JPS62280986A - Pattern identifying device - Google Patents

Abstract

PURPOSE:To improve the accuracy of identification even if an input pattern deviates from its normal position by obtaining projection curves in a direction splited uniformly with 360 deg. to the input pattern and obtaining input information on Perceptron through the extraction of the peaks of the curves. CONSTITUTION:Two types of data stored through general processing by a general processing part 12 and a peak number detection part 14 and through local processing by a local processing part 13, a correlation calculation part 16 and a partial feature extraction filter group 17 are inputted to the Perceptron. As a result an input pattern group is identified and learned, and a newly inputted pattern can be identified to which pattern groups stored it belongs. The titled identifier can efficiently be applicable to the storage and identification of various product patterns.

Description

Detailed Description of the Invention 3. Detailed Description of the Invention [Field of Industrial Application] The present invention relates to a pattern identification device applied to pattern storage and identification of various products, and particularly relates to a pattern identification device that is applied to pattern storage and identification of various products. The present invention relates to a pattern identification device that is required to store patterns and identify new input patterns to the stored patterns.

[Prior Art] Conventionally, a representative document regarding this type of pattern recognition device is Omori-Nakano's “Visual System Model Containing Self-Organizing Cells” (Proceedings of the Society of Instrument and Control Engineers).

VOI 20. No, 1. p5B-69). In this model, by differentiating the input pattern and detecting its peak, points where the shape of the input pattern changes drastically are found as feature points, and a predetermined filter is applied to those feature points to extract features in the vicinity. By repeatedly showing a plurality of input patterns, a filter is formed that extracts various representative feature patterns included in the input pattern group.

Therefore, when one pattern is presented to the model for the above input pattern group, the correlation between the neighborhood of each feature point, the filter formed as shown above, and the existing filter is calculated. , are added to the storage device corresponding to each filter, and the same processing is performed on all feature points. By inputting the data in the storage device obtained as a result to the perceptron, a group of input patterns can be identified and learned.

With the above steps, the pattern identification device is completed, and it is possible to identify which of the already stored input pattern groups a new input pattern belongs to. To summarize the above, first, a filter is learned to extract features included in a group of input patterns, and second, the features are extracted using the filter for each pattern of the input pattern group that is repeatedly presented, and By using the results as input to the perceptron, it becomes possible to perform discriminative learning on a group of input patterns.

[Problems to be Solved by the Invention However, this type of device has the following problems. That is, ■ The main input data of the perceptron is only the statistics of the amount of correlation between the local feature filter and the feature points, and the pattern identification ability is poor because information that constrains the relative positional relationship of the local features is not input. .

■ Since feature extraction filters are formed through learning, redundancy occurs between them and existing filters, and as a result, an optimal filter cannot be formed.

The present invention has been made in consideration of the above circumstances, and its purpose is to store patterns of various shapes to be identified, to be able to identify new human cover patterns to the stored patterns, and to be able to input It is an object of the present invention to provide a pattern identification device that can obtain high identification ability regardless of positional deviation of a pattern.

[Means for solving problems] The main points of the present invention are as follows.

(1) As information that constrains the relative positional relationship of local features, the projection curves from the directions equally divided into 360′ of the input pattern are obtained, and by extracting the peaks of these curves, the lines in that direction are calculated. Find out how many minutes there are in the input pattern, and use that information as the human power of Bersebtron.

(2) In order to reduce the redundancy of local feature extraction filters, all filters prepared in advance are used. already,(
Since global line segment elements are extracted in step 1), the prepared filter has the property of detecting edges and endpoints.

That is, in the present invention, in a pattern identification device, projection curves from directions equally divided into 360° are obtained for the input pattern, and by detecting the peaks of these curves, the number of line segments in that direction is determined in the input pattern. By memorizing whether there is a pattern or not, and by differentiating human cover turns and detecting their peaks, we can find places where the shape of the input pattern changes rapidly as feature points, and use the information about those feature points that have already been prepared in the vicinity. By calculating the correlation with the filter group that exists, multiple filters with high correlation are selected, and the total number of filters selected for all feature points is memorized. By inputting type data into a perceptron, the perceptron learns to identify input pattern groups, and further identifies which of the previously stored input pattern groups a new input pattern belongs to.

[Function] With the above configuration, the rotation angle is 3130° with respect to the human cover turn.
Since the projection curves from the directions in which the image is equally divided are obtained, and the input information of the perceptron is obtained by extracting the peaks of these curves, it is possible to improve the discrimination ability even with respect to positional deviation of the input pattern. In addition, since all the local feature extraction filters necessary for identifying human cover patterns are prepared in advance, it is possible to reduce the redundancy of the filters.

[Example code] Hereinafter, details of the present invention will be explained with reference to illustrated embodiments.

FIG. 1 is a schematic configuration diagram showing a pattern recognition device according to an embodiment of the present invention. Reference numeral 11 in the figure is an image memory for storing input patterns, and the pattern data stored in this image memory 11 is supplied to a global processing section 12 and a local processing section 13, respectively.

In the global processing unit 12, 38 (1
The projection curves from the directions in which `` is equally divided are calculated, and the peak values of these curves are calculated.Then, among all the peak values, a particularly large contribution rate near the peak value is detected, and the corresponding For each projection curve, peaks with a height greater than or equal to the average height are determined.Then, the number of peaks determined is calculated by the global feature counter 15.
supplied to

On the other hand, the local processing section 13 detects feature points of the input pattern by differentiating the input pattern, and supplies these feature points to the correlation calculation section 16 . The correlation calculation unit 16 calculates the correlation between the input pattern information near the feature point and the filter group prepared in advance in the local feature extraction filter group 17, and selects a certain number of local feature extraction filters with high correlation for each feature point. be done. The selection information is then supplied to the local feature counter 18.

The global feature counter 15 and the local feature counter 18 constitute a first counter (layer A), and the stored information of this first counter is supplied to a second counter (layer R) 19 . The information stored in the second counter 19 is used as an input to the bersebutron.

Note that the local feature extraction filter group 17 is formed, for example, in a 5×5 matrix, as shown in FIG.

However, in the figure, the mark represents (+1) and the space represents (-1). Further, the global feature counter 15 and the local feature counter 1B are configured as shown in FIGS. 4(a) and 4(b), respectively. That is, the global counter 15 is formed of a number of counters corresponding to the N filters shown in FIG. 4(a). Regarding the P1 filter, if the P1 filter is selected 0 times in the input pattern, the O bit flag is set to 1. When the P1 filter is selected once in the input pattern, the 1-bit flag is set to 1.
Set to . Further, if the P1 filter is selected KP times or more in the input pattern, the KP bit flag is set to 1. The same applies to the P2 to PM filters. Further, the local feature counter 18 is formed of a number of counters corresponding to M filters shown in FIG. 4(b). Regarding the F1 filter, if the F1 filter is selected 0 times in the input pattern, the 0 bit flag is set to 1. When the F1 filter is selected once in the input pattern, the 1-bit flag is set to 1.
Set to . Further, if the F1 filter is selected KF times in the input pattern, the KF bit flag is set to 1. The same applies to the F2 to FM filters.

The operation of the apparatus configured as described above will be explained with reference to FIG. Now, assuming that a certain pattern is set as a digital image input, the following processing is performed.

Local processing is performed by the following steps (Ll) to (L5).

However, median fil is used to remove noise such as isolated points.
Apply direct ones such as ter.

(Ll) input pattern S stored in image memory 1;
Divide it into two aircraft, find the absolute value, and let this be Sl.

(L2) Detect all peaks using S. Input this as pattern S. are called feature points.

(L3) A correlation calculation is performed between the input pattern figure near the feature point of the input pattern So and the local feature extraction filter group 17 prepared in advance.

(L4) For each feature point, local feature extraction filter group 1
A certain number of highly correlated filters are selected from 7.

(L5) Calculate the total number of times each local feature extraction filter has been selected for all feature points, and set them in the local feature counter 8.

The global processing is performed in the following steps G1) to (G4).

(G1) For the input pattern So, obtain projection curves from directions equally divided into 360 degrees.

(G2) Find the peak value for each projection curve, and calculate the contribution rate P + of the vicinity of the peak value to the area surrounded by the projection curve.
Find (t-1 to N).

(G3) Using statistical data of the distribution of contribution rate P (i-1 to N) (mean μ1 variance σ2, etc.), a uniquely large P is found. For example, P, 〉μ〉1σ.

J
J However, σ>k2μ (k, , 2 is a positive constant)
.

(G4) P selected in (G3) is (P...
,P,)j jL' ,un, for the projection curve corresponding to jL,...,jLiu, find all the peaks above the average height of each, and calculate the number of peaks as the corresponding global feature. Set it on counter 15.

As described above, all the feature extraction results obtained by local processing and global processing are set in the feature counter.

The structure of the feature counter is such that a local feature counter 8 and a global feature counter 5 are installed in parallel, and this is named the A layer. Therefore, layer A has one vector α−
It is expressed as (al, . . . , aX), and a,=Oorl. There is an R layer for the A layer, which is also represented by one vector 7, -(zl,...,zy), and Z,=Oo
Takes the value of rl. The relationship between a and Z is Zi-1, (Σω,,・a,-h,)...
■j, given by 11j J 1. However, to determine ω, , h in the formula, the so-called bar set IJ is used.
Adopt J Ron's learning method. That is, for a given pattern group, if you want to set Zl-1 for a certain pattern and Zl-0 for the remaining patterns, input these patterns to this pattern recognition device and set γ- Y-Z engineering ・・・Calculate ■. However, Y is Y-1 for the input pattern that you want to set to Zl-1, and Y for the pattern that you want to set to Z1-0.
-0. Then, if γ is 0, there is no change; if γ-1, then a, and for j >Q, where C is a positive constant.

■Procedure j to change ω and h according to the formula
If l is repeated for all patterns, it becomes possible to distinguish a particular pattern from other patterns. If you want to separate and identify multiple patterns, use Zl, Z2
．． . . . It is sufficient to prepare as many input patterns as the number of R layer components from which you want to separate.

Thus, according to this embodiment, the global processing by the global processing unit 12 and the peak number detection g514, and the local processing by the local processing unit 1
3. By inputting the two types of data stored through local processing by the correlation calculation unit 16 and the local feature extraction filter group 17 into the Bersebutron, it learns to identify the human cover pattern group, and furthermore, the already stored patterns are used for new input patterns. You can identify which group it belongs to. Therefore, it can be effectively applied to pattern storage and identification of various products.

Note that the present invention is not limited to the embodiments described above, and can be implemented with various modifications without departing from the gist thereof. For example, the structure of the local feature extraction filter group is not limited to that shown in FIG. 3, but can be changed as appropriate depending on the pattern to be identified.

[Effects of the Invention] As described in detail above, according to the present invention, patterns of various shapes to be identified can be stored, and new input patterns can be identified as the stored patterns. In addition, global processing is performed by calculating projection curves from directions equally divided into 360 degrees for the input pattern, and local processing is performed based on local feature extraction filters prepared in advance. Effects can be obtained.

(1) In order to understand the feature values that are invariant to parallel movement,
Strong against displacement due to parallel movement.

(2) Even enlarged/reduced patterns can be recognized by rough normalization of the entire pattern.

[Brief explanation of the drawing]

FIG. 1 is a schematic configuration diagram showing a pattern identification device according to an embodiment of the present invention, FIG. 2 is a schematic diagram for explaining the operation of the device according to the embodiment, and FIG. FIG. 4 is a schematic diagram showing an example of a feature extraction filter, and FIG. 4 is a schematic diagram showing an example of global and local feature counters. 11... Image memory, 12... Global processing unit, 13
...Local processing unit, 14...Peak number detection unit, 15
...Global feature counter, 16...Correlation calculation unit, 1
7... Local feature extraction filter, 18... Local feature counter, 19... Counter. Applicant Sub-Agent Patent Attorney Takehiko Suzue Figure 1

Claims (1)

[Claims] Projection curves from directions equally divided into 360° are calculated for the input pattern, and by detecting the peaks of these curves, the number of line segments in that direction is stored in the input pattern, and the input pattern is By differentiating and detecting the peaks, we can find locations where the shape of the input pattern changes drastically as feature points, and then calculate the correlation between those feature points and the filter group that has already been prepared. , by selecting multiple highly correlated filters, memorizing the total number of filters selected for all feature points, and inputting the two types of data stored above into the perceptron. A pattern identification device characterized by learning to identify a group of input patterns and further identifying to which group of input patterns already stored a new input pattern belongs.