JPH0271381A - Sorting processor - Google Patents

Abstract

PURPOSE:To select the desired number of higher rank categories that have small degrees of difference at recognition of a pattern at a high speed by carrying out a necessary arithmetic process and the selection of the higher rank categories at one time and in parallel with each other. CONSTITUTION:The correct answer value corresponding to the features and the feature value extracted out of an input pattern are added to an arithmetic circuit 13 from a dictionary memory 11 and a feature memory 12 respectively. Then a degree of difference between categories is obtained via a prescribed arithmetic. The maximum value, i.e., the initial value of a maximum value memory 15 is detected by a maximum value detecting circuit 16. A comparator 17 compares the obtained maximum value with the arithmetic result of the circuit 13. If the degree of difference, i.e., the arithmetic result is smaller than the maximum value, the set value of the memory 15 is replaced together with replacement of the category number of a sorting result memory 19. As a result, an arithmetic process to obtain the degree of difference and the selection of a higher rank category are performed at one time and in parallel with each other. Thus the higher rank category is selected at a high speed. The processing speed is more increased by multiplexing a sorting process function part having a small hardware quantity.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a classification processing device, and particularly to a classification processing device for performing turn recognition processing.

[Conventional technology]

Generally, when performing classification processing in P-turn recognition, the feature value (x) is compared with the dictionary content (y) of the valley category, and the degree of dissimilarity Σf(x, y) for each valley category is computed. After the calculations for all categories are completed, a search is made for the degree of dissimilarity corresponding to these full-scale decorations. Then, the first step is to find the category with the smallest degree of dissimilarity.
category. Furthermore, the degree of dissimilarity is searched for the categories excluding the first place, and the category with the smallest degree of difference is set as the category of the second place. By repeating this series of processes in order, the top categories were selected.

[Problem to be solved by the invention]

However, classification processing using such a method has a problem in that it takes a long processing time when there are many target categories. For example, processing kanji requires an extremely large amount of processing time because there are thousands of characters in different categories.

The present invention focuses on this point, and provides a classification processing device that can quickly select t higher-rank categories from among a large number of categories.

[Means to solve problems]

According to the present invention, the features 1ift(x) extracted from among the categories and the dictionary contents (y) of each category are
The degree of dissimilarity Σf (x ry ) k is calculated by comparing the values.

In the classification processing circuit that selects -L category I/t with small dissimilarity, the dissimilarity Σf(x
=y) Multiple calculation units that perform all calculations and t upper power outputs after one classification (/maximum value memory that stores those calculation results, and among the calculation results stored in these maximum value memories) A maximum value detection circuit prepared to find the maximum value in , a classification result memory that stores the names of each of the upper categories, and a maximum value detection circuit prepared to find the maximum value in The high-speed classification processing is characterized by being equipped with a counter circuit that performs at-to-V-tuning, and a multi-stage control circuit that repeats the N-stage processing from 1st time to the Nth time, dividing the total number of times into 2 classification processing times. A device is obtained.

〔Example〕

The following is a detailed description of the invention with reference to one drawing.

FIG. 1 is a block diagram of an embodiment of the present invention. This classification processing device is composed of the following circuit parts: ■ A dictionary memory 110 that stores the correct answers corresponding to each feature; ■ A feature memory 12 that stores all the features extracted from manual input. ■ A two-system arithmetic circuit 13-1, 13-2 that performs distance meter jIj: by inputting all the outputs 101A and 10IB of the dictionary memory 11 and the output 102 of the feature memory 12 ■ A counter circuit that simultaneously performs addressing of various memories 14゜■ Output 103A of two arithmetic circuits 13-1 and 13-2
, 103B, the maximum value of the two systems is manually stored, and the maximum value is stored. Direct memory 15-1.15-2゜■ Output 104A of maximum value memory 15-1.15-2,
Two systems of maximum 1 direct detection circuit 16-1.16-2゜■ Output 105A of maximum value detection circuit 16-1.16-2 which inputs the corresponding one among 104B and detects the maximum value
, 105B and the corresponding arithmetic circuit 13-1.1
A two-system comparison circuit 17 that compares all corresponding outputs 103A and 103B of 3-2 and determines the category.
-1.17-2゜■ Output 107A of comparison circuit 17-1.17-2.

107B etc. are manually processed and classified by V-1 classification processing equipment.
ζ A Control circuit 18 for controlling the umbrella body Classification result memory 19 for storing classification processing results.

■ Classification processing is divided into two stages: 1IB and 2nd stage 2
Step control circuit 20゜02 Output of 1IB as output of step control circuit 20 and 2
Note that in this embodiment, the switching circuit 21 changes the output from the arithmetic circuit 13. Maximum value memory 15
Although we have arranged two systems, as long as there are multiple systems, it is possible to arrange three or more systems.

Next, let's talk about the classification processing circuit with such a configuration.

Each circuit portion will be specifically explained.

(Arithmetic circuit) The arithmetic circuit 13 calculates the features extracted from the human cover pattern,
This is the part that calculates all the degrees of difference from the dictionary for each category corresponding to each feature, and the distance is calculated using the linear equation (1).

where n; number of features X; Features (direct y; Dictionary output (direct It is.

(Maximum (direct memory, maximum 1 direct detection circuit)) Originally, in one classification process, this d1 calculation of Σf(xyy) is executed for all categories. If we simply calculate the distance a1 for all categories and select the top candidates from the distance results based on the full power, it would require a huge amount of processing time. become.

The present invention is intended to shorten the nine-classification processing time, and in this regard, the maximum value memory 15 and maximum value detection circuit 16, which are most important in this embodiment, will be explained in detail.

In this embodiment, a large shift is written into the maximum value memory 15 as an initial value before classification processing. Then, each time the distance is calculated for each category, the maximum value memory 15 (i
It is detected, and the comparison circuit 17 compares the maximum value with the distance result of one category. As a result, if the distance is greater than the first maximum value detection, the distance calculation for the next category is performed.

Conversely, if the distance 1st shift is smaller than the maximum value detected value, the maximum value in the maximum 1st distance memory 15 in the 9 classification result memory 19 is replaced with the maximum value in the maximum value memory 15 in the 9 classification result memory 19. Swap the category number corresponding to the distance with the current category number. This process 'f: is performed for each category and the maximum value memory 15 is updated in order. If you do this.

After calculating the distance of the final force lever, the content of the maximum value memory 15 is finally the distance meter nosebleed of the upper category, that is, the difference degree Σf(x+y).

The category number of the top candidate will remain in the classification result memory 19. In other words, the group of categories with the lowest degree of dissimilarity is selected from a large number of categories.
The classification process will be completed.

(Counter circuit) In particular, in this embodiment, there are two: dictionary memory 11 and feature memory 1.
The counter circuit 14 for addressing the maximum value memory is arranged as shown in row 5. By sharing the same counter circuit 14 in this way,
At the same time as the distance calculation is being performed, the contents of the maximum 1-direction memory 15 can be searched to perform maximum (fi'c detection processing).This result.

The time required to rearrange the order of the categories becomes zero. In other words, the processing time required for the two-classification process is approximately equal to the time obtained by multiplying the processing time of Σf (x + y) by the number of categories in the distance calculation, thereby enabling high-speed classification processing.

Furthermore, in general, a dictionary requires a very large capacity compared to other circuit parts such as the feature memory 12.

The ratio of hardware will increase. Therefore, in this embodiment, the classification processing circuit, which has a smaller amount of hardware than the dictionary memory 11, is multiplexed into two, thereby making it possible to perform classification processing in parallel without increasing the overall hardware size. Due to this.

Classification processing will be further accelerated. Of course, one circuit may be multiplexed into three or more.

(Multi-stage control circuit) In addition, in this embodiment, the two-stage control circuit 19 is arranged, so that the nine-classification process can be repeated in several stages, including the first, second, and Nth times. . In other words, the first time is a big classification.

In other words, a small amount of features are compared at full speed, and the number of candidates is reduced to a fraction of the total number of candidates. In the second time, the candidates narrowed down in the first time are compared with all the entries in the dictionary, one step at a time, and the candidates are classified in detail. Furthermore, it becomes possible to carry out all the same narrowing down operations from the third time onward to the Nth time and output all the final classification results.

If the first, second, and Nth stages are efficiently distributed and processed, the time required for the nine classification process will be further shortened.

Next, the explanation will be supplemented by comparing the processing times for one classification with respect to the examples described above.

In the conventional processing circuit, the time T required for classification processing is as follows (2
) as shown in the formula.

T=nk t1+ tk tz −
(2) Here, n; number of features; number of categories to be categorized before classification, t; number of upper categories after classification, tl; calculation time for one feature, t2; distance be.

In this equation (2), nktl is the minimum necessary processing time and cannot be reduced. On the other hand, tkt2i
d: There is a possibility of decrease. Furthermore, when the number t of upper categories after classification increases to the number of categories before classification, tkt2 increases, and the processing time T increases. Therefore, in the present invention, as already explained, when the distance value search and all distance calculations are executed, they are performed at the same time so that tkt2-0 is reached. In other words, the operation Σf(x
.

y) time T2=Ω1. At time T3 during the processing of
- Execute tt2 and exit. The number of features is larger than the number t of upper categories after classification, and in the classification processing circuit of this embodiment, j,=j2. Therefore, in this embodiment, time T3
Since the actual one shift is considered to be O, the two-class processing time T
(/i is obtained from the following equation (3).

T = n kt 1 - (3
) As described above, according to the two embodiments, the classification processing time can be significantly shortened. Further, in the classification processing circuit of this embodiment, parallel processing is performed by dividing one large-capacity word -J and processing it in two multiplexed classification processing circuits. Generally, as a result of parallel processing being performed by two or more classification processing circuits, it is possible to further increase the processing time T to -L by more than twice.

〔Effect of the invention〕

As described above, in the present invention, the maximum value memory and the maximum direct conversion circuit can be used to perform arithmetic processing and high-level category selection processing at the same time, which has the effect of enabling high-speed classification processing. In addition, by dividing the nine classification processing from the first to the Nth time, and dividing the classification from the major classification to the subclassification, and distributing it over multiple times, there is an effect that the speed of the one classification processing can be further increased.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of an embodiment of the classification processing device of the present invention. [Explanation of symbols] 11...Dictionary memory, 12...Characteristic memory, 13.
...Arithmetic circuit, 14...Counter circuit, 15...Maximum value memory, 16...Maximum value detection circuit, 17...Comparison circuit. 18... Control circuit, 19... Classification result memory, 20
...Two-stage control circuit.

Claims (1)

[Claims] 1. Feature values (x
) and the dictionary content (y) of each category to calculate the degree of difference Σ
Find f(x, y) and choose the upper category with the smaller degree of difference as l
In the classification processing circuit for selecting the items, a plurality of calculation units that calculate the degree of dissimilarity Σf (x, y) for each category, and a maximum value memory that stores the calculation results for the l upper categories after classification; A maximum value detection circuit prepared to find the maximum among the calculation results stored in these maximum value memories, a classification result memory that stores the names of each of the l upper categories, and the maximum value memory. and a multi-stage control circuit that divides the classification process into several times and repeats the N-stage processing from the first time to the Nth time. classification processing device.