JPWO2008075592A1 - Image discrimination method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To accurately determine, for example, a partially missing character or number in an image determining method for determining whether a character, number or other image has the same category as a sample image. When a number is expressed in units of 7 × 7 = 49 pixels, statistics are obtained from a plurality of sample images displaying the same number, and similarity is determined from a statistical constant and an image to be determined. To do. At this time, constants are determined by all the statistics of 49 pixel units, and the image is divided into a plurality of areas, and correlation coefficients in the respective areas are obtained. Thus, by determining the constants in two steps, for example, an image in which a part of a character is missing can be determined as an image of the same type as the sample image. [Selection] Figure 4

Description

  The present invention relates to an image discriminating method capable of discriminating whether an image obtained by capturing characters, numbers, symbols, or other objects to be discriminated is of the same category as a sample image.

  Images of characters, numbers, symbols, and other objects should be photographed with an image sensor such as a CCD, and the image should be determined to be the same as the sample image by analyzing the image on the computer software. An image discriminating method for discriminating whether or not there is being realized.

  The pattern detection method based on image recognition described in Patent Document 1 below divides a pattern image of a detection object, and compares the divided pattern with a reference pattern so that the pattern of the detection object is the same type as the reference pattern. It is determined whether it is a thing. A normalized correlation coefficient is used as this discrimination method.

The character recognition method described in Patent Document 2 also divides a pattern image of an object and calculates a spatial distance between the image of each divided area and a reference image. The Euclidean distance or Mahalanobis distance is used as this spatial distance. In the invention described in Patent Document 2, by multiplying a predetermined coefficient for each area obtained by dividing a pattern image and making a similarity determination with a reference image, for example, even if there is a portion that is easily crushed in a character, It is to enable judgment of similarity.
Japanese Patent Laid-Open No. 11-160043 Japanese Patent Laid-Open No. 11-238099

  The conventional image discrimination method including the invention described in the above-mentioned patent literature performs image similarity discrimination based on one kind of calculation. However, the conventional discrimination method using the normalized correlation coefficient and the discrimination method using the Euclidean distance or the Mahalanobis distance cannot always accurately perform the target image discrimination.

  FIGS. 3A, 3B, and 3C show results of attempts to discriminate images by different methods. In this test method, as shown in FIG. 2, assuming a 7 × 7 pixel unit on the computer, one of the pixel units is set to black, and the numbers “0” to “9” are expressed. To do. Then, the number display of “6” shown in FIG. 2 is used as a reference image, and each of “0”, “1”, “2”, “3”, “4”, “5”, “6”, “7”, “8”, and “9” is displayed. The numerical display is used as a discrimination target, and the spatial distance between each discrimination target and the numeric display of “6” as the reference image is calculated.

  FIG. 3A shows a normalized correlation value between the reference image and the discrimination target. 1 / (normalized correlation value) is taken on the vertical axis, and “0”, “1”, “2”, “3”, “4”, “5”, “7”, “8”, and “9” that are discrimination target images are taken on the horizontal axis. Each numerical display is shown. When the reference image and the discrimination target image are the same “6”, the normalized correlation value is 1.

  The vertical axis in FIG. 3B represents the Mahalanobis distance. The Mahalanobis distance in FIG. 3B is calculated from the sum of data in 7 × 7 pixel units and the number of items. When the reference image and the discrimination target image are the same “6”, the Mahalanobis distance is 0-1.

  In FIG. 3C, a proportional constant and a standard deviation are calculated from 7 × 7 pixel unit data, and the Mahalanobis distance is obtained using two items of the proportional constant and the standard deviation. In this case, when the reference image and the discrimination target image are the same “6”, the Mahalanobis distance is 1.

  As shown in FIG. 3A, in the discrimination method using the normalized correlation value, the correlation value between each numeric value that is a discrimination target and the number “6” that is the reference image is close, and the reference image It can be understood that it is difficult to distinguish the object from the image. On the other hand, as shown in FIG. 3B, the Mahalanobis distance calculated using all image data in pixel units is longer than that in FIG. 3A, and as shown in FIG. The Mahalanobis distance calculated using the two items of the proportionality constant and the standard deviation is further increased.

  That is, when the image to be discriminated is compared with the reference image, the probability that the method shown in FIG. 3B is confused and determined is lower than that in FIG. In the method (C), the probability of being further confused is determined. However, on the other hand, when an image that should be judged to be “6” when it is viewed with the human eye, for example, only a part of the numeral “6” is missing, In the methods of (B) and FIG. 3 (C), it is determined that the spatial distance is far from the reference image, and the result is that the number “6” in which only a part is missing is excluded as being different. It becomes.

  As described above, as shown in FIGS. 3B and 3C, there is a problem that the sensitivity of image discrimination is increased, and conversely, what is originally determined to be the same or similar is determined to be non-identical.

  The present invention solves the above-described conventional problems. When the reference sample image is different from the image of the object to be discriminated, confusion discrimination is performed by increasing the spatial distance. On the other hand, an object of the present invention is to provide an image discriminating method capable of discriminating an image having a chipped portion from the same type as a reference image.

According to a first aspect of the present invention, there is provided an image determination method for determining similarity between a target image as a determination target and a sample image.
One sample image has image data of k pixel units, and n sample images that can be determined to be the same type of image are used, and each one sample image is divided into a plurality of regions for each of a plurality of pixel units. Segment
A correlation coefficient is calculated from image data in units of pixels located in the same area of n sample images, and the correlation coefficient is obtained for each area.
Constants obtained as statistics of image data of k × n pixel units of all sample images and the correlation coefficient for each region of the sample image, image data of k pixel units of the target image, A spatial distance between the target image and the sample image is obtained from a correlation coefficient of each region of the target image, and an analogy between the target image and the sample image is determined based on the obtained spatial distance. is there.

  In the first aspect of the present invention, constants obtained as statistics of image data of k × n pixel units can be obtained as follows.

An average value M _j (j = 1, 2, 3,..., k) of n image data in pixel units at the same position in each of n sample images, and each pixel for each sample image A linear form L _i (i = 1, 2, 3,..., N) obtained by adding the product of the unit image data and the average value M _j, and an effective divisor r obtained by adding each average value M _j. Obtain Y ₁ (i) = L _i / r which is a proportionality constant,
An error variance Ve (i) is obtained from image data in units of pixels at the same position in each of the n sample images, and the SN ratio η (for each sample image is obtained from the error variance Ve (i) and the effective divisor r. i), and the square root of the reciprocal of this SN ratio η (i) is calculated to calculate Y ₂ (i), which is the standard deviation for each sample image,
Two items, an average value of n Y ₁ (i) and an average value of n Y ₂ (i), are defined as the constants.

In the first aspect of the present invention, the correlation coefficient of each region is
The number of pixel units in one region of the sample image is p, and an average value X _q bar (q = 1, 2, 3, 3) of n image data in pixel units at the same position in each of the n sample images. .., P), the average value Y _i bar (i = 1, 2, 3,..., N) of the image data of p pixel units for each sample image, and the same for all sample images The average value X bar of image data of n × p pixel units in the region can be obtained.

  In the first aspect of the present invention, the spatial distance is a Mahalanobis distance. Alternatively, the spatial distance is Euclidean distance.

  In the first invention, the spatial distance can be calculated by assigning a coefficient to each of the constant and the correlation coefficient of each region. In this case, it is preferable that the information amount of each region is compared, and the coefficient assigned to the correlation coefficient obtained in the region where the information amount is small is made smaller than the coefficient assigned to the correlation coefficient obtained in the other region. .

  As described above, by adjusting the weighting of the correlation coefficient obtained from each region according to the information amount of each region divided in the sample image, an optimal image corresponding to the characteristics of the sample image Discrimination can be made.

According to a second aspect of the present invention, there is provided an image determination method for determining similarity between a target image that is a determination target and a sample image.
Using n sample images that have image data for each pixel unit and can be determined as the same type of image, each sample image is divided into a plurality of regions for each pixel unit, and one sample image And k is the total number of pixel units and the number of segmented areas in
X _ij (i = 1, 2, 3,..., K) (j = 1, 2, k) is used as the correlation coefficient of the image data for each pixel and the image data obtained for each divided area. 3, ..., n)
The average value of the image data of each pixel unit in the number n of sample images and the average value of the correlation coefficient in each region in the number n of sample images are both _mi (i = 1, 2, 3). , ..., k),
Σ _i (i = 1, 2, 3) is the standard deviation of the image data of each pixel unit in the number n of sample images and the standard deviation of the correlation coefficient in each region in the number n of sample images. , ..., k)
Obtain a normalized value x _ij = (X _ij −m _i ) / σ _i ,
The spatial distance between the target image and the sample image is obtained from the image data in pixel units of the target image, the correlation coefficient in each region of the target image, and the normalized value x _ij , and is calculated according to the obtained spatial distance. In this case, an analogy between the target image and the sample image is discriminated.

In the second invention, the correlation coefficient of each region is
An average value X _q bar (q = 1, 2, 3,...) Of n image data in pixel units at the same position in each of n sample images, where p is the number of pixel units in one region in the sample image. .., P), the average value Y _i bar (i = 1, 2, 3,..., N) of image data of p pixel units for each sample image, and the same area of all sample images And the average value X bar of the image data of n × p pixel units.
Furthermore, in the second aspect of the present invention, the spatial distance is a Mahalanobis distance.

  In the present invention, the spatial distance between the target image to be discriminated and the sample image is obtained by two types of calculations. First, by calculating the Mahalanobis distance, etc., from the entire sample image and the entire image that is the discrimination target, the spatial distance between the image that should not be determined to be the same type as the sample image and the sample image can be calculated far away. Furthermore, the sample image and the image to be discriminated are divided into a plurality of regions, and the spatial distance between the regions can be calculated with a short distance using a correlation coefficient or the like. Then, by giving predetermined weights to the two calculation results, the spatial distance between the image of the determination object in which a part of the character is missing and the sample image can be shortened.

  Therefore, it is possible to determine with high accuracy the similarity to the sample image even if a part of the character is missing or an image that is captured slightly blurred by the image sensor.

FIG. 1 is a block diagram showing an image discriminating apparatus according to an embodiment of the present invention.
The image discrimination device has a camera 1. The camera 1 includes an image sensor such as a CCD or a CMOS. The image of the discrimination target 10 photographed by the camera 1 is converted into digital signal image data by the A / D converter 2. The image data of the digital signal is calculated by the calculation unit 3 and sent to the determination unit 4. The memory 5 connected to the calculation unit 3 and the determination unit 4 holds various constants calculated from the sample images to be used as the determination reference. The arithmetic unit 3 obtains the spatial distance between the sample image and the image of the discrimination target object from each constant held in the memory 5 and the image data obtained from the A / D conversion unit 2. It is determined whether or not the spatial distance obtained in step is smaller than a threshold value.

  FIG. 2 shows an example of a sample image that is a reference image. The sample image shown in FIG. 2 is configured in units of 7 × 7 pixels, and numbers “0” to “9” are indicated by the brightness (darkness) of each pixel unit. Here, each pixel unit of 7 × 7 = 49 may be a unit of one sensor such as a CCD, or images taken by a plurality of sensors are collectively handled as one pixel unit. Also good.

  In this embodiment, the sample image shown in FIG. 2 is a number, and by calculating the spatial distance between the image to be determined and each sample image, the image to be determined is “0” to “9”. It is possible to determine which number of "." Therefore, the image data of each pixel unit of the sample image shown in FIG. 2 may be handled as binary data of white or black, or the light and dark image data of one pixel unit has a gradation such as 256 gradations. The spatial distance between the image to be discriminated and the sample image may be obtained including this gradation.

  A plurality of sample images with the same number shown in FIG. 2 are prepared, a constant based on statistics is calculated from each sample image, and the constant is stored in the memory 5. For example, assuming a case where a number is printed, this number is read by the camera 1 and A / D converted, and it is determined to which number the image belongs, Variations in color or variations in images captured by the camera 1 that occur each time shooting is assumed. In this way, image data with variation is used as a plurality of sample images, the characteristics of variation in data in images expressing the same number are calculated by statistics, and a constant based on the statistics is calculated. Then, this constant is held in the memory 5.

  A sample image used for this statistic is obtained by photographing a plurality of actually printed numbers with the camera 1 and A / D converting. Alternatively, in calculating this constant, instead of actually obtaining the number by photographing with the camera 1, for example, in a computer, a number formed in units of 7 × 7 pixels shown in FIG. Is set to “0”, the black pixel unit is set to “1”, and random numbers are applied to the “0” and “1” to cause variation in the image data of each pixel unit. It is also possible to use image data as a plurality of sample images and calculate a constant based on statistics.

  Table 1 below indicates image data of a plurality of sample images. The sample image data shown in Table 1 shows, for example, variation in image data of a plurality of sample images when the number “6” shown in FIG. 2 is shot a plurality of times, or a number “6” artificially created by a computer. This means variation in image data.

“N” at the leftmost end of the uppermost column in Table 1 means the number of sample images, and the number of sample images is “n”. X ₁ to X _k arranged in a row horizontally in the uppermost column mean pixel units included in one sample image. When the number of pixel units is 7 × 7 as shown in FIG. 2, “k” in Table 1 is “49”.

In Table 1, the image data of each pixel unit in each sample image having sample numbers 1 to n is indicated by “X _ij ”. “I” represents the number of k pixel units in one sample image, and “i = 1, 2, 3,..., K”. “J” represents the number of n sample images, and “j = 1, 2, 3,..., N”. For example X ₁₂ means that an image data in the first pixel in the sample image of the sample number "2". The image data X _ij is, for example, data of 256 gradations when the image is actually taken by a camera, and is data having random variations when a sample image is created on a computer.

Using the image data _Xij of each pixel unit of each sample image shown in Table 1, statistics of the image data are calculated. This calculation is performed by the arithmetic unit 3 and the discriminating unit 4 shown in FIG. 1, and finally calculated constants are stored in the memory 5. Alternatively, using the image data of the sample image, a constant based on statistics may be calculated by a computer different from the image discrimination device shown in FIG. 1 and the constant may be stored in the memory 5.

In the following formula 1, average values M ₁ , M ₂ ,..., M _k are calculated. The average values M ₁ , M ₂ ,..., M _k are average values of image data in pixel units at the same position in all n sample images. That is, “i” in the image data X _ij is an average value of n pieces of image data having the same numerical value. Therefore, the number of average values calculated is k, which is the same as the number of pixel units in one sample image.

In the following equation 2, fluctuations S _T (1), S _T (2),..., S _T (n) of image data included in one image sample are calculated for each sample image. This variation is obtained by squaring each of k pixel data included in each sample image and adding all the square values in one sample image, and has a unique value for each sample image. Therefore, the calculated variation is n, which is the same as the number of sample images.

In the following formula 3, the line formats L ₁ , L ₂ ,... L _n are calculated. The line format is obtained by calculating the product of the image data X _ij of each pixel unit in one sample image and the average value M _i calculated in Equation 1 based on the pixel unit, and adding the products. . The line format has a unique value in each sample image, and the number of line formats is n, which is the same as the number of sample images.

Formula 4 calculates the effective divisor r, and is a value obtained by squaring and adding each of the average values M ₁ , M ₂ ,..., M _k obtained in Formula ₁ .

Equation 5 is obtained from the linear form L _j obtained in Equation 3 and the effective divisor r obtained in Equation 4, and the variation S _β (1), S _β (2) _,. .., S _β (n) is obtained. Equation 6 obtains proportional constants β (1), β (2),... Β (n) from the line format L _j and the effective divisor r.

Equation 7 calculates error variances Ve (1), Ve (2),..., Ve (n) of image data in each sample image. This error variance Ve (j) is the total variation obtained from all the image data S _T (j) in each sample image obtained by Equation 2, and the proportional term of the sample image obtained by Equation 5. Is obtained from the fluctuation S _β (j).

  As shown in Equation 8 below, the SN ratios η (1), η (2),..., Η (n) are obtained from the reciprocal of the error variance Ve (j) and the effective divisor r. .

In this embodiment, the image data obtained from n sample images having k pixel units, that is, the statistics of the image data of X ₁₁ , X _{21 ,.} Two items Y ₁ (j) and Y ₂ (j) shown in Equation 9 and Equation 10 are calculated.

The first items Y ₁ (1), Y ₁ (2),..., Y ₁ (n) obtained in Expression 9 are the proportionality constants shown in Expression 6, which are the respective sample images. It represents the sensitivity of the obtained image data. The second items Y ₂ (1), Y ₂ (2),..., Y ₂ (n) obtained by Equation 10 are standard deviations of the image data of each sample image.

Table 2 below shows the constants calculated so far.

The average value I _{1 of} the first items Y ₁ (1), Y ₁ (2),..., Y ₁ (n) calculated from n × k image data obtained from n sample images, The average value I ₂ of the second items Y ₂ (1), Y ₂ (2),..., Y ₂ (n) means the zero point of the image data group having variations shown in Table 1. (See Table 4 for I ₁ and I ₂ ). This embodiment is based on obtaining a spatial distance between an image to be determined and the zero point.

FIG. 3 (C), using the average _{I 2} of the first item _Y 1 mean _{I 1} and the second item _Y 2 of (j) (j) as a constant, the image and the sample image as a determination target The results of the Mahalanobis distance, which is the spatial distance between and. When the spatial distance is calculated using the average value I ₁ and the average value I ₂ , the spatial distance is increased only by slightly different images. Therefore, as shown in FIG. 3C, the spatial distance between the sample image “6” and the image of a number other than “6” is long, and the numbers other than “6” are the same as “6”. The probability of being judged can be considerably reduced. However, on the other hand, since the spatial distance is increased only by slightly changing the character shape of the sample image and the image to be discriminated, for example, an image to be discriminated that lacks a part of the number is classified into the category of the number. It is possible that it cannot be determined.

Therefore, in the image determination method of this embodiment, obtains the spatial distance on the basis of only the average value I ₂ of the first item Y ₁ mean I ₁ and the second item Y ₂ of _{(j) (j)} Instead, the sample image and the image to be discriminated are both divided into a plurality of regions, and the spatial distance for each region is obtained using a normalized correlation coefficient. It is possible to determine that an image with only a missing part is the same image as the sample image.

  The number of divisions of the image area is arbitrary, but in this embodiment, the image is divided into four areas. In the example shown in FIG. 2, the number of pixel units of one image is 7 × 7 = 49, but the number of pixel units included in each region when one image is divided into four regions Area is 4 × 4 = 16 pixel units, the lower left area is 4 × 4 = 16 pixel units, and the upper right area and lower right area are 4 × 4 = 16 pixel units. That is, in the area divided into four, there are pixels that overlap each other. However, the divided areas may not be overlapped and may be divided so that no common pixel exists in both of the divided areas.

  The region is set for each of the n sample images shown in Table 1, and a correlation coefficient for each region is obtained.

  The leftmost column N in Table 3 below shows the numbers 1 to n of the sample images. Each sample image has 7 × 7 = 49 pixel units. As described above, each sample image is divided into four regions. Table 3 shows 4 × 4 = 16 pixel unit image data located in the upper left region of each of the sample images 1 to n. Since the number of image data included in the upper left area of one sample image is 16 and the number of sample images is n, the number of image data appearing in Table 3 is 16 × n.

Equation 11 below represents all of the image data X _ij (i = 1, 2, 3,..., 16) (j = 1, 2, 3,..., N) described in Table 3. The average value is indicated by an X bar (in this specification, a symbol having a line drawn at the upper end of the symbol is referred to as a “bar”). The average value of 16 pieces of image data for each sample image is indicated by Y _j bar (j = 1, 2, 3,..., N). Furthermore, the mean value of _{X i} is indicated by a bar of n image data in pixel units at the same position of each sample image (i = 1,2,3, ···, 16 ).

Equations (12) to (15) are used to obtain the correlation coefficient r ₁ (1) of 16 pieces of image data included in the upper left area of the sample image of sample number 1 (N = 1) shown in Table 3. Yes. The correlation coefficient r ₁ (2) of the 16 pieces of image data included in the region with the folds of the sample image of sample number 2 (N = 2) shown in Table 3 is expressed by Equations 12 to 15. using the _{X i, 2} instead of X _{i, 1,} may be calculated using the _{Y 2} bar instead of _{Y 1} bar. Correlation coefficients r ₁ (3), r ₁ (4),... R ₁ (n) of 16 pieces of image data included in the upper left segmented area of the sample image of sample number 3 (N = 3) or less ) Is the same.

Similarly, the correlation coefficient r ₂ (j) of the image data in the lower left area of each sample image, the correlation coefficient r ₃ (j) of the image data in the upper right area of each sample image, and each sample image The correlation coefficient r ₄ (j) of the image data in the lower right region at (j = 1, 2, 3,..., N) can be obtained.

Table 4 below shows, for each of the n sample images, the first item Y ₁ (j), Y ₂ (j), and the correlation coefficient r ₁ for each of the upper left, lower left, upper right and lower right regions. (J), r ₂ (j), r ₃ (j), r ₄ (j) are shown (j = 1, 2, 3,..., N). In the memory 5 shown in FIG. 1, each constant shown in Table 4 is stored as template data for discriminating an image to be discriminated.

  Next, calculation for obtaining the spatial distance between the image to be discriminated and the sample image will be described.

  Image data can be obtained by converting an image obtained by photographing the discrimination target 10 with the camera 1 shown in FIG. 1 into a digital signal by an A / D converter. This image data is given to the calculation unit 3, and the calculation unit 3 performs the next calculation.

The image to be discriminated has image data of 7 × 7 = 49 pixel units. Image data of each pixel unit of the image to be discriminated is represented by X ₁ , X ₂ , X ₃ ,... X _k .

The average values M ₁ , M ₂ , M ₃ ,..., M _k of the image data of the sample image obtained in Table 1 and Equation 1, and the image data X ₁ , X ₂ , X _{3 of the} discrimination target image , · · · _{X k} and a, by performing the same calculation with the number 3, is a particular value in the image of the determination target linear equation _{L 0 = M 1 X 1 +} M 2 X 2 + ··· + M k X k Ask for. Using this linear format L ₀ and the effective divisor r obtained from the sample image as shown in Equation 4, β ₀ = L ₀ / r is calculated in the same manner as in Equation 6. This β ₀ is the first item y1 in the image to be discriminated.

Then, from the linear equation L ₀ of the image is determined subject to the effective divisor r obtained by the number 4, in the same manner as the number 5, the proportional term S _beta (0) of the image as a determination target = L ₀ ² / r is calculated. Further, using the image data X ₁ , X ₂ , X ₃ ,... X _k of the image to be discriminated, the total variation S _T = X ₁ ² + X peculiar to the image to be discriminated in the same manner as Equation ² . ₂ ² + X ₃ ² +... + X _k ² is obtained. This and total variation S _T, and proportional term S _beta (0), with the number k of pixels, by performing the same calculation with the number 7, the error variance Ve image data of an image to a determination target (0) Ask for. Further, the SN ratio η (0) of the image data of the image to be determined is obtained from the error variance Ve (0) and the effective divisor r obtained in Equation 4 in the same manner as in Equation 8. Similarly to the above, by taking the square root of the reciprocal of the SN ratio η (0), the sensitivity of the image data of the image to be discriminated, that is, the second item y2 is obtained.

Further, the image to be discriminated is divided into four areas in the same manner as the sample image, the correlation coefficient r ₁ (0) of the 16 pixel data in the upper left area of the image to be discriminated, and 16 in the lower left area. Correlation coefficient r ₂ (0) of 16 pieces of pixel data, correlation coefficient r ₃ (0) of 16 pieces of pixel data in the upper right area, and correlation coefficient r ₄ (16 pieces of image data in the lower right area) 0).

The method for obtaining the correlation coefficient r ₁ (0) of the upper left region of the image to be determined is the same as the method for obtaining the correlation coefficient for any one of the n sample images shown in Table 3. In the same manner as Y ₁ bar that is described in equation 11 First, the average value Y ₀ of 16 image data of the upper left area of the image to be determined target.
Y ₀ = (1/16) (X ₁ + X ₂ + X ₃ +... + X ₁₆ )

Then, the image data X _i is used instead of X _{i, 1} in Equation 15 (i = 1, 2, 3,..., K), and the value of Y ₀ is used instead of Y ₁ bar in Equation 15. By using it, the correlation coefficient r ₁ (0) of the upper left area of the image to be determined can be obtained. Similarly, the correlation coefficient r _{2 (0)} of the lower left region, the correlation coefficient r _{4 (0)} of the correlation coefficient r _{3 (0)} and the lower right area of the upper right region can be obtained.

As described above, the first item y1, the second item y2, and the correlation coefficients r ₁ (0), r ₂ (0), r ₃ (0), r ₄ obtained from the image to be determined. The following calculation is performed from (0) and the average values I ₁ , I ₂ , I ₃ , I ₄ , I ₅ , and I _{6 of the six} items listed in the bottom row of Table 4.

Y ₁ = y ₁ −I ₁
Y ₂ = y ₂ −I ₂
Y ₃ = r ₁ (0) −I ₃
Y ₄ = r ₂ (0) −I ₄
Y ₅ = r ₃ (0) −I ₅
Y ₆ = r ₄ (0) −I ₆

The correlation coefficient matrix shown in Formula 16 is obtained from I ₁ , I ₂ , I ₃ , I ₄ , I ₅ , and I ₆ that are average values of the items 1 to 6 shown in Table 4. In this correlation coefficient matrix, r _xy is a correlation coefficient between I _x and I _y . R ₁₁ , r ₂₂ , r ₃₃ , r ₄₄ , r ₅₅ , r ₆₆ in the matrix mean a correlation coefficient between I ₁ and I ₁ , a correlation coefficient between I ₂ and I ₂ , etc. In both cases, the correlation coefficient is “1”. Equation 17 is an inverse matrix of the correlation coefficient matrix shown in Equation 16.

As shown in Equation 18, the Mahalanobis distance is obtained from the inverse matrix of Equation 17 and the six items Y ₁ to Y ₆ obtained from the discrimination target image.

  The discrimination unit 4 shown in FIG. 1 sets the threshold value of the Mahalanobis distance shown in Formula 18, and if the Mahalanobis distance is equal to or less than the threshold value, the discrimination target image is of the same category as the sample image. to decide. For example, if the sample image is the number “6” shown in FIG. 2, if the Mahalanobis distance is less than or equal to the threshold value, it is determined that the image to be determined indicates the number “6”.

  As shown in Table 1, the Mahalanobis distance compares the statistics of all the image data of a plurality of sample images with the entire image data to be discriminated, and further determines the plurality of sample images and the discrimination target. The image is divided into a plurality of regions, and the correlation coefficient for each region is taken into consideration. For this reason, there is a slight difference in characters, when the image to be identified is the same number or character as the sample image, and some of the number or character is missing or the image is blurred. However, it is possible to adjust so that the zero point of the sample image and the spatial distance of the image to be discriminated are not greatly separated. Therefore, it is possible to determine that a part of the character or the like is the same type as the sample image. Further, for example, when the sample image has the number “6” and the image to be determined is an image having a number other than “6”, the zero point of the statistical value of the sample image and the image to be determined The spatial distance can be separated, and the probability that different numbers are judged to be the same type of image is extremely low.

  4 and 5 are simulations when the Mahalanobis distance is calculated by the method of the embodiment of the present invention and when the Mahalanobis distance is calculated by the method of the comparative example similar to FIG. And the results.

  In this simulation, as shown in FIG. 4, a 7 × 7 pixel unit is set on the computer, the image data when each pixel unit is black is “1”, and the image data when white is “1”. An image expressing a number as “0” was generated. As the sample image, as shown in FIG. 4A, an image in which the number “6” is expressed by “1” “0” data in a 7 × 7 pixel unit is used.

Then, in order to obtain the statistics of the sample image, a random number in the range of 0 to 0.05 is added to the data “1” or “0” of each pixel of 7 × 7 shown in FIG. N sample images having variations are generated in a pseudo manner, and using these n sample images, constants Y ₁ (j) and Y ₂ (j) based on statistics are calculated using Equations ₁ to 10. Further, each sample image is divided into four regions each having 4 × 4 = 16 pixels, and the operations of Equations 11 to 15 are performed for each region, and the four regions are divided into four regions. Correlation coefficients r ₁ (j), r ₂ (j), r ₃ (j), and r ₄ (j) were obtained.

Then, based on the sample image of the number “6” with pseudo variation, average values I ₁ , I ₂ , I ₃ , I ₄ , I ₅ , I ₆ shown in Table 4 are obtained. It was.

  Next, as a target image to be discriminated, a 7 × 7 pixel unit is set on the same computer, and each pixel data is expressed as a number with “1” (black) or “0” (white). Using.

  Here, one image representing the number “8” shown in FIG. 2 is used as a first target image serving as a guideline for Mahalanobis distance.

  Further, as the second target image serving as a guideline for character discrimination, one image expressing the number “6” as shown in FIG. 4A is used, and as shown in FIG. 4B. , One image unit of “6” with “black” set to “white” to create “missing”, and two numbers of “6” as shown in FIG. One image in which “black” in the pixel unit is set to “white” to create “missing”, and “black” in three pixel units of the number “6” as shown in FIG. Each image used was “white” and “chip” was used.

For each of the first target image representing the number “8” and the second target images shown in FIGS. 4A, 4B, 4C, and 4D, the constants y1 and y2 are calculated as described above. did. Further, each target image is divided into four regions every 4 × 4 = 16 pixels, and correlation coefficients r ₁ (0), r ₂ (0), r ₃ (0) are assigned to each region. ), R ₄ (0). Then, the average values I ₁ , I ₂ , I ₃ , I ₄ , I ₅ , and I ₆ shown in Table 4 obtained from the sample image of the number “6” having variations are obtained from the respective target images. Y ₁ , Y ₂ , Y ₃ , Y ₄ , Y from the constants y1, y2 and correlation coefficients r ₁ (0), r ₂ (0), r ₃ (0), r ₄ (0) ₅ and Y ₆ were obtained.
Then, the Mahalanobis distance was obtained by the following two methods (a) and (b).

(A) Comparative example (calculation using only Y ₁ and Y ₂ )
As a comparative example, the Mahalanobis distance was calculated using only Y ₁ and Y ₂ . That is, the number 9 to the first item _Y 1 obtained from the sample image by the number 10 (j) and the second item _Y 2 (j), and _{Y 1} obtained from the constant y1, y2 of the respective target image , using Y _2, without using the correlation coefficient calculated by dividing the image into four and Ji order number 16 to number 18, was calculated Mahalanobis distance.

In this calculation, the following spatial distances (a) (b) (c) (d) (e) were obtained.
(A) Spatial distance between the sample image with the number “6” and the first target image with the number “8”;
(A) Spatial distance between the sample image of the number “6” and the target image of the number “6” shown in FIG.
(C) Spatial distance between the sample image of the number “6” and the target image lacking only one pixel from the number “6” as shown in FIG.
(D) The spatial distance between the sample image of the number “6” and the target image lacking only two pixels from the number “6” as shown in FIG.
(E) the spatial distance between the sample image of the number “6” and the target image lacking only 3 pixels from the number “6” as shown in FIG.
The diagram (a) of FIG. 5 shows the tendency of the spatial distance obtained by the above calculation, where “0” on the horizontal axis is the ratio (i / a) of the spatial distance, and “1” is the space. The distance ratio (U / A), “2” is the spatial distance ratio (E / A), and “3” is the spatial distance ratio (O / A). The vertical axis in FIG. 5 is (%), and “0%”, “5%”, “10%”, “15%”... “40%” from the bottom.

(B) Embodiment of the present invention (calculation using Y ₁ , Y ₂ , Y ₃ , Y ₄ , Y ₅ , Y ₆ )
Next, using all of Y ₁ , Y ₂ , Y ₃ , Y ₄ , Y ₅ , and Y ₆ , the Mahalanobis distance was calculated in accordance with Equations 16 to 18. That is, not only the first item Y ₁ (j) and the second item Y ₂ (j) shown in Equations 9 and 10, but also the correlation coefficient r ₁ (j) calculated by dividing the image into four. , R ₂ (j), r ₃ (j), and r ₄ (j) were used to determine the Y ₁ , Y ₂ , Y ₃ , Y ₄ , Y ₅ , and Y ₆ to calculate the Mahalanobis distance. .

Also in this calculation, the ratio of the spatial distances of the following (f) (ki) (ku) (ke) (ko) was obtained.
(F) Spatial distance between the sample image with the number “6” and the first target image with the number “8”;
(G) Spatial distance between the sample image of the number “6” and the target image of the number “6” shown in FIG.
(H) The spatial distance between the sample image of the number “6” and the target image lacking only one pixel from the number “6” as shown in FIG.
(K) Spatial distance between the sample image of the number “6” and the target image lacking only two pixels from the number “6” as shown in FIG.
(K) Spatial distance between the sample image of the number “6” and the target image lacking only 3 pixels from the number “6” as shown in FIG.
The diagram (b) in FIG. 5 shows the tendency of the spatial distance obtained by the above calculation, where “0” on the horizontal axis is the spatial distance ratio (ki / ka), and “1” is the spatial distance. The ratio (K / K), “2” is the spatial distance ratio (K / K), and “3” is the spatial distance ratio (K / K).

(C) Results of FIG. 5 In the comparative example shown in the diagram (a) of FIG. 5A, the Mahalanobis distance is calculated without using the correlation coefficient calculated by dividing the image into four. is doing. Therefore, as shown in FIGS. 4B, 4C, and 4D, a part of the number “6” is missing, and as the number of missing pixels increases, the spatial distance from the sample image increases, The spatial distance becomes closer to the spatial distance between the sample image and the image “8”. As shown in FIG. 5, in the diagram (a), when the number of missing pixel units becomes three as shown in FIG. 4 (D), the spatial distance from the sample image is the image “8” and the sample. It is close to 35% of the spatial distance from the image.

  In other words, in this calculation method, if a threshold value is set to any value between 0% and 100% in order to identify the numbers “6” and “8”, the sample image and FIGS. The spatial distance with the image shown in D) may exceed the threshold value, and the probability of determining that the number “6”, which is partially missing, is different from the sample image increases.

On the other hand, in the embodiment of the present invention, not only the first item Y ₁ (j) and the second item Y ₂ (j) but also the correlation coefficient r ₁ calculated by dividing the image into four. The Mahalanobis distance is calculated using (j), r ₂ (j), r ₃ (j), and r ₄ (j). Therefore, as shown in the diagram (b) of FIG. 5, even if the number “6” is missing, the spatial distance between the sample image and the sample image is not as large as the spatial distance between the number “8”. For example, as shown in FIG. 4D, even if “missing” occurs in the three pixel units of the number “6”, the spatial distance between the image and the sample image is the number “8” and the sample image. Less than 4% of the spatial distance.

  When the sample image is the number “6” and the Mahalanobis distance is calculated according to the embodiment, the spatial distance of the missing image as shown in FIGS. 4 (B), (C), and (D) is obtained. This means that the distance of the space “8” is short and can be increased. According to FIG. 3C, when the sample image is the number “6”, the number “8” has the closest spatial distance to “6”. However, in the embodiment, the spatial distance between the sample image and the number “8” image can be kept far, and the number “8” can be determined to be different from the sample image. In addition, as shown in FIGS. 4B, 4C, and 4D, the probability that a missing number can be determined to be the same number as the sample image is increased.

  Therefore, if the threshold value of the spatial distance is set in the range of 3 to 5% in the determination unit 4 of the image determination apparatus shown in FIG. 1, it is the same as the sample image that is partially missing with the same numbers and characters. It can be determined as a number or a character, and for example, the characters “6” and “8” can be clearly identified.

Next, another embodiment of the image discrimination method of the present invention will be described.
In this embodiment, as shown in FIGS. 4B, 4C, and 4D, a part of the numeral “6” is deleted, and Y ₁ , Y ₂ , Y ₃ , Y ₄ , Y ₅ , to obtain a calculated value of the five items of Y _6.

In the following equation 19, the Euclidean distance is calculated using the calculated values Y ₁ , Y ₂ , Y ₃ , Y ₄ , Y ₅ , Y ₆ of the _six items obtained from the image to be discriminated as described above. Seeking. From this Euclidean distance, it is possible to determine whether the image to be determined falls within a predetermined pattern category.

Further, as shown in the following Expression 20, the calculated values Y ₁ , Y ₂ , Y ₃ , Y ₄ , Y ₅ , Y ₆ of the _six items obtained from the image to be discriminated are uniquely weighted. Coefficients a1, a2, a3, a4, a5, and a6 may be added to determine whether the image to be determined falls within a predetermined pattern category. a1 is a weighting coefficient for calculating values _{Y 1,} a2 is the weighting factor for the calculated value _{Y 2, a3, a4, a5} , a6 is for each calculated value _{Y 3, Y 4, Y 5} , Y 6 It is a weighting factor.

That is, when discriminating the image to be discriminated, the weight of the discriminating method using the average value of the first item Y ₁ (j) and the average value of the second item Y ₂ (j) shown in Table 4 By dividing the image and obtaining the correlation coefficient, the weight with the discrimination method can be set as needed according to the type of the image. Accordingly, the coefficient added to Y ₁ and Y ₂ and the coefficients a 1, a 2, a 3, a 4, a ₅ , a ₆ added to Y ₃ , Y ₄ , Y ₅ , Y ₆ are arbitrarily changed without being limited to Equation 20. , Y ₁ , Y ₂ , Y ₃ , Y ₄ , Y ₅ , Y ₆ may be added to obtain the spatial distance.

  For example, when the sample image is the number “6” shown in FIG. 6A, the areas of the four regions (i), (ii), (iii), and (iv) to be divided to obtain the correlation coefficient are As shown in FIG. 6B, the pixel data can be changed according to the amount. Then, the weighting coefficient a5 corresponding to the region (iii) with the smallest amount of information is made smaller than the other coefficients. For example, a1 = 1, a2 = 1, a3 = 1, a4 = 1, a5 = 0.5, a6 = 1, etc. Also, when the sample image is the number “9” shown in FIG. 9A, the areas of the four regions (i), (ii), (iii), and (iv) are reduced according to the amount of information in units of pixels. As shown in FIG. 7B, the pixel data is changed according to the amount of pixel data. Then, the coefficient corresponding to the area (ii) with a small amount of information may be set to a smaller number of a4 = 0.5, and all other coefficients a1, a2, a3, a5, a6 may be set to 1.

  In this way, in accordance with the bias in the information of the sample image, the area of the partition area for obtaining the correlation coefficient is changed, or the weighting coefficient corresponding to the correlation coefficient is changed, so that the characteristics of the image are changed. High-precision characters can be distinguished.

  Further, the spatial distance between the image to be discriminated and the zero point of the sample image may be obtained by the following calculation method.

  In this embodiment, it is determined whether or not an image having, for example, 7 × 7 = 49 pixel units as shown in FIG. 2 is the same type as the sample image. As shown in Table 5 below, the number of sample images is n. In Table 5, the number of pixel units of one sample image is k-4. Therefore, the number of all pixel units of n sample images, that is, the total number of image data is n × (k−4).

“Coefficient 1” in Table 5 is a correlation coefficient of the upper left region when each sample pixel is divided into four regions. Similarly, “coefficient 2” is a correlation coefficient in the lower left area, and “coefficient 3” and “coefficient 4” are a correlation coefficient in the upper right area and a correlation coefficient in the lower right area. That is, X _k-31 in Table 5 is the same as r ₁ (1) in Table 4, and X _{k-21 in} Table 5 is the same as r ₂ (1) in Table 4. X _{k2 in} Table 5 is the same as r ₄ (2) in Table 4. Thus, one sample image has k-4 image data and four correlation coefficients, and in Table 5, both the image data and the correlation coefficient are indicated by X _ij (i = 1, 2, 3,..., K) (j = 1, 2, 3,..., N).

Table 5 shows the average value of n of the image data in units of pixels of the same position of the respective sample images, the average value of n correlation coefficients of the same region of the sample image together with m _i. Also, the standard deviation of n image data in pixel units at the same position in each sample image and the standard deviation of n correlation coefficients in the same region of the sample image are both indicated by σ _i (i = 1). , 2, 3, ..., k). The average value m _i and the standard deviation σ _i can be obtained by the following equation 21.

Each data X _ij of each item shown in Table 5 is normalized by Equation 22 using the average value m _i and the standard deviation σ _i . The normalized x _ij is listed in Equation 23 (i = 1, 2, 3,..., K) (j = 1, 2, 3,..., N).

  In Table 6 below, the image data of each pixel unit of n sample images and the correlation coefficient of each region are standardized and described. The average value of the n normalized values of the pixel unit at the same position of the sample image and the average value of the normalized values of the correlation coefficient of the same region of the sample image are all “0”, and the standard deviation Are all “1”.

Next, a correlation matrix R is obtained by Expression 24.

  The following formula 25 is a general formula for obtaining the correlation coefficient between the item x and the item y. In this embodiment, as shown in Table 6, the average value = 0 and the standard deviation = 1. Therefore, the correlation coefficient is obtained by Equation 26.

Next, an inverse matrix A of the correlation matrix R is obtained by the following equation (27). Then, the inverse matrix A, the average value _mi and the standard deviation σ _i shown in Table 5 are both stored as vectors in the memory 5 shown in FIG.

Table 7 shows the image data of each pixel of the image to be discriminated and the correlation coefficient of each of the four areas divided by X ₁ , X ₂ , X ₃ ,..., X _k . .

Then, according to Equation 28, X ₁ , X ₂ , X ₃ ,..., X _{k of the} images to be discriminated shown in Table 7, the average value _mi and the standard deviation σ _i stored in the memory 5. Y ₁ , Y ₂ , Y ₃ ,... Y _k are calculated from the respective vector values.

Then, the Mahalanobis distance is calculated in Equation 29 from the value of Equation 28, the inverse matrix A stored in the memory 5, and the vector value of the average value _mi and standard deviation σ _i shown in Table 5. Based on the Mahalanobis distance, it is determined whether the image to be determined is in the category of the sample image.

Block diagram of image discrimination device, An explanatory view showing an example of an image, (A), (B) and (C) are graphs showing differences in spatial distance based on the respective calculation methods; (A), (B), (C), and (D) are explanatory diagrams of images used for simulation, Graph showing simulation results (A) (B) is explanatory drawing which shows the other example of a division of a sample image, (A) (B) is explanatory drawing which shows the other example of a division of a sample image,

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Camera 2 A / D conversion part 3 Calculation part 4 Discrimination part 5 Memory 10 Discrimination target object

Claims (10)

In an image discrimination method for discriminating the similarity between a target image that is a discrimination target and a sample image,
One sample image has image data of k pixel units, and n sample images that can be determined to be the same type of image are used, and each one sample image is divided into a plurality of regions for each of a plurality of pixel units. Segment
A correlation coefficient is calculated from image data in units of pixels located in the same area of n sample images, and the correlation coefficient is obtained for each area.
Constants obtained as statistics of image data of k × n pixel units of all sample images and the correlation coefficient for each region of the sample image, image data of k pixel units of the target image, An image discrimination characterized in that a spatial distance between the target image and the sample image is obtained from a correlation coefficient of each region of the target image, and an analogy between the target image and the sample image is discriminated based on the obtained spatial distance. Method. 2. The image discriminating method according to claim 1, wherein a constant obtained as statistics of image data of k × n pixel units is obtained as follows.
An average value M _j (j = 1, 2, 3,..., k) of n image data in pixel units at the same position in each of n sample images, and each pixel for each sample image A linear form L _i (i = 1, 2, 3,..., N) obtained by adding the product of the unit image data and the average value M _j, and an effective divisor r obtained by adding each average value M _j. Obtain Y ₁ (i) = L _i / r which is a proportionality constant,
An error variance Ve (i) is obtained from image data in units of pixels at the same position in each of the n sample images, and the SN ratio η (for each sample image is obtained from the error variance Ve (i) and the effective divisor r. i), and the square root of the reciprocal of this SN ratio η (i) is calculated to calculate Y ₂ (i), which is the standard deviation for each sample image,
Two items, an average value of n Y ₁ (i) and an average value of n Y ₂ (i), are defined as the constants. The correlation coefficient for each region is
The number of pixel units in one region of the sample image is p, and an average value X _q bar (q = 1, 2, 3, 3) of n image data in pixel units at the same position in each of the n sample images. .., P), the average value Y _i bar (i = 1, 2, 3,..., N) of the image data of p pixel units for each sample image, and the same for all sample images An average value X bar of image data of n × p pixel units in the region;
The image discrimination method according to claim 1, wherein the image discrimination method is obtained from:   The image discrimination method according to claim 1, wherein the spatial distance is a Mahalanobis distance.   The image discrimination method according to claim 1, wherein the spatial distance is a Euclidean distance.   The image discrimination method according to claim 1, wherein the spatial distance is calculated by assigning a coefficient to each of the constant and the correlation coefficient of each region.   7. The image discrimination according to claim 6, wherein the information amount of each region is compared, and a coefficient assigned to a correlation coefficient obtained in a region having a small amount of information is made smaller than a coefficient assigned to a correlation coefficient obtained in another region. Method. In an image discrimination method for discriminating the similarity between a target image that is a discrimination target and a sample image,
Using n sample images that have image data for each pixel unit and can be determined as the same type of image, each sample image is divided into a plurality of regions for each pixel unit, and one sample image And k is the total number of pixel units and the number of segmented areas in
X _ij (i = 1, 2, 3,..., K) (j = 1, 2, k) is used as the correlation coefficient of the image data for each pixel and the image data obtained for each divided area. 3, ..., n)
The average value of the image data of each pixel unit in the number n of sample images and the average value of the correlation coefficient in each region in the number n of sample images are both _mi (i = 1, 2, 3). , ..., k),
Σ _i (i = 1, 2, 3) is the standard deviation of the image data of each pixel unit in the number n of sample images and the standard deviation of the correlation coefficient in each region in the number n of sample images. , ..., k)
Obtain a normalized value x _ij = (X _ij −m _i ) / σ _i ,
The spatial distance between the target image and the sample image is obtained from the image data in pixel units of the target image, the correlation coefficient in each region of the target image, and the normalized value x _ij , and is calculated according to the obtained spatial distance. An image discrimination method comprising discriminating an analogy between a target image and a sample image. The correlation coefficient for each region is
An average value X _q bar (q = 1, 2, 3,...) Of n image data in pixel units at the same position in each of n sample images, where p is the number of pixel units in one region in the sample image. .., P), the average value Y _i bar (i = 1, 2, 3,..., N) of image data of p pixel units for each sample image, and the same area of all sample images An average value X bar of image data of n × p pixel units in
The image discrimination method according to claim 8, which is obtained from:   The image discrimination method according to claim 8, wherein the spatial distance is a Mahalanobis distance.

Images