, and the final stage has a field of view large enough to observe information about the entire input cover pattern at once.

As described above, the input cover pattern is divided into small fields of view and compared with the reference pattern, so in the putter/recognition method of the present invention, there is a considerable difference in the shape of both patterns, but when observed in a large field of view, it is not the same. Even if it is difficult to determine the nature of the pattern, if both patterns are compared minute by minute and their shapes match within an acceptable range, they can be considered to be the same pattern.

In principle, such an operation is equivalent to preparing multiple reference patterns with different positions and observing whether the input cover pattern matches any one of them. become.

In this case, even if, for example, the upper part of the figure matches the reference pattern at one position and the lower part of the figure matches the reference pattern at another position, each observation is performed for each small field of view. Therefore, the input cover pattern and the reference pattern are considered to match.

Therefore, even if the input cover pattern has some shape distortion, correct pattern recognition can be performed.

In the above explanation, we used the expression of comparing the input cover pattern with the reference pattern, but in the circuit configuration of a specific recognition device, the reference pattern is recorded as is somewhere in the circuit device. isn't it.

A basic arithmetic element with multiple inputs and one output that has characteristics abstracted from a neuron will be simply referred to as a "cell" below.
By setting the distribution of the coupling coefficient values for each input terminal of a plurality of inputs in such a "cell" in advance in correspondence with the shape of the reference pattern, the input pattern can be supplied to that cell. When this happens, an operation corresponding to comparison with a reference pattern is performed, and the result is taken out as a recognition output.

In order to determine this coupling coefficient, all possible types of patterns are artificially set in advance by investigating the nature of the input pattern to be recognized. It is also possible to set the coupling coefficients of a plurality of input lines and form them in a self-organizing manner so as to set a desired distribution.

FIG. 1 schematically shows an overview of the overall configuration of a circuit device that performs pattern recognition according to the present invention based on the above-mentioned basic concept.

The pattern recognition device shown in the first diagram includes an input photoelectric conversion layer U.

It consists of a combination of a feature extraction plate Usl (1=1 to L) and a feature alignment plate Ucl (1=I to L) coupled thereto.

Here, the input photoelectric conversion layer U.

is a device consisting of photoelectric conversion elements on a two-dimensional substrate, and the first feature extraction plate U8□, that is, U
The 81st layer is an input photoelectric conversion layer U.

It consists of the same number of circuit elements, and the individual circuit elements are, for example, disclosed in Japanese Patent Application Laid-Open No. 51-35 by the present inventor.
A circuit element having a shunt-type suppression input disclosed in Japanese Patent No. 255 is used.

One circuit element in this feature extraction board Us1 is:
Photoelectric conversion layer U.

A photoelectric conversion layer U that is coupled to a plurality of photoelectric conversion elements in a specific region of the photoelectric conversion layer U to which each circuit element is coupled.

In the corresponding regions above, the photoelectric conversion element groups are sequentially overlapped with each other while being shifted one element at a time, and the combinations of photoelectric conversion elements are different for each area.

The mode of such binding is schematically shown in FIG. In the bonding mode shown in FIG. 2, for example, U.

Area U of the layer.

The output of the photoelectric conversion element included in No. 1 enters the circuit element U8□1 of the U8□ board, the output from the No. 2 area enters the circuit element U312, and so on. There is.

In such a manner, U8, all circuit elements on the board are U.

They overlap each other on the layer and are combined in correspondence with all different regions.

The recognition device of the present invention includes the above-mentioned primary feature extraction board U.
8. One feature extraction plate U8□ is provided with a photoelectric conversion layer U in the above-described manner.

are connected in parallel to the individual photoelectric conversion elements.

One feature extraction board for each of them has (kl-1
The explanation will be given by adding the symbol 1) to .

Therefore, the Usl(1 to Usi(K1)) layers discriminate different types of patterns from each other.

In other words, for example, several patterns constituting a character are categorized into types, and each of these types is classified into U8□ (S~
Assign them to the U8□ (K1) layer respectively, and
The (1) to U8□ (K1) layers are provided with the ability to discriminate whether or not they are of the assigned type.

For example, as shown in Figure 3, in the U layer, 51 (1
) Discriminates one letter, U layer discriminates T type, Ma・s
1 (2) In addition, the U8□(3) layer discriminates between vertical and letters, and up to the Ust (K1) layer, shapes corresponding to specific patterns are created to be able to handle a variety of shapes. Have the ability to discriminate.

In addition, although the number of feature extraction boards is reduced in Fig. 3 for simplicity, in reality, the more the number of feature extraction boards is increased, the more
It can be configured to accommodate different patterns at different locations.

The pattern discrimination function described above is provided by the following configuration.

That is, as mentioned above, the individual circuit elements on the U8□ board are configured as described in Japanese Patent Application Laid-Open No. 5135255, but how to set each type according to the above-mentioned pattern is explained below. For the sake of explanation, its configuration is simplified and shown in FIG.

In the configuration shown in FIG. 4, if the circuit element of the U8□ layer that extracts information from a specific region X of the Uo layer is U8□, then the photoelectric conversion element U in the region X of the Uo layer.

The outputs of x1 to IJox9 are added as positive inputs to the linear summation element Σ.

At the same time, a suppression force v8 consisting of the average value of the outputs from the elements of the entire Uo layer is added as a negative input.

Now, among the circuit elements U8□ and one U81 board of the U8□ board, if we assume that it belongs to U which has the function of discriminating the horizontal bar, for example, the photoelectric converter 5l(1) converter element U.

Gain adjusting elements RX4.K4, UOX5, and UOX6 are provided on the lines for taking out their outputs. Rx5. The gain of Rx6 is set to dog so that when a horizontal bar pattern comes in, a large output is taken out from the linear summation element Σ.

The above-mentioned similar pattern settings are such that each circuit element of the Us□ layer, that is, the Us, (+) board to the USI (K1) board, is connected in parallel to each element of the U layer. This is done for all circuit elements, and the similar patterns set for each are different for the US1 (1) board to the U31 (K1) board.

Furthermore, the same U8□ board, i.e., a specific U8□(kl
) The circuit elements of the board are all capable of discriminating the same similar pattern.

For example, all circuits 51(1) on the U board can be distinguished no matter where on layer 06 there is a pattern of horizontal bars, and the area on layer 06 where the pattern of horizontal bars was. Take out the output from the corresponding circuit element.

U as described above.

The type of pattern on the layer and the position where the similar pattern exists are determined by which circuit element of the U8□(k,) board the output is taken from.

The output of such U8□(kl) layer is then converted into each Usi(
A primary feature alignment plate U provided corresponding to kl).

1, that is, U. U to feed one board.

1 board has fewer circuit elements than the U8□ board, and
U8□U for each specific area on the board.

Specific circuit elements on one board correspond to each other.

Therefore, as shown in Figure 3, on the Uo1 board,
The output is taken out from the circuit element corresponding to the area on the UU8□ board where the circuit element from which the output is taken exists.

That is, in that case, the light image that entered the U8 layer is divided, and the similar patterns included in the light image are determined in the U8 and plate, and the positions where each of the similar patterns exist are also determined. and its position will be normalized on the Uo1 board.

In the configuration example shown in FIG. 3, the primary feature extraction plate U
81 and the primary feature alignment board U.

1, the letters ITJ are recognized, but in reality, it is necessary to recognize a variety of patterns using a large number of circuit elements, so as shown in Figure 1,
The above-mentioned combination of feature extraction board and feature alignment board is
It is necessary to arrange them in series over a plurality of stages, including the second, third, . . . , first stages, and perform feature extraction and feature alignment L times over L stages.

For example, information is transferred from the Uo1 board to the second feature extraction board U8° as shown in FIG. 5, for example.

In the example of information migration shown in FIG. 5, the number of circuit elements on the U82 board is U.

U equal to the number of circuit elements on one board.

The bonding relationship between the 1 plate and the U8° plate is that the photoelectric conversion layer U
.

The connection relationship between the U81 plate and the U81 plate is the same.

However, while the U8 layer was just one layer, the Uo1 layer was U.

1 (1) to 'JCI (K1), and the U32 board consists of two sheets to U8□ (1) to US2 (K2), so the U is made up of multiple sheets.

1(k+) and Us2(K2) intersect with each other in a 0 mutual coupling relationship, and are connected in parallel to all of the other parties.

Therefore, K2 secondary feature extraction plates US2(1) to U
8□ (K2) is similar to U8□ (kl) board,
It is configured so that different types of patterns are divided and discriminated, but in this case, the types of patterns that are divided respectively are different from those on the U8□(kl) board,
It consists of a combination of various similar patterns set on the U8□ (kl) board.

For example, in the Us2(1) board, as shown in the figure, the combination of the horizontal bar on the left and the T-shape on the right is discriminated, and the combination of the horizontal bar on the left and the T-shape on the right is discriminated,
Assuming that the combination of the T-shape on the left side and the horizontal bar on the right side is to be discriminated in the two boards, as shown in FIG. Its location is displayed.

U8□ board and U.

The mutual relationship between the two plates is U81 plate and U. The structure is similar to that of the first board, and furthermore, position normalization and similar pattern discrimination are performed more precisely.

The recognition operations performed by each feature extraction board and feature alignment board in recognizing the pattern of, for example, the character rAJ as described above will be described with reference to FIG. 6, which schematically shows the recognition operations.

FIG. 6 shows a tertiary feature alignment plate U.

3 shows an example of the recognition process when the pattern of the character rAJ is finally recognized by the output of 3.

Generally, by repeating the recognition operation in the procedure described above multiple times and accumulating the results, the number of circuit elements on the feature extraction board U and the feature alignment board Uc rapidly decreases at each successive stage. Then, the final L-th feature alignment board U is created.

In L, the number of circuit elements on the board is one, and instead, the number of feature alignment boards that are responsible for discriminating each type of pattern from 1 to KL is determined by the number of feature alignment boards that handle the discrimination of each type of pattern from 1 to KL. Each feature alignment plate Uc increases to the same extent as the number of pieces.
Each of L (KL) corresponds to all the target characters.

That is, in that case, each feature alignment board U.

The similar patterns that L(1) to U and L(KL) each share are patterns of characters and the like.

A general explanation of the operation of the pattern recognition device according to the present invention having such a configuration is as follows.

Input photoelectric conversion layer U.

Since the photoelectric conversion elements corresponding to the photoreceptor cells are arranged in a matrix in a two-dimensional base circuit, the output of the nX-th photoreceptor cell in the ny row is U.

(n). Note that the symbol is a two-dimensional coordinate representing the position of the cell, n-(nx, ny
).

The output of the Usl board included in the first stage basic circuit is
(kl, n) and the output of the ucl board is UcDkL
n), where U is a two-dimensional coordinate representing the position of the center of the receptive field of those cells on the Uo layer, as described above.

Therefore, as described above, the input photoelectric conversion layer U.

In , cells exist at all positions corresponding to the integer values of kl nX, ny, but as the number of stages increases, the density of cells becomes coarser, and cells do not necessarily exist at all integer coordinate positions. .

Further, the coefficient kl is a parameter that specifies the type of optimal stimulation for the receptive field of the cell, that is, the type of feature extracted by the cell, and takes an integer value.

The U8 board or U mentioned above.

The board consists of a set of S cells, ie, S circuit elements, or C cells, ie, C circuit elements, each having the same coefficient kl.

Thus, S cells and C cells are excitatory cells, that is, cells that send signals to their partner cells when there is a stimulus input, but the U8 plate and U.

Additionally present in the plate are the inhibitory cells v81(n) and vcl(n), respectively.

Now, using the above-mentioned symbols, the output of each cell, that is, each circuit element, can be expressed as follows.

S cells have shunt-type inhibitory inputs JP-A-51-35
This is the circuit element described in the No. 255 publication, and its output is given by the following equation.

Here, al(kl-1, ν, kl) represents the strength of excitatory connections, and bl(n) represents the strength of inhibitory connections.

Furthermore, rl is a positive parameter that controls the strength of inhibitory binding, and the larger the value of rl, the better the selectivity of the cell response.

Therefore, the value of rl actually used is set to an appropriate value depending on the ability to distinguish between similar patterns and the ability to tolerate shape distortion.

Note that in the case of 1-1 in equation (1), ucl-□(
kl-1, n) is interpreted as representing uo(n), and by the way, the total number of 1's Kl in the case of ■=0 is 1.

Next is the inhibitory V.

The cell has a root-mean-square (r, m, s) type input-output characteristic, and forms an output as shown in the following equation.

Here, 01-1 (1) represents the strength of excitatory fixed synaptic connections.

In addition, in equations (1) and (3), the total range of ν,
That is, Sl, which specifies the range in which one cell receives and receives input, is small at the first stage 1 = 1, and increases as it goes to the later stage, that is, as ■ becomes larger.In the final stage, the input photoelectric conversion layer U.

Make it large enough to cover the entire surface.

Next, like the S cell, the C cell is a circuit element having shunt type input/output characteristics, but its output exhibits saturation characteristics.

The output of such a C cell can be expressed mathematically as follows.

In addition, the inhibitory cell v8 is a circuit element that simply outputs the arithmetic mean of its inputs as an output, and its output is given by the following equation.

In equations (4) and (6), the total range of ν, that is, 1
Dl, which specifies the range in which cells receive and take input, is Sl
Similarly, it is specified that the first stage is narrower and the later stages become wider.

Therefore, as can be seen from equation (1) above, the value of the excitatory connection al (kl-Lν, kl) is - It does not depend on the position n of the receptive field.

This means that each S cell has input connections with the same spatial distribution.

Therefore, the only difference between the S cells included in the same -S plane is that the position of the previous cell that sends the input signal to each cell is different.

Next, the coupling coefficient between each cell is set as follows.

That is, cl-1(y) and di(+') are set to be monotonically decreasing functions of distance 1v.

The magnitude of the spread is respectively Sl and DI, and outside thereof O.

That is, there is no coupling outside of Sl or DI.

On the other hand, the coupling coefficients al(kl-1, v, kl) and bl(kl) are artificially determined in advance and incorporated as fixed bonds, but they are formed self-organized according to the input pattern. You can also do that.

However, in the artificial construction method or in the state at the time when self-organization is completed, the construction method of these coupling coefficients is the same for any stage of S cells, but it is possible to I have selected all the iso-symptoms.

As described above, the S cell in the subsequent stage is designed to be able to extract a combination of a plurality of features extracted in the previous stage as a new feature.

Therefore, when focusing on one S cell on the Usl plate shown in FIG. 4, the feature that the U81 plate should extract is the input photoelectric conversion layer U.

The value of the excitatory connection that reaches the S cell from the photoreceptor cell, which produces an output when the characteristic is presented, is selected to be a large positive value, and conversely, the value of the excitatory connection that reaches the S cell does not output when the characteristic is presented. Since the value of excitatory connections from photoreceptor cells is set to O,
As can be seen from equations (1) and (3), the input layer U is connected to the S cell.

cells and the V that sends inhibitory signals to their S cells.
Input layer U connected to cells.

Since this S cell is a perfect match with the cell of , this S cell is V from the Uo cell which does not output for that feature.

It will only receive inhibitory signals via cells.

Therefore, if the ratio between the strength of inhibitory connections and the strength of excitatory connections is selected to an appropriate value, each S cell will be able to connect to the input layer U, as described above.

It is possible to make the robot respond only to one type of feature presented in the image, or to a slight variation of that feature.

Also, as can be seen from equation (4), the input of U is the collection of all the outputs of the circuit elements existing within a small area of the U8□ board.

It is the basic circuit element. Therefore, like the corresponding S cells on the U8 plate, C cells on the Uo plate produce output in response to only one feature, but they are not very sensitive to changes in the presentation position of that feature. , its characteristics are the input layer U.

As long as it is presented within a certain area within the body, a reaction output will be produced.

In the pattern recognition device according to the present invention, the basic circuits as described above are connected in cascade, so that
As mentioned above, the receptive field of the circuit element in the final stage feature alignment board UcL is the human power layer U.

Thus, each circuit element responds only to a specific shape of the input cover turn, regardless of the presentation position of the input cover turn.

As is clear from the above description, according to the present invention, in the process of pattern recognition, the comparison with the input cover pattern is performed step by step, and at each step of the comparison, the position normalization operation is performed in parallel. Therefore, there is no need to independently normalize the position of the input pattern to be presented to the pattern recognition device, and even if the presentation position of the input pattern shifts, it will not be affected by the position shift. It is possible to correctly recognize patterns without any problems.

Further, even if the size of the input pattern changes somewhat or the shape is slightly distorted, correct pattern recognition can be performed.

The pattern recognition method of the present invention can be widely applied not only to the recognition of patterns such as characters and figures, but also to object recognition and auditory pattern recognition, that is, voice recognition.
When applied to speech recognition, for example, input patterns converted into spatial patterns on the frequency axis or reaction patterns of the basilar membrane can be used as input.

[Brief explanation of drawings]

Fig. 1 is a diagram schematically showing the process of pattern recognition according to the method of the present invention, Fig. 2 is a diagram schematically showing an example of information transfer in the organizational process, and Fig. 3 is a diagram schematically showing the pattern recognition process. A diagram schematically showing another example of the information transfer mode in the process, FIG. 4 is a diagram schematically showing an example of the discrimination mode of similar patterns, and FIG. 5 is a diagram schematically showing another example of the similar pattern discrimination mode. Diagram 6 schematically showing another example of the manner of determination of
The figure is a diagram schematically showing an example of a recognition operation mode in a pattern recognition device according to the present invention. Uo...Input photoelectric conversion layer, Usn(Ki)...
...Feature extraction board, Uon (K'i)...
・Feature alignment plate, uoXY... photoelectric conversion element,
U, ... feature extraction circuit element, XI ucxi ...- feature alignment circuit element, Σ...
・Linear summation element.

Claims (1)

[Claims] 1. A plurality of first feature extraction boards each capable of discriminating between the plurality of types corresponding to the plurality of types formed by classifying the plurality of partial patterns constituting the information pattern are provided, and input A pattern representing information is divided into a plurality of sections, parts of which overlap with each other, and each section is supplied to the plurality of first feature extraction boards, so that the type of the similar pattern corresponds to a partial pattern included in the pattern. and the position, respectively, and supply the first discrimination output information respectively extracted from the plurality of first feature extraction boards to the plurality of first feature alignment boards to obtain the first discrimination output information. The first spatial pattern information is arranged in parallel while allowing positional deviation within a certain range, and the first spatial pattern information respectively taken out from the plurality of first feature alignment plates is arranged in a plurality of second feature alignment plates. and determining the types and positions of at least two combinations of the similar types corresponding to the partial patterns included in the pattern, respectively, from the plurality of second feature extraction boards. The extracted second discrimination output information is supplied to a plurality of second feature alignment plates, and the second discrimination output information is converted into second spatial pattern information in parallel while allowing positional deviation within a certain range. A pattern recognition method characterized in that a recognition operation is performed by arranging input information, and a pattern representing the input information is recognized by repeating such recognition operation at least twice.