JP2515732B2 - Pattern matching device - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a pattern collation device used for pattern recognition, and particularly, it is highly efficient in collating an input pattern that allows deformation in handwritten Chinese character recognition and the like with a large number of standard patterns. The present invention relates to a pattern matching device that is executed in.

[Conventional technology]

With an optical handwritten kanji reading device (handwritten kanji OCR), which is a high-speed input means to a large number of documents and forms calculators written by handwritten kanji kanaji sentences, input patterns of character images obtained by photoelectric conversion processing are used. On the other hand, a recognition processing process is performed in which preprocessing, feature extraction processing, and pattern matching processing are performed and a category name of a character type such as Japanese JIS code is output as a recognition result.

Here, pre-processing is processing such as noise removal of input patterns, position normalization, size normalization, and the like, and feature extraction processing is to generate an input feature vector from a pre-processed input pattern according to a predetermined algorithm. In the pattern matching process, the candidate categories are extracted based on the similarity relationship between the input feature vector and the standard feature vector generated in advance from the standard pattern for each recognition target category. Processing.

Conventionally, a character recognition device such as a handwritten Chinese character OCR has several hundred vector elements for each of the input pattern and the standard pattern in order to accurately absorb the complex input pattern unique to the Chinese character and accurately match it with the standard pattern. Expressed as a high-dimensional feature vector, the pattern matching device calculates the similarity between these feature vectors using the inter-vector distance as a scale, and the category name of the standard feature vector that minimizes the inter-vector distance is the first candidate. The method of extracting as a category has been taken. Also, in the case of handwritten Chinese characters, it is difficult to narrow down the categories to be matched with a simple classification rule, so a method to calculate the vector distance by brute force with the input feature vector and the standard feature vectors of all categories to be recognized is available. Has been used.

Each vector element of the feature vector is a physical element value that reflects a local or global character line structure or background structure, and is generally quantized within a range of about 1 byte.

A specific method of calculating the distance between vectors is as follows. The number of elements (number of dimensions) of the feature vector is m, the number of categories is n, and the input feature vector I is I = (i ₁ i ₂ ...... i _m ) No. 1 If the standard feature vector J _y of the y-th category in the categories from No. to n is J _y = (J _1y J _2y ... J _my ), the input feature vector I and the standard feature vector J _y The vector distance D _y is [City block distance] [Euclidean distance] Calculation formulas such as are generally used. Where W _xy is
It is a vector of W _y = (W _1y W _2y ...... W _my ), which is a weight vector representing the weight between the elements of the feature vector.

Such a vector distance is obtained for each standard feature vector, a category name is added, and a sorting process is performed to extract the category name having the smallest vector distance as the first candidate category.

[Problems to be solved by the invention]

As described above, in the conventional pattern matching device, it is necessary to perform the cumulative calculation of the number of elements m × the number of categories n (times) for the pattern matching for one character, and the number of elements is extremely large due to a high-dimensional feature vector such as handwritten Chinese character recognition. When a candidate category is extracted from among the above categories, the total calculation amount becomes enormous, and in the character recognition device, the pattern matching device has a problem of limiting the recognition speed.

On the other hand, as a conventional technique for reducing the total calculation amount,
In the large classification pattern matching process, the vector distances are calculated by brute force using feature vectors with a small number of elements, and in the next detailed classification pattern matching process, about 100th from the first candidate category obtained from the large classification result. A pattern matching processing method of two stages is known in which the correct vector category is extracted by calculating the inter-vector distance with a high-dimensional feature vector for detailed classification only for the category. However, in this method, if the number of elements of the large classification feature vector is reduced, there is a risk that the correct category is rejected from the candidate categories of the large classification result, and the total calculation amount cannot be effectively reduced.

SUMMARY OF THE INVENTION An object of the present invention is to provide a pattern matching device capable of extracting candidate categories from all categories with high accuracy and high speed by cumulative calculation for some categories instead of all categories.

[Means for Solving Problems] A pattern matching device according to the present invention includes a dictionary memory that stores a dictionary in which category names are sorted in order of element value for each vector element of a standard feature vector; Based on the selection processing means for selecting and reading a part of the category name from the dictionary memory for each vector element according to the size of the element value of the feature vector, and the category name read for each vector element by this selection processing means Evaluation value accumulation processing means for accumulating evaluation values for each category,
The evaluation value accumulation processing means is configured by a sorting processing means for extracting candidate categories according to the magnitude relation of the accumulated value of the evaluation values for each category.

[Action]

According to the present invention, by adopting the above-mentioned configuration, the evaluation value cumulative value corresponding to the inter-vector distance between the input feature vector and the standard feature vector corresponding to each category is selected for some categories selected for each vector element. Therefore, the total calculation amount in the pattern matching process can be significantly reduced.

〔Example〕

 Hereinafter, details of the present invention will be described with reference to embodiments.

FIG. 1 is a block diagram showing the first embodiment of the pattern matching device of the present invention. In the figure, 1 is a dictionary memory, 2
Is a selection processing means, 3 is an evaluation value accumulation processing means, 4 is a sorting processing means, 5 is a processor, 6 is a program memory, 7
Is a general-purpose memory, 8 is a 2-boat memory, 9 is an input bus, 10
Is a connection bus, 11 is an output bus, 12 is an address bus, 13 is a data bus, 14 is a program address line, and 15 is a program read line.

When the input feature vector generated from the unknown input character pattern is input from the input bus 9, the pattern matching device of FIG. 1 performs a pattern matching process with the standard feature vector corresponding to the category, and outputs a candidate as a matching result from the output bus 11. It has a function to output the category name. Here, the number of elements (number of dimensions) of the feature vector is m, the number of categories is n, and each element value of the feature vector is quantized to L + 1 level from O to L. The configuration and function of each means will be described below.

The dictionary memory 1 is connected to the selection processing means 2 by an address bus 12 and a data bus 13. The dictionary table shown in FIG. 2 and the dictionary management table shown in FIG. 3 are stored in the dictionary memory 1. These two tables correspond to dictionaries created by sorting the standard feature vectors of all categories in the standard feature table shown in FIG. 4 by sorting the category name for each vector element in the order of the element value. Incidentally, the arrangement state of the category names shown in the dictionary table of FIG. 2 is an example, and does not show the state peculiar to this embodiment.

The selection processing means 2 comprises a processor 5 for executing the selection processing described later, a program memory 6 storing a control program for the selection processing, and a general-purpose memory 7 storing the input feature table shown in FIG. The processor 5 is configured for high-speed control, and is independently connected to the program memory 6 by a program address line 14 and a program read line 15, and to the general-purpose memory 7 by an address bus 12 and a data bus 13. There is. The input feature table of the general-purpose memory 7 can receive an input feature vector from the outside via the input bus 9 and the processor 5. Further, the processor 5 and the dictionary memory 1 have an address bus 12 and a data bus 13 which are common to the general-purpose memory 7.
Connected by. With the above configuration, the selection processing means 2 realizes a selection processing function of selecting and reading a part of category names from the dictionary memory 1 for each vector element according to the size of the element value of the input feature vector.

The evaluation value accumulating processing means 3 is a processor 5 for executing an evaluation value accumulating process described later, a control program for the evaluation value accumulating process, a program memory 6 storing an evaluation value table shown in FIG. 6, and an evaluation value shown in FIG. It consists of a two-vot memory 8 storing a cumulative table. The two-voted memory 8 and the processor 5 are connected by two sets of address buses 12a, 12b and data buses 13a, 13b of port a and port b. As a result, the 2-port memory 8 can perform the read operation from the port a and the write operation to the port b in parallel with different address designations. The processor 5 of the evaluation value accumulation processing means 3 and the processor 5 of the selection processing means 2 are connected by a connection bus 10.
With the above configuration, the evaluation value accumulation processing means 3 realizes the evaluation value accumulation processing function of accumulating the evaluation values for each category according to the category name read from the dictionary memory 1 for each vector element by the selection processing means 2.

The sort processing means 4 includes a processor 5 for executing sort processing described later, a program memory 6 storing a control program for the sort processing, and a general-purpose memory 7 storing a candidate category table shown in FIG. The processor 5 of the sort processing means 4 and the processor 5 of the evaluation value accumulation processing means 3 are connected by a connection bus 10. The contents of the candidate category table in the general-purpose memory 7 can be output to the outside via the processor 5 and the output bus 11. With the above configuration, the sort processing means 4 realizes the function of extracting the candidate categories based on the magnitude relation of the cumulative value of the evaluation values for each category output from the evaluation value cumulative processing means 3.

Next, the principle and procedure of the pattern matching process in this embodiment will be described.

FIG. 9 is a principle diagram of the selection processing and the evaluation value accumulation processing in the first embodiment of the present invention. In this figure, for a dictionary of sorted standard feature vectors consisting of the dictionary table of FIG. 2 and the dictionary management table of FIG. 3, a histogram of the number of categories for each element value size from 0 to L in each vector element is shown. , And the relationship between these histograms and the element values of the input feature vector, the selection range, the non-selection range, and the evaluation value is shown. In FIG. 9, only element 1, element 2, and element m among the m vector elements are shown as a representative. The distribution of the number of categories shown by the histogram is an example, and does not show the state peculiar to this embodiment.

First, in the selection process, for example, as shown in the histogram of element 1 in FIG. 9, the element value ix of the input feature vector
The element value of vector 1 whose lower limit corresponds to ix−Δ and whose upper limit corresponds to ix + Δ according to (indicate any of i ₁ , i _2, ... Im)
Calculate j _lx and j _hx, and range j _{lx ≤j x ≤j hx}
Is the selected range, and the range of the element value larger than this is the non-selected range. Here, Δ is a value determined according to the variation range of the element value of the feature vector and the like. In the selection processing means 2 of FIG. 1, category names included in the selection range are read from the dictionary memory 1 for each vector element. In the dictionary table of FIG. 2, the category names are arranged in order of magnitude of element values for each vector element. Since they are sorted, the dictionary management table of FIG. 3 is referenced for each element value j _x included in the selection range, and the start address of the category name in the dictionary table is referenced.
By determining Aj _x and the number of readout categories Nj _x ,
The category name can be read only by accessing the required part of the dictionary table. The outline flow of the above selection process is
Shown in Figure 10.

Next, in the evaluation value accumulation process, as shown in the principle diagram of FIG. 9, the category included in the selected range is associated with the evaluation value e, and the category included in the non-selected range is associated with “0”. Generally, a positive or negative numerical value other than "0" can be used as the evaluation value e, but in this embodiment, e is "1" and the selection range is "1", as shown in the evaluation value table of FIG. A binary evaluation value with a non-selected range of "0" is used. In the evaluation value accumulation processing means 3 shown in FIG. _1, from V ₁ of the evaluation value accumulation table shown in FIG.
"0" is written in V _n in advance, and the selection processing means 2
This table is accessed with the category name of the selection range given by and the evaluation value "1" is accumulated. For the categories included in the non-selected range, since the evaluation value is "0",
The evaluation value accumulating process can be omitted. FIG. 11 shows a schematic flow of the above evaluation value accumulation processing. Note that the category name reading in the selection processing and the evaluation value accumulation in the evaluation value accumulation processing can be executed by pipeline processing.

In the next sort processing, in the sort processing means 4 of FIG. 1, the category names corresponding to the evaluation value cumulative values for all categories given from the evaluation value cumulative processing means 3 are assigned the eighth category in descending order of the evaluation value cumulative value. The process of rearranging to the candidate category table in the figure is performed. When the rearrangement is completed, the pattern matching process for one input pattern is completed by outputting the candidate categories from the first candidate category having the largest evaluation value cumulative value to the predetermined rank. FIG. 12 shows a schematic flow of the above sort processing.

In this embodiment, when the category name is read out from the dictionary table of FIG. 2 stored in the dictionary memory 1 for each vector element in the selection process, each element value j _x included in the selection range by the selection processing unit 2 is read. The structure for determining the start address Aj _x and the number of categories Nj _x corresponding to the element value j _x in the dictionary table in FIG. 2 is shown with reference to the dictionary management table in FIG. In addition to this configuration, the start address Si _x (Aj _lx ) of the smallest element value j _lx in the selection range corresponding to the size of the element value i _{x and} the number of categories Hi _x (each Nj _x (Cumulative value) and a selection range table shown in FIG. 13 is created, which is stored in the dictionary memory 1 in place of the dictionary management table shown in FIG. It is also possible to determine the start address and the number of read categories in the dictionary table of FIG. 2 by referring to the selection range table of FIG.

Also, in this first embodiment, the selection is showing a structure for defining with a component value j _lx, j _hx standard feature vector determined by the width Δ of element values i _x and element values of the input feature vector. In addition to such a configuration, a configuration defined by the number of categories included between j _lx and j _hx is also possible.

In the pattern matching device of the first embodiment described above, the evaluation corresponding to the inter-vector distance between the input feature vector and the standard feature vectors of all categories is performed by cumulative calculation for the number of categories included in the selection range of each vector element. Since the value cumulative value is obtained, the processing can be executed at a higher speed than the conventional pattern matching processing for calculating the inter-vector distance by brute force.

Next, a second embodiment of the pattern matching device of the present invention will be described.

The pattern collating apparatus of the second embodiment has the block configuration shown in FIG. 1 as in the first embodiment, but the selection processing means 2 selects and reads from the dictionary memory 1 for each vector element. The dictionary memory 1 and the selection processing means 2 are configured so that the selection range of names is determined based on the relationship between the magnitude of the element value of the input feature vector and the standard deviation of the element value of each vector element obtained in advance. It is characterized by that.

FIG. 14 is a principle diagram of the selection processing and evaluation value accumulation processing in the second embodiment of the pattern matching apparatus of the present invention. Similar to FIG. 9 described above, this figure is a histogram of the number of categories by element value size for the xth vector element of the dictionary table of FIG. 2 in which category names are sorted in order of element value size. And the element value of the input feature vector, the selection range, and the non-selection range. In this figure, according to the relationship with the standard deviation of the element values, the case where the element value i _x of the input feature vector is large, the selection range is wide, and the value i _x is small, the selection range is narrow is illustrated in the same drawing. are doing. As an example of this, in the handwritten Chinese character recognition, the element value of the standard feature vector of each category uses the average value of many samples of the same character type, but the larger the average value of the element value, the more the variation of the element value of each sample. That is, the deviation tends to increase.

FIG. 15 is a diagram showing the relationship between the magnitude of the element value and the standard deviation for the x-th vector element of the standard feature vectors of all categories, which is approximated by a one-dimensional function. The relationship in this figure is that not only the categories of element values j _{x in} the vicinity of i _x but also the categories of element values i _x of the input feature vector are large in FIG.
This corresponds to the possibility that the category of the element value j _x distant from i _x may indicate the value of i _x due to handwriting fluctuations. Therefore, in the second embodiment, Δ _lx and Δ _hx that determine the lower limit and the upper limit of the selection range are standard deviations δ (j _lx ) and δ (of the element values j _lx and j _hx of the standard feature vector of the lower limit and the upper limit, respectively. j _hx )
As a of alpha _{times, j lx = i x -Δ lx} , j hx = i x + Δ hx Δ lx = α × σ (j lx), showed _{Δ hx = α × σ (j} hx) in FIG. 15 From the approximate expression, σ (j _lx ) = k _x × j _lx ＋ t _x σ (j _hx ) ＝ k _x × j _hx ＋ t _{x For the} element values j _lx and j _hx , the variable i _x and the coefficients α, k _x , t Arranging by _x , j _lx = (i _x −α × t _x ) / (1 + α × k _x ) j _hx = (i _x + α × t _x ) / (1-α × k _x ).

In this second embodiment, the standard deviation calculation table shown in FIG. 16 is created by obtaining the coefficients k _x and t _x for all vector elements in advance when creating the standard feature vector, and this is created as a general-purpose memory of the selection processing means 2. It is stored in 7. Also α
The value of is incorporated in the control program for the selection process. In the selection processing, the coefficients k _x and t _x are obtained by referring to this standard deviation calculation table for each vector element, and j is calculated from the above equation according to the value of α incorporated in the control program of the selection processing.
_lx and j _hx are calculated, and for each element value j _{x in} the range of j _lx ≦ j _x ≦ j _hx , the dictionary management table of FIG. 3 is referred to, and the start address A _jx in the dictionary table of FIG. And the number of read categories N _jx are determined, the category name is read only by accessing a necessary portion of the dictionary table.

In the second embodiment, the evaluation value accumulation processing and the sorting processing other than the selection processing are the same as those in the first embodiment described above.

In the second embodiment, as in the case of the first embodiment, the minimum element value j _lx of the selection range corresponding to the magnitude of the element value i _x is set in advance by the procedure shown in the selection processing above. category number Hi _x contained in all of the selection head address Si _x create a selection table shown in FIG. 13 is calculated and stored in place it in the third diagram of the dictionary management table of the dictionary memory 1 In the selection process, the start address and the number of read categories in the dictionary table of FIG. 2 can be determined by referencing the selection range table once for each vector element. As for the processing procedure for reflecting the standard deviation in the selection range, in addition to the processing procedure using the one-dimensional approximation formula shown in this embodiment, the standard deviation is set in correspondence with the magnitude of the element value of each vector element of the standard feature vector. It is possible to change to a processing procedure using the table shown, a processing procedure using a statistic similar to the standard deviation, and the like.

In the pattern matching device of the second embodiment described above, the selection range is determined according to the relationship between the magnitude of the element value of the input feature vector and the standard deviation of the element value of the standard feature vector. Even if the element values of the feature vector fluctuate unevenly, the pattern matching process can be executed with high accuracy and at high speed.

Next, a third embodiment of the pattern matching device of the present invention will be described.

The pattern matching device of the third embodiment has the block configuration shown in FIG. 1 as in the first embodiment, but the evaluation value accumulating means 3 accumulates evaluation values for each category for each vector element. The dictionary memory 1, the selection processing means 2, and the evaluation value accumulating processing means 3 are configured so as to make the determination based on the relationship between the element value of the input feature vector and the element value of the standard feature vector of each category. There is. FIG. 17 is a principle diagram of the selection processing and evaluation value accumulation processing in the third embodiment of the pattern matching apparatus of the present invention. Similar to FIGS. 9 and 14 described above, this figure shows the histogram of the number of categories by element value size and the elements of the input feature vector for the x-th vector element of the dictionary table of FIG. It shows the relationship between the value, the selection range, and the non-selection range. As shown in this figure, in the pattern matching device of this embodiment, the selection range is defined by a region of the element value j _x of the standard feature vector similar to the element value i _x of the input feature vector and a slightly similar element value j _x . Areas are set and e1 is accumulated as an evaluation value for category names included in similar areas, and e2 is accumulated as an evaluation value for category names included in slightly similar areas. The values of e1 and e2 are set to "2" and "1" in the third embodiment as shown in FIG. Since the evaluation value of the non-selected range is "0", cumulative calculation can be omitted. Therefore, in this invention,
The case where a three-valued evaluation value is used is illustrated.

In the selection processing means 2 of the third embodiment, the element values j _lx 1 and j _hx 1 and j _lx 2 and j _hx 2 indicating the limits of these areas are set in the same manner as in the second embodiment described above. Then, it is calculated from the coefficients α ₁ and α ₂ which are in the relationship of α ₁ <α ₂ . The evaluation value accumulation processing means 3 accumulates the evaluation values shown in FIG. 18 according to the area information and the category name given from the selection processing means 2.

The configuration and processing procedure other than the selection processing and the evaluation value accumulation processing are the same as those in the first and second embodiments described above.

In the third embodiment, as in the case of the first and second embodiments, the dictionary management table of FIG. 3 is replaced with the selection range table of FIG. 13 for the areas within the selection range. The configuration is possible. Also, in this embodiment, "2",
Although the configuration using three evaluation values of “1” and “0” is shown, the configuration using different evaluation values such as using “4”, “1”, and “0” obtained by squaring these values, or It is possible to change to a configuration using evaluation values of three or more values.

As described above, in the pattern matching apparatus of the third embodiment, the evaluation value is determined according to the relationship between the element value of the input feature vector for each vector element and the element value of the standard feature vector of each category. Even if the number of feature vectors is small (the number of dimensions), the pattern matching process can be executed with high accuracy and at high speed.

Next, a fourth embodiment of the pattern matching device of the present invention will be described.

The pattern matching device of the fourth embodiment has the block configuration shown in FIG. 1 as in the first embodiment, but has a category name having an element value of the standard feature vector similar to the element value of the input feature vector. Depending on the number of categories to be selected, a positive selection mode in which is selected, and a negative selection mode in which the category name of the element value of the standard feature vector that is not similar to the element value of the input feature vector is selected Either the positive selection mode or the negative selection mode is designated to configure the dictionary memory 1 and the selection processing means 2 so as to select and read a part of the category names from the dictionary memory 1. It is characterized in that the evaluation value accumulating processing means 3 is configured to accumulate the evaluation values given according to the case and the negative selection mode.

FIG. 19 is a principle diagram of the selection process and the evaluation value accumulation process in the pattern matching device according to the fourth embodiment of the present invention.
Similar to FIGS. 9 and 14 and 17 described above, this figure shows a histogram of the number of categories by element value size for the x-th vector element in the dictionary table of FIG. The relationship between the element value of the feature vector, the selection range, and the non-selection range is shown. As shown in this figure, in the pattern matching apparatus of this embodiment, the positive selection mode in which the category name having the element value of the standard feature vector similar to the element value of the input feature vector is selected and the element of the input feature vector are selected. There is provided a negative selection mode in which the category name of the element value of the standard feature vector that is not similar to the value is selected. One of these modes is designated for each vector element in the selection process, but both modes are shown in the same drawing in FIG. 19 for convenience. In this embodiment, the evaluation values accumulated in the positive selection mode and the negative selection mode have positive and negative polarities different from each other.
It shows the case where and -e are used.

In the pattern matching device of the fourth embodiment, similar to the first and second embodiments described above in the selection process,
When the element values j _lx and j _hx for _{discriminating the} selected range from the non-selected range are calculated by the selection processing means 2, first, assuming the positive selection mode, refer to the dictionary management table of FIG. 3 to refer to the dictionary of FIG. The start address and the number of read categories in the table are determined. Here, the selection processing means 2 determines whether the number of read categories is 1/2 or less of the total number of categories. If it is 1/2 or less, it is a positive selection mode, and if it is 1/2 or more, it is a negative selection mode. Take In the positive selection mode, the range determined by the previously determined start address and the number of read categories is set as the selection range, and the category names included in this selection range are read from the dictionary memory 1 to specify the positive selection mode. At the same time, it is transferred to the evaluation value accumulation processing means 3. In the evaluation value accumulating processing means 3, the evaluation value e for positive selection in the evaluation value table shown in FIG. 20 incorporated in the control program for the evaluation value accumulating process is designated, and the evaluation shown in FIG. 7 is performed according to the given category name. The evaluation value e is accumulated in the value accumulation table. On the other hand, in the case of the negative selection mode, the selection processing means 2 refers to the dictionary management table of FIG. 3 again, and the head address in the dictionary table of FIG. 2 is read out with the selection range other than the selection range of the positive selection mode as the selection range. Determine the number of categories. Next, the category name included in this selection range is read from the dictionary memory 1 and transferred to the evaluation value accumulation processing means 3 together with the positive / negative flag designating the negative selection mode. In the evaluation value accumulating processing means 3, the first value incorporated in the control program of the evaluation value accumulating processing is used.
The evaluation value −e for negative selection in the evaluation value table shown in FIG. 20 is designated, and this evaluation value −e is accumulated in the evaluation value accumulation table in FIG. 7 according to the given category name. As a result, even in the negative selection mode, it is possible to execute a process equivalent to accumulating the evaluation value e in the positive selection mode. In the fourth embodiment, the value of e is set to "1".

The configuration and processing procedure other than the selection processing and the evaluation value accumulation processing described above are the same as those in the first and second embodiments described above.

In the fourth embodiment, as in the case of the first and second embodiments, the dictionary management table of FIG. 3 is replaced with the positive / negative flag F _{ix in} the selection range table of FIG. It can be configured to use the selection range table of FIG. Further, in this embodiment, "1" and "0" in the positive selection mode and the negative selection mode,
By using a binary evaluation value consisting of "1" and "-1", the evaluation value accumulating process accumulates the evaluation values "1" and "-1" which are different in positive and negative polarities in the positive selection mode and the negative selection mode. Although the configuration is shown, a configuration in which these numerical values use different evaluation values,
In the positive selection mode, it is possible to use a configuration in which two or more evaluation values similar to those in the third embodiment are used.

In the pattern matching device of the fourth embodiment described above, since the evaluation value cumulative calculation can be performed by designating the mode in which the number of categories included in the selection range is small in the positive selection mode and the negative selection mode, most standard feature vectors are calculated. The high-speed pattern matching apparatus can be executed when a feature vector in which the magnitude of the element value is concentrated on a specific value is used, such as when the element value of 0 is zero.

In each of the above-described first to fourth embodiments, the device configuration of FIG. 1 is illustrated, but in this device, the dictionary memory 1, the selection processing means 2, the evaluation value accumulation processing means 3, the sorting processing means 4 are used.
In the processor 5 and the various memories 6 to 8 constituting the
Commercially available LSI etc. can be used. Regarding the function sharing among the respective means, the connection relation, and the block configuration of each means, the processor 5 of the selection processing means 2 and the processor 5 of the evaluation value accumulation processing means 3 are shared in addition to the configuration of FIG. Or the two-vote memory 8 of the evaluation value accumulation processing means 3
Various modified configurations are possible, such as a general-purpose memory.

Further, in the first to fourth embodiments, individual pattern matching devices are illustrated, but it is also possible to configure the first pattern matching device by combining the configurations of the second, third and fourth embodiments, for example. Is.

〔The invention's effect〕

As described above, according to the pattern matching apparatus of the present invention, a dictionary formed by sorting category names in order of magnitude of element values is used for each vector element of the standard feature vector,
By performing cumulative calculation of evaluation values only for some category names selected for each vector element, the evaluation value cumulative value corresponding to the inter-vector distance between the input feature vector and the standard feature vectors of all categories can be obtained. Therefore, the pattern matching process can be executed faster than the conventional pattern matching device that calculates the inter-vector distance by performing cumulative calculation for all vector elements between the input feature vector and the standard feature vectors of all categories.

Further, the pattern matching device according to the present invention can be applied to various character recognition devices, image recognition devices, voice recognition devices, etc. as well as handwritten Chinese character OCR as long as it is a pattern matching process using a feature vector. Can be greatly improved.

[Brief description of drawings]

FIG. 1 is a diagram showing the device configuration of an embodiment of the present invention, and FIGS. 2 to 8, FIG. 13, FIG. 16, FIG. 18, FIG. 20, and FIG. 21 are the devices of FIG. Configuration diagram of the table used, Fig. 9, Fig.
FIGS. 14, 17, and 19 are principle diagrams for explaining the processing in the apparatus of FIG. 1, FIGS. 10, 11, and 12 are flowcharts of the processing in the apparatus of FIG. 1, and FIG. It is a graph explaining a standard deviation. In the figure, 1 is a dictionary memory, 2 is selection processing means, 3 is evaluation value accumulation processing means, 4 is sorting processing means, 5 is a processor,
6 is a program memory, 7 is a general-purpose memory, 8 is a 2 port memory, 9 is an input bus, 10 is a connection bus, 11 is an output bus,
12 is an address bus, 13 is a data bus, 14 is a program address line, and 15 is a program read line.

Claims (4)

(57) [Claims] 1. A pattern matching device for extracting a candidate category based on a magnitude relation of similarity between an input feature vector generated from an input pattern and a standard feature vector previously generated from a standard pattern corresponding to a category. A dictionary memory storing a dictionary in which category names are sorted in order of magnitude of element values for each vector element of the standard feature vector, and the dictionary for each vector element according to the magnitude of element values of the input feature vector Selection processing means for selecting and reading a part of category names from the memory,
The evaluation value accumulation processing means for accumulating the evaluation values for each category based on the category name read out for each vector element by the selection processing means, and the accumulation value of the evaluation values for each category output from the evaluation value accumulation processing means. A pattern matching device, comprising: a sort processing means for extracting a candidate category based on a magnitude relationship. 2. The selection range of the category name selected and read from the dictionary memory for each vector element in the selection processing means is the magnitude of the element value of the input feature vector and the standard deviation of the element value of each vector element obtained in advance. The pattern matching device according to claim (1), wherein the dictionary memory and the selection processing means are configured so as to be determined based on the relationship with. 3. The evaluation value accumulating means determines the evaluation value accumulated for each category based on the relationship between the element value of the input feature vector of each vector element and the element value of the standard feature vector of each category. The pattern collating apparatus according to claim 1, further comprising a dictionary memory, a selection processing means, and an evaluation value accumulation processing means. 4. A positive selection mode in which a category name having an element value of a standard feature vector similar to an element value of an input feature vector is selected, and an element of a standard feature vector not similar to an element value of the input feature vector. A negative selection mode in which the category name of the value is selected is provided, and either the positive selection mode or the negative selection mode is designated according to the number of categories to be selected, and some categories are selected from the dictionary memory. The dictionary memory and the selection processing means are configured to select and read a name, and the evaluation value accumulating processing means is configured to accumulate the evaluation values given according to the case of the positive selection mode and the case of the negative selection mode. The pattern matching device according to claim (1), characterized in that