JPS61211783A - Image signal processing system - Google Patents

Abstract

PURPOSE:To obtain a strong and high speed pattern recognition with respect to a misregistration of a pattern by dividing an input image into small areas overlapping with one another or not overlapping and converting the input image into a pattern of a spatial frequency power spectrum or a self-correlation function. CONSTITUTION:A light and darkness of a pattern image information inputted from a television camera or an image scanner is represented by a numeric value every picture element and stored in an image memory. The inputted image is divided into small areas and respectively fed to processors exclusively used of them and converted into a pattern of a spatial frequency power spectrum every small area. Then the image is supplied to a pattern judging unit, subjected to a similar processing in advance, and a ten plate matching by using a standard pattern to perform a pattern judgement. Thereby, with respect to a misregistration of an indicated pattern, a strong and a high speed processing can be done by a deletion of a calculating amount.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to an image signal processing method applied to a pattern recognition device such as OCR.

(Prior art and its problems) The recognition methods of pattern recognition devices such as OCR can be broadly divided into analog matching methods (template matching methods) and structural analysis methods.

The former method recognizes a pattern by superimposing a standard pattern and an unknown pattern, and is said to be highly versatile and particularly useful for reading printed characters.

However, this method has the drawback that if there is a positional shift during superimposition or if the pattern is deformed such as rotation or expansion/contraction, the overlap with the standard pattern will be reduced and the pattern will not be recognized correctly. Conventionally, to deal with positional deviations, the recognition rate was improved by artificially blurring the standard pattern, or by moving the unknown pattern up, down, left and right, and searching for the position with the greatest overlap. It was improving.

However, although this method improves the recognition rate for small displacements in the position of unknown patterns, it is ineffective for recognizing patterns with large displacements, and therefore requires further alignment, etc. There was a drawback that pretreatment was required. Furthermore, the method of artificially adding blur has the disadvantage that spatial resolution decreases.

“The Proceedings of the Ieeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeee.

1970.58 Volume & l (“ProceedinJs o
EE from f the.

G, G, published in 197 (L vol s 8″)
Renderis゛('G, G, Lenaariθ) et al.''
The document “Diffraction-Pattern Sampling Flow Automation Pattern Recognition” (“Di
ffraction-Patl;ern Sample
inJ for Automatic Pattern
RecoJlnition'') describes a method in which a Franhofer diffraction image of a pattern is created using holography, and the pattern is recognized from the diffraction image.This method involves performing two-dimensional Fourier transformation on the pattern. , which corresponds to converting a pattern into a power spectrum of its spatial frequency, and since the power spectrum remains unchanged even if the original pattern is translated, it is a method that is robust against pattern displacement. However, this method has been difficult to put into practical use because the pattern to be presented needs to be a film or the like that transmits light.

(Object of the Invention) An object of the present invention is to eliminate such conventional drawbacks of the analog matching method and to provide a high-speed image signal processing method for pattern recognition that is resistant to pattern displacement.

('Structure of the Invention) The requirement to be strong against pattern displacement is to divide the input image into small regions that overlap or do not overlap each other, and to analyze the spatial frequency power spectrum or autocorrelation of the input image for each small region. This is satisfied by converting it into a function pattern. By checking the overlap using an analog matching method using a standard pattern that has been previously subjected to the same processing as the set of patterns obtained, high-speed pattern recognition that is resistant to positional deviations of patterns can be realized.

(Detailed Description of Configuration) FIG. 1 is a flowchart of a pattern recognition method using the method of the present invention. Pattern image information input from a television camera or image scanner is converted into numerical values for the brightness of each pixel and stored in an image memory. In the processing according to this method, the following conversion is performed on this image information using a processor or the like.

Now, assume that the number of pixels of the input image is NXN. First, this input image is divided into n2×n2 small regions each having n and xn pixels. If the images are to be divided so that there is no overlap, N-n.

xn, but they may be divided so that they overlap each other. Next, each image divided into these small regions is subjected to appropriate processing so that even if the presented image is translated, the transformed pattern remains unchanged. For example, since the spatial frequency power spectrum of an image remains unchanged even when the image is translated, processing to obtain the spatial frequency power spectrum may be performed. Alternatively, since the autocorrelation function of the image also remains unchanged with respect to the translational movement of the image, a process of determining the autocorrelation function may be used.

Subscript each small area (X, Y) (X, Y =
The position of a pixel in each small area is represented by a subscript (X).

7) Represented by (X, 7=1 to n,). Also, the value of the pixel at the position (x, y) in the small area represented by (X, Y) is expressed as (X, 3", XIY)1. Then, the spatial frequency power spectrum described above・Mother tongue is calculated by the following formula.

・・・・・・・・・ (1) P (R,, R2; X, Y)-F(R11R2:
X, ' Y) ” 1., (2) (1) Formula c, I (x
, 7 ; xt Y) 2 regarding the subscript (x+y)
Represents the discrete value-to-IJ transformation of the dimension. Regarding the calculation of equation (1), it is possible to use fast Fourier transform algorithm, and in that case, the processing time can be further reduced. P in equation (2) is the spatial frequency/9worth vector. Also, in the above process, when calculating the autocorrelation function, it is calculated by the following equation: 07+d, :X,
Y) If the effect at the boundary of the current season region can be ignored, P or C will be a nozotan that remains unchanged with respect to the translational movement of the input image/mother tongue, so by such processing, the input image A very stable pattern with respect to positional shift can be obtained (this condition is satisfied when the size of each pattern in the input image is sufficiently small compared to the size of the small area nt figure)). Therefore, by sending the mother tongue that has undergone such processing to the next pattern judgment unit and performing template matching with a standard pattern that has undergone similar processing in advance, it is possible to A mother tongue recognition method that is resistant to positional deviations (in terms of size) will be realized.

If the input image is relatively large compared to the size of the inner pattern or small area, the effect at the boundary cannot be ignored in the processing of each small area, so P(R11R29XI
Y) Strictly speaking, a certain IhaC(d+s d2 :X, Y) is no longer invariant with respect to the translational movement of the input/9 tongue. However, in this case as well, P or C has a stable pattern with respect to translational movement when compared with the input image itself.

The reason for this will be explained using FIG. 3 (the patterns that the present invention mainly targets are two-dimensional patterns;
For simplicity, we will explain using a one-dimensional signal, 9 tangs).

As is well known, when a moving average is taken over a certain width (nl) for a signal that changes drastically, the high frequency components of the signal are removed and the signal changes gradually. This method of using an appropriate weighting function to take the moving average of the mother tongue and then performing ten rate matching corresponds to the method of artificially blurring the mother tongue.
This method is resistant to misalignment.

The biggest difference between the method using such a moving average and the processing method according to the present invention is that the former discards high frequency components of the pattern in order to make it resistant to positional fluctuations, whereas the latter discards the high frequency components of the pattern. In this case, ・Stability regarding the positional fluctuation of the 9-tongue is achieved without discarding the information on the high-frequency components of the 9-tongue. The former simply calculates a scalar quantity called the average from an input signal with a width of n, whereas the latter calculates its spatial frequency power spectrum (or autocorrelation function) from the same input signal. This is because we are calculating a vector quantity called the pattern of ). Therefore, in the former case, the allowable size of positional deviation is in a trade-off relationship with resolution, so the allowable size of positional deviation is limited by the resolution, but in the latter case, these two Since the two quantities are independent, the allowable magnitude of positional deviation can be set freely.

As was achieved by Lendari et al. using holography, a method of using a processor or the like to obtain the spatial frequency power spectrum of the entire input image can also be used to obtain image data in the same way as the method according to the present invention. It is conceivable that it is possible to achieve yataan recognition that is strong against changes in the position of the armpit tongue without dropping high-frequency components, but in that case, the amount of calculation would be enormous, so it is not practical.
No 〇 Now N2 To find the spatial frequency power spectrum or autocorrelation function of this image consisting of pixels, N
Suppose we need this calculation (multiplication). As in the present invention,
When processing is performed by dividing into small head regions, 111Xl'l,...
ΣNxn1 This amount of calculation is enough, and the reduction in calculation amount is N
X n, / N αl / n; If fast Fourier transform is used, the reduction in calculation amount will be smaller than on this evaluation side, but a significant reduction can still be achieved. Therefore, by using the method according to the present invention, high-speed pattern recognition that is resistant to pattern position fluctuations can be realized.

(Embodiment) FIG. 4 is a flow chart showing an embodiment in which the processing according to the present invention is realized by a parallel processing method using a multi-grocer configuration. Image information of a pattern input from a television camera or an image scanner is converted into numerical values of chiaroscuro for each pixel and stored in the image memory. Next, input image 1
M (Let lI!Ii be a prime number N×N), the number of pixels n1
Xn, is divided into "2xnt" sub-regions. Each stroke divided into these sub-regions is sent to a dedicated processor, and for each sub-region, its spatial frequency, It is converted into 9-tan of the 9-war spectrum.At this time, fast Fourier transform may be used for the processing.

Spatial frequency 7f for each small area obtained in this way
The set of four patterns of worth vectors is sent to the next pattern judgment unit, where the same processing is performed in advance and pattern judgment is performed by template matching using a standard pattern. Since the processing according to the present invention can be executed independently for each divided small area, parallel processing as shown in FIG. 4 is possible.

Of course, processing speed increases by parallelizing processing.

(Effects of the Invention) By performing the processing using this method to convert the nof tongue and then performing the no amount tongue matching, the tongue recognition method is resistant to misalignment of the presentation and mother tongue and can be executed at high speed. can be realized. Compared to the conventional method of taking the moving average of the f tongue and artificially adding blur to the nine tongue, this method does not discard information on the high frequency components of the presented arm tongue, and is effective against positional shifts. It has the advantage of being able to handle large displacements without compromising resolution.

Furthermore, compared to a method that converts the entire input image into a pattern of its spatial frequency power spectrum, the method of the present invention can achieve a significant reduction in the amount of calculation and can perform high-speed processing. However, since processing can be executed independently for each divided small area, it is possible to use a method suitable for increasing speed by using multiple processors as shown in the embodiment.

[Brief explanation of drawings]

Figure 1 is a flowchart of the arm tongue recognition method using image signal processing according to the present invention, and Figure 2 is an example of a zfJn that is converted by translational movement using the processing according to the present invention and does not change at all. FIG. 3 is a diagram showing a comparison between the moving average operation and the processing according to the present invention, and FIG. 4 is a flowchart showing an embodiment of the present invention using a multiprocessor configuration. Agent Patent Attorney Shinsuke Honjo Figure 3
xI(x)dx(Sho22 I(X) P(R,X) Fig. 4

Claims (1)

[Claims] An image signal processing method characterized by dividing an input image into small regions that overlap or do not overlap each other, and converting the input image into a pattern of a spatial frequency power spectrum or an autocorrelation function for each of these small regions. .