JPH0291779A - Information recognizing system - Google Patents

Abstract

PURPOSE:To include a correct recognized result and to more efficiently contract the number of candidates by generating the threshold to contract the candidates by the maximum value and the minimum value of the distance of a pattern inputted to a contraction function and a standard pattern. CONSTITUTION:The distance of respective standard vector information groups of the pattern information of input information extracted by an extracting means 2 is calculated by a calculating means and based on the minimum distance, the threshold is determined by a threshold determining part 8. Thereafter, the distance group calculated by the determined threshold is contracted and the candidate group based on the standard vector corresponding with selection outputting means 7 and 9 is outputted. Thus, when the information to be a recognizing object is recognized and a candidate is outputted, the correct recognized result can be included in a small number of output candidates with high probability. By generating the threshold contracting the candidates by the maximum value and the minimum value of the distance of the pattern inputted to the contraction function and the standard pattern, the correct recognized result is included and the contraction of the number of candidates is made more effective.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a pattern recognition method, and more particularly to an information recognition method that inputs recognition target information and outputs a corresponding group of candidates.

[Prior Art] Conventionally, a pattern recognition device has a configuration as shown in FIG.

In the figure, 51 is an input unit that converts input patterns such as audio signals, character images, graphic images, etc. into electrical signals, and 52 is a feature extraction unit that extracts features specific to the input pattern from the input signal and generates a feature vector. .

Note that the configurations of the input section 51 and the feature extraction section 52 are well-known techniques, so the details thereof will be omitted. Reference numeral 53 denotes a dictionary section in which standard values (standard vectors) serving as templates of patterns to be recognized are registered. The standard vector is made up of values such as the average value and standard deviation obtained by performing feature extraction in advance from a plurality of learning data according to the same algorithm as the feature extraction unit 52 and by statistical processing. 54 is a matching unit that compares the feature vector of the input pattern with the standard vector in the dictionary and calculates the distance to each standard pattern. 55 is the distance obtained by the matching unit 54 in order to select a standard vector similar to the input pattern. 56 is an output section that displays the output from the sorting section 55 on a display or transfers it to other devices. .

[Problem to be solved by the invention] Now, in the conventional example with the above-mentioned configuration, sorting is performed to select the pattern that is most similar to the input pattern, but in the case of sorting that performs complete ordering, the recognition target When the number of patterns is N, no matter what algorithm is used, a calculation amount of O(N logN) is required. Also, pattern 1 with the smallest distance
If you only select the pieces, the amount of calculation is 0(N), but
In reality, multiple patterns are required as candidates, so
(N-Nlog N) calculation (2) is required, and as N becomes large, processing time becomes extremely long, resulting in a disadvantage that the overall throughput is greatly reduced.

The present invention has been made in view of the related art, and aims to provide an information recognition method that makes it possible to include correct recognition results with a high probability among a small number of output candidates.

[Means for Solving the Problem] In order to solve this problem, the present invention includes the configuration shown below.

That is, in an information recognition method that inputs recognition target information and outputs a corresponding candidate group, there is an extraction means for extracting input information-specific features as input pattern information, and a storage means for storing a standard vector information group for the recognition target information. a calculating means for calculating a distance of the input pattern information to each of the standard vector information groups; a detecting means for detecting the calculated minimum distance; and a threshold determining means for determining a threshold based on the calculated minimum distance. , a narrowing means for narrowing down the distance group calculated by the calculating means based on the threshold determined by the determining means, and selecting a candidate group based on standard vector information corresponding to the distance group narrowed down by the narrowing means; and a selection output means for outputting the output.

[Operation] In the configuration of the present invention, the calculation means calculates the distance between the pattern information of the input information extracted by the extraction means and each standard vector information group, and the threshold value is determined based on the minimum distance. Thereafter, the calculated distance group is narrowed down using the determined threshold value, and the selection output means outputs a candidate group based on the corresponding standard vector.

[Embodiments] Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings.

<Description of the first embodiment (FIGS. 1 to 4)> FIG. 1 shows a block diagram of the pattern recognition apparatus of this embodiment.

The input section 1, feature extraction section 2, dictionary section 3 in which standard patterns are registered, and output section 9 in the figure are the same as 51, 52, 53, and 56, respectively, in the conventional example shown in FIG.

The specific processing contents of the collation unit 4 will be explained below.

The matching unit 4 combines the feature vector f, (j = 1.2..., NUN is the number of dimensions) obtained from the input pattern by the feature extraction unit 2 with a candidate character registered in advance in the dictionary as a standard pattern. standard vector mk, (j=t.

2,...,N=1.2. ..., M: M is the number of categories to be recognized) and the distance d (k) is calculated by the city block distance defined by the following formula.

d (k) = ΣlfJ mk, l ・
・'■wl This distance calculation is performed on the average vector of all recognition patterns, and the minimum value detection unit 5 calculates the obtained d(k)
The minimum value d1□ is selected from (k・1.2. . . , M). Specifically, the minimum value detection unit 5 unconditionally stores the first calculation result of equation 0 (・d(1)) in the minimum value storage unit 6 as the minimum value CI+in, and all k The smaller value is compared with d (k) according to the formula 0 and the smaller value is stored and updated in the minimum value storage unit 6. That is, the value stored in the minimum value storage unit 6 at the time when all the standard vectors have been compared becomes the target minimum value d1n.

In this way, when the minimum value d1° of the standard vector is determined, the threshold value determination unit 8 determines the threshold value dt for candidate selection based on the minimum value dw+In.

A method for determining the function γ(dln) for determining the threshold value dt will be described below.

FIG. 2 is a graph in which d(k) is plotted on the vertical axis and dmin is plotted on the horizontal axis. However, the distance to the correct answer (the real pattern represented by the input pattern) is indicated by a "○" mark, and the other distances are indicated by a "Δ" mark. still,
In the following explanation, the distance d from the correct answer marked with “○”
Let's call it c.

Now, since the distances between one input pattern and M standard patterns are determined, M points are plotted for one dmln, as in the area surrounded by the broken line 202 or 203. However, in the figure, points that fall outside the frame are not plotted.

Since dmln is the minimum value in d(k) (k=1.2.-...,M), all points (“○°°mark” or “△”
mark) is never below the straight line 201. If feature extraction and distance calculation were performed ideally, the distances dc and dmln for the correct answer would match, as in the region 202, but in actual recognition processing, they do not always match, and the distances in the region 203 d (i) for an incorrect pattern that is not correct may be dmln (distance de
and dln do not match). However, even in such a case, the distance d to the correct answer is usually not very large with respect to dmln.

Here, FIG. 3 shows the relationship between dc and d, 1°, which was obtained as a result of experimental analysis for many input patterns. still,
In the figure as well, d(i) for the correct answer (the real pattern represented by the input pattern) is shown with a "○" mark, and the others are shown with a "△" mark.

Therefore, when outputting the recognition results for input information (patterns) via the output unit 9, when determining how to reduce the number of candidates to be output, it is necessary to use a curve that includes "○" marks for many of these cases. It can be seen that it is sufficient to use this as the threshold value.

In this example, a function of the following form was selected in order to include almost all the points (marked with "O") representing the distance dc and to avoid including as many points representing other distances as possible.

y (dmln) = a, rgo τ−+b
... ■(a, b: constants) The curve 208 shown in the figure is shown by this formula ■, and the threshold value determination unit 8 calculates the threshold value dt(-γ(d
, to)).

Then, the candidate selection unit 7 follows the threshold value dt output from the threshold value determination unit 8 and satisfies d (k)<dt.
k) is selected and output from the output unit 9 as a candidate.

Note that if you change the values of constants a and b in equation (2) and raise the entire graph as shown in curve 207 in Figure 4, the probability of not being included in the candidates, that is, selecting the wrong candidate, decreases, but the average It is clear from the distribution in FIG. 3 that the number of candidates increases and the efficiency of narrowing down candidates decreases. On the other hand, curve 2
If you move the graph down like 08, you will see that the opposite trend occurs. In this way, it is possible to adjust the parameters a and b according to the application target of the apparatus, and to provide the optimum candidate selection rate and average number of candidates.

For example, the parameters a and b may be changed as appropriate depending on the type of information to be input (voice, text, graphic image, etc.) or the input environment (in the case of handwritten character input, the distinction of the person writing the information, etc.).

Note that the candidate group outputted via the output unit 9 is outputted to other devices including a display.

Description of the second embodiment> In the embodiment described above, dmin is used as a function to narrow down the candidates.
Although only dt was used, there is a possibility that most of d(k) (k·1...N) falls within the threshold dt. Therefore,
In the embodiment of this section 2, an example will be described in which even if such a situation occurs, narrowing down the search results more efficiently.

FIG. 6 is a block diagram showing the configuration of the second embodiment, in which a human power section 61, a feature extraction section 62, a dictionary section 63, a collation section 64, a candidate selection section 67, and an output section 69 correspond to those of the first embodiment. It has the same structure as the part.

The maximum value/minimum value detection unit 65 detects the maximum value dmaX and the minimum value dllll11 among the distances d (k) (k-1, 2, ..., M) from each standard vector (minimum value detection is As in the first embodiment, the maximum value is detected by reversing the principle, so the explanation will be omitted), and the maximum value is stored in the maximum value/minimum value storage section 66. In the threshold value determination unit 68, d□.

A threshold value dt for candidate selection is determined according to the following equation using the values of and dmax.

dt=d1. +(d,,, dmln)Xθ However,
θ is a real number that satisfies O×θ×1, and in this example, θ;0.
It is set as 1.

As described above, according to this embodiment, when recognizing information to be recognized and outputting candidates, it is possible to include correct recognition results with a high probability and narrow down the number of output candidates. becomes.

Moreover, since there is no double loop logic, high-speed processing is possible.

In addition, by generating threshold values based on the maximum values d, □ and the minimum value d+aln of the distance between the pattern input to the narrowing down function and the standard pattern, it is possible to include correct recognition results and to narrow down the number of candidates more efficiently. becomes possible.

Although the input information to be recognized in the above-mentioned first and second embodiments is not particularly limited, it can be applied to the recognition of speech, characters, graphics, etc. by using the principles of the present invention. It's for a reason.

In addition, the arithmetic expressions related to determining the threshold value are limited to the formulas ① and ② of the first and second embodiments described above, since the point is that any function that includes the correct recognition result is sufficient. It's not a thing.

Further, the present invention can be modified as appropriate within the scope of the claims, and the configuration is not limited to the present embodiment.

[Effects of the Invention] As explained above, according to the present invention, when recognizing information to be recognized and outputting its candidates, it is possible to include correct recognition results with a high probability in a small number of output candidates. becomes.

In particular, by generating thresholds for narrowing down candidates based on the maximum and minimum distances between the pattern input to the narrowing down function and the standard pattern, correct recognition results can be included and the number of candidates can be narrowed down more efficiently. becomes possible.

[Brief explanation of the drawing]

Fig. 1 is a block diagram of the first embodiment, Fig. 2 is a diagram showing an example of the distribution of distance between the input pattern and the standard pattern, and Fig. 3 is a diagram showing the distribution of distance and a function curve for determining the threshold value. Figure 4 is a diagram showing changes in the threshold function due to parameter changes, Figure 5 is a block diagram of a conventional information recognition device, and Figure 6 is a block diagram of a second embodiment. . extraction part, 3 and 63... dictionary part, 4 and 64... collation part, 5... minimum value detection part, 6... minimum value storage part,
7 and 67...Candidate selection unit, 8 and 68...Threshold value determination unit, 9 and 69...Output unit, 65...Minimum value/maximum value detection unit, 66...Minimum value/maximum value This is the storage section. In the figure, l and 61...input section, 2 and 62...characteristic drnln funeral map min figure

Claims (2)

[Claims] (1) An information recognition method that inputs recognition target information and outputs a corresponding candidate group, which includes an extraction means that extracts the input information-specific features as input pattern information, and a memory that stores a standard vector information group for the recognition target information. means, a calculating means for calculating a distance of the input pattern information to each of the standard vector information groups, a detecting means for detecting the calculated minimum distance, and a threshold determining means for determining a threshold based on the calculated minimum distance. a narrowing means for narrowing down the distance groups calculated by the calculating means based on the threshold determined by the determining means; and selecting a candidate group based on standard vector information corresponding to the distance group narrowed down by the narrowing means. , and a selection output means for outputting. (2) Furthermore, a second detection means for detecting the maximum distance of the distance group calculated by the calculation means is provided, and the threshold value determination means is configured to detect the minimum distance and the maximum distance detected by the detection means and the second detection means. 2. The information recognition method according to claim 1, wherein the threshold value is determined based on .