JPH0628476A - Processor for image signal - Google Patents

Abstract

PURPOSE:To provide the image processing method for recognizing a pattern, which is strong against enlargement/reduction, rotation and misalignment, and also, strong against superposition of a background noise. CONSTITUTION:After a pre-processing is executed to an input image, such a transformation as a Fourier/Melan-transformation, etc., is performed thereto by a converting part 14, and an image is converted to a pattern being invariable against enlargement/reduction, rotation and misalignment. A matching part 15 compares the pattern to which such a conversion is performed, with a standard pattern, and determines a candidate pattern. A reverse converting part 16 executes a reverse conversion processing of the conversion by the converting part to the candidate pattern, and calculates a feedback image. An image memory part 13 eliminates a background noise from an input image, based on the feedback image. By repeating these processings, the pattern is decided.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an image signal processing apparatus for performing image processing for pattern recognition.

[0002]

2. Description of the Related Art Conventional recognition methods for an apparatus for performing image processing for pattern recognition such as OCR can be roughly classified into an analog matching method (template matching method) and a structure analysis method. The former is a method for recognizing a pattern by superimposing a standard pattern and an unknown pattern, which is excellent in versatility and is particularly advantageous for reading a printed character.

[0003]

However, in this conventional method, if the pattern is deformed by scaling, rotation, or misaligned, the overlap with the standard pattern is reduced and the pattern cannot be recognized correctly. There was a flaw. “Applied Optics, 1976, vo
l. 15 "by D. Casasent et al.," Position, rotation, and sc. "
ale invariant optical cor
“Relation” includes a Fourier-Melan transform of a pattern by an optical method.
The method of performing pattern matching after performing the (transform) is described. In this method, the amplitude of the Fourier transform of the pattern is invariant even if the position of the original pattern is shifted, and the amplitude of the expanded Melan transform of the pattern is obtained by scaling or rotating the original pattern. Is also invariant, and the original pattern is converted into a pattern that is invariant to misalignment, scaling, rotation, etc., and then matched with the standard pattern. This method is strong against deformation and displacement. However, this method has a drawback that matching is significantly deteriorated when background noise is superimposed on the posted pattern.

An object of the present invention is to eliminate such drawbacks of the prior art and process an image signal for pattern recognition, which is strong against enlargement / reduction, rotation, displacement and at the same time strong against background noise superposition. To provide a device.

[0005]

According to a first aspect of the present invention, a Fourier-Melan transform is applied to an input image to shift the position of the image,
A first conversion unit that converts into a pattern that is invariant to scaling and rotation operations; a matching unit that compares the pattern obtained from the first conversion unit with a standard pattern to determine a candidate pattern;
A first inverse transform unit that performs an inverse transform of the transform performed by the first transform unit on the candidate pattern to calculate a feedback image;
An image memory that removes background noise irrelevant to the pattern from the input image based on the feedback image is provided, and pattern determination is performed by repeating loop processing.

According to the second aspect of the invention, the input image is divided into small regions, multi-stage hierarchical Fourier transform is performed, and the image is invariant with respect to the positional deviation by a process of obtaining the amplitude of the Fourier transform of the image. A second conversion unit for converting to a second conversion unit, and a third conversion for converting the pattern obtained from the second conversion unit into a pattern that is invariant to scaling and rotation operations by performing extended Melan conversion. Section, a matching section that determines a candidate pattern by comparing the pattern obtained from the third converting section with a standard pattern, and the third converting section based on the candidate pattern determined by the matching section and the data from the third converting section. Based on the feedback image; and a second inverse transform unit that calculates the inverse transform of the second inverse transform unit, a third inverse transform unit that calculates a feedback image from the data from the second inverse transform unit and the data from the second transform unit Pattern from the input image And a image memory to remove extraneous background noise is characterized by performing a pattern determined by iteration of the loop processing.

Further, a third aspect of the present invention is a fourth transformation part for subjecting an input image to coordinate transformation represented by complex logarithmic transformation, and a fourth transformation part.
The pattern obtained from the transformation unit of is divided into small regions and subjected to a multi-stage hierarchical Fourier transform, and the image is subjected to small displacement, scaling,
A fifth conversion unit for converting into a pattern invariant to a rotation operation, a matching unit for comparing a pattern obtained from the fifth conversion unit with a standard pattern to determine a candidate pattern, and the multi-stage for the candidate pattern. A fourth inverse transform unit for performing a feedback image by performing the inverse Fourier transform and the inverse transform of the amplitude calculation of the Fourier transform, and an image memory for removing background noise irrelevant to the pattern from the input image based on the feedback image And is characterized by performing pattern determination by repeating loop processing.

[0008]

In the first invention, the Fourier transform is applied to the input image and the amplitude thereof is calculated. As is well known, the amplitude of the Fourier transform of an image is invariant even if the position of the image is shifted, so that a position invariant pattern is obtained. At this time, it is well known that when the original image is enlarged a times, the Fourier transform amplitude pattern is reduced to 1 / a, and when the original image is rotated, the Fourier transform pattern is also rotated by the same angle. There is. Therefore, next, this pattern is represented by new coordinates, the vertical axis represents the deviation angle, and the horizontal axis represents the radius vector on a logarithmic scale. The pattern expressed in this way translates in the vertical direction when the original pattern is rotated,
Further, it is understood that when the original pattern is enlarged or reduced, the original pattern is translated in the horizontal axis direction. Then, the Fourier transform is performed again on this pattern and the amplitude thereof is calculated to obtain the extended Melan transform. As pointed out by Casasent et al., It is invariant to rotation or scaling of the original pattern.

Next, template matching is performed with a standard pattern which has been previously subjected to the same processing as the processing so far. By this process, the presented pattern can be correctly recognized even if it is displaced from the standard pattern, scaled, or rotated. However, in general, the presented pattern is not limited to the target pattern. Since background noise is superimposed, it is difficult to perform accurate recognition with such processing. Therefore, when the similarity obtained by the template matching described above is smaller than a preset threshold value, the determination is suspended, and the standard pattern having a relatively high similarity is set as the candidate pattern. Based on this candidate pattern, the above processing is followed in reverse, a feedback image is calculated, and the input image is corrected based on this feedback image. If the candidate pattern is the correct pattern, the feedback image can be used to determine the part of the pattern to be targeted and the part of the background noise that is not related to the pattern. The background noise part of can be suppressed. The background noise is suppressed by such correction, and the forward conversion process and template matching are repeated again. In this way, when the candidate pattern is a correct pattern, the background noise is suppressed by the feedback processing, so the result of the second template matching is greatly improved, and the above loop processing is repeated several times. If it is repeated, the matching is almost perfect, and it is possible to perform highly accurate pattern determination even in the presence of background noise.

Next, in the second invention, the amplitude calculation process of the Fourier transform, which is the first process in the first invention, is replaced with a process divided into multiple stages as follows. That is,
Instead of computing the Fourier transform of the entire input image at once,
The input image is divided into a number of small areas, and the Fourier transform of the image and its amplitude are calculated for each small area. This gives the amplitude of the local Fourier transform at many positions in the image, which is then considered a new pattern,
The Fourier transform of the position variable is performed again, and its amplitude is calculated. The pattern obtained by such a process has approximately the same property as the amplitude of the Fourier transform for the original image, and therefore, by adding the remaining process of the first invention to this process, misalignment and enlargement can be achieved. It is possible to perform pattern recognition processing that is strong against reduction and rotation and at the same time strong against superposition of background noise.

Next, in the third aspect of the invention, first, the input image is expressed by the coordinate axes subjected to complex logarithmic transformation. As is well known, the pattern obtained by this processing approximately translates when the original image is scaled or rotated.
Next, this pattern is divided into a large number of small areas, and the Fourier transform of the image and its amplitude are calculated for each small area. Here, if the position of the original image is shifted, the image in each small area also moves, but if the magnitude of the displacement in each small area is smaller than the size of the small area, the amplitude of the Fourier transform is reduced. Due to the invariance with respect to the positional deviation of, the pattern obtained by such processing is approximately invariant. Next, the amplitudes of the local Fourier transforms at a large number of positions obtained in this way are regarded as new patterns, and the Fourier transform for the position variable is performed again, and the amplitudes are calculated. The pattern obtained by such a process has approximately the same property as the amplitude of the Fourier transform of the input image represented by the coordinate axis subjected to the complex logarithmic transformation, and therefore, even if the pattern is translated on this coordinate axis, That is, even if the original input image is scaled or rotated, it is approximately unchanged.
Therefore, by the above processing, a pattern that is substantially invariable with respect to a small displacement, scaling conversion, and rotation can be obtained.

Next, template matching is performed with a standard pattern which has been previously subjected to the same processing as the processing so far. Similar to the first invention, when background noise is superimposed on the bulletin pattern, it is difficult to perform accurate recognition by such processing. Therefore, similar to the first invention, when the similarity is smaller than a preset threshold value, a standard pattern having a relatively high similarity is fed back as a candidate pattern, and based on this candidate pattern, the above-mentioned In the reverse order, the feedback image is calculated and the input image is corrected based on this feedback image. After such correction, the forward conversion process and template matching are repeated again.
Similar to the first invention, if the above loop processing is repeated several times, the matching is almost completed, and even if there is background noise, the accuracy is high, and the position shift, enlargement / reduction,
A strong pattern judgment is possible even for rotation.

[0013]

Embodiments of the present invention will now be described with reference to the drawings.

FIG. 1 is a block diagram showing an embodiment of the first invention. The embodiment shown in FIG. 1 includes an image input unit 11 using an image scanner or a television camera,
A pre-processing unit 12 for performing necessary pre-processing such as digitizing data and edge enhancement, an image memory 13 for storing digital data, and Fourier transform for image data read from the image memory 13 and extended Melan. A conversion unit 14 that performs conversion to convert an input into a pattern that is invariable with respect to displacement, rotation, and enlargement / reduction change, and template matching between the output of the conversion unit 14 and a standard pattern is performed to determine a pattern and determine a candidate pattern. The matching unit 15 for performing the above and the inverse transform unit 16 for calculating a feedback image from the candidate pattern determined by the matching unit 15 and the data of the transform unit 14 are provided.

Next, the operation of the embodiment of the first invention will be described. The image information input from the image input unit 11 is
The pre-processing unit 12 receives the A / D conversion processing, digitizes it, and stores it in the memory. If necessary, preprocessing such as enhancement of the edge portion in the image may be performed in this portion so as to be convenient for the subsequent recognition processing. Next, this processing result is sent to the image memory 13, and the transform unit 14 calculates the Fourier transform F (Kx, Ky) and its amplitude P (Kx, Ky) for the pattern in the image memory 13. . The above processing will be referred to as conversion A.
The amplitude thus obtained does not change when the input image is translated, as is well known. Depending on the convenience of calculation, the squared power may be calculated instead of the amplitude. Since the phase of the Fourier transform is used in the subsequent processing, it is stored in the memory in the conversion unit 14.

Next, the transforming unit 14 transforms the amplitude of the Fourier transform into a polar coordinate display, displays the deviation angle on the vertical axis, and the radius vector on a logarithmic scale on the horizontal axis. Expressed as a mathematical formula

## EQU00001 ## rcos .theta. = Kx s = log (1 + r / r
c) The coordinate transformation (Kx, Ky) → (s, θ) represented by rsin θ = Ky is calculated, and the amplitude pattern of the Fourier transform is converted to the s-θ axis pattern. Here, rc is a positive constant for preventing the logarithmic function from diverging at r = 0.

The pattern P (s, θ) thus obtained is
It can be seen that when the original pattern is rotated, it is translated in the θ-axis direction, and when the original pattern is enlarged or reduced, it is approximately translated in the s-axis direction. Therefore, when the Fourier transform is performed again on this pattern and the amplitude thereof is calculated, it becomes invariant to the rotation or enlargement / reduction of the original pattern, as is already known.

The processing from the conversion A to the processing up to this point will be called conversion B. The amplitude obtained by the conversion B is the matching unit 15
Sent to. Also, the phase of the Fourier transform obtained by the conversion B is stored in the memory for use in the subsequent processing.

The matching unit 15 stores the standard patterns which have been subjected to the above-mentioned conversion in advance, and performs template matching between these standard patterns and the patterns sent from the conversion unit 14 to determine the pattern. I do. As a result of the processing in the conversion unit 14, even if the presented pattern is displaced from the standard pattern, scaled or rotated, the matching unit can correctly recognize it, but in general, it is displayed in the display pattern. Since background noise is superimposed on patterns other than the target pattern, it is difficult to perform accurate recognition with such processing. Therefore, if the similarity obtained by the template matching described above is better than a preset threshold value, the matching unit 15 ends the process with the result as a determination result. The standard pattern having a high degree of similarity is fed back to the inverse conversion unit 16 as a candidate pattern.

The inverse transformation unit 16 reverses the processing of the transformation unit 14 to calculate a feedback image. That is, first, an inverse Fourier transform is calculated from the amplitude of the candidate pattern fed back from the matching unit 15 and the phase calculated by the transform B in the transform unit 14, and s−θ
Calculate the feedback pattern on the coordinate axes. further,
On the other hand, the inverse transformation of the coordinate transformation expressed by Equation 1
The pattern P '(Kx, Ky) is calculated. Further, an inverse Fourier transform is performed again from the feedback amplitude P ′ thus obtained and the phase calculated by the conversion A in the conversion unit 14 to obtain the feedback image I ′. In the above inverse transform processing, threshold processing may be added after each inverse Fourier transform processing in order to prevent the amplitude or the intensity which should originally be positive from becoming negative or becoming too large.

The image memory unit 13 corrects the input image I ₀ on the basis of the feedback image I ′ sent from the inverse conversion unit 16, for example, by the following equation, and the corrected image I to the conversion unit 14 again. And send.

[Number 2] I (x, y) = ( I 0 (x, y) a + (1-a)) · (I '(x, y) b + (1-b)) x, y is the position of the pixel It is a specified parameter.

After this conversion, appropriate thresholding may be added to prevent the values from becoming too large. In Equation 2, a and b are parameters that determine the weights of the input image and the feedback image, respectively, and 0
Set between 1 and 1. Now, it is assumed that the input image is a translation, scaling, or rotation of any of the standard patterns, and there is no background noise. In this case, since the determination result is obtained in the first matching process, the inverse conversion process is not performed. However, assuming that the inverse conversion process is performed, in this case, the process performed by the inverse conversion unit 16 is accurately converted. This is the inverse conversion of the processing in the unit 14, and the feedback image exactly matches the input image subjected to the preprocessing.

Even when noise is superimposed on the input image,
Approximately, the feedback image is the input image with the noise part removed. Therefore, for example, in Equation 2, a
Setting b and b to 1 results in significant suppression of noise from the modified image. This is the number 2
Even if the pixel has a non-zero value in the input image, if the pixel is in the noise portion, the feedback signal I '
This is because it is zero because is zero. When a and b are set to 1, a process for taking a square root is added after the formula 2. As a result, if the input image and the feedback image are equal, the modified image will be equal.

The conversion of the conversion unit 14 and the matching process of the matching unit 15 are repeated for the image I thus modified. This loop is repeated until the similarity by matching exceeds a predetermined threshold value, but if sufficient similarity is not obtained even after reaching the predetermined number of iterations, the first candidate pattern is If it is not correct, change the candidate pattern to the next one and continue the same operation.

FIG. 2 is a block diagram showing an embodiment of the second invention. The embodiment shown in FIG. 2 includes an image input unit 21 using an image scanner or a television camera,
A pre-processing unit 22 that digitizes data and performs necessary pre-processing such as edge enhancement, an image memory 23 that stores digital data, and a multi-stage hierarchical Fourier transform to form a pattern that is invariable with respect to positional deviation. A conversion unit A24 that converts an input, a conversion unit B25 that performs an extended Melan transform to convert an input into a pattern that is invariant to rotation and scaling, and template matching between an output of the conversion unit B25 and a standard pattern. And the matching unit 26 that performs the pattern determination and the candidate pattern determination.
An inverse transform unit B27 that calculates the inverse transform of the transform unit B25 from the candidate pattern determined in 6 and the data of the transform unit B25;
An inverse transform unit A28 that calculates a feedback image based on the data from the inverse transform unit B27 and the data from the transform unit A24 is provided.

Next, the operation of the second embodiment of the invention will be described. The operation of the embodiment of the second invention is almost the same as the operation of the first invention, but the amplitude calculation processing of the first Fourier transform in the embodiment of the first invention is divided into the following multi-stage processing. The difference is that it is replaced with. In addition, the inverse conversion unit is also divided into multiple stages corresponding to this.

Similar to the first embodiment of the present invention, the image information input from the image input section 21 is subjected to necessary preprocessing by the preprocessing section 22 and stored in the memory. In the second embodiment of the invention, the conversion A portion of the processing in the conversion unit in the first invention is replaced with a hierarchical Fourier transform divided into multiple stages. That is, first, in the conversion unit A, the preprocessed image is divided into a large number of small regions, and the Fourier transform of the image and its amplitude are calculated for each small region.
These values are saved in the memory in the conversion unit A for use in the subsequent processing. This processing is expressed in the following formula.

Now, let the number of pixels of the input image be N × N. First, this input image is represented by n ₂ consisting of the number of pixels n ₁ × n _1.
It is divided into × n ₂ small areas. Here, when the images are divided so that they do not overlap each other, N = n ₁
Although it is × n ₂ , they may be divided so that they overlap each other.

Next, each image divided into these small regions is subjected to the Fourier transform and its amplitude determination process as follows. Note that, hereinafter, X ^* represents an X vector, x ^* represents an x vector, and k ^* represents a k vector.

Here, each sub-region is assigned a subscript X ^* = (X,
Y) (X, Y = 1 to n ₂ ) and the position of the pixel in each small area is subscript x ^* = (x, y) (x, y = _{1 to} n ₁ ).
It is represented by. Further, the value of the pixel at the position of x ^{* in} the small area represented by X ^* is represented by I (x ^* ; X ^* ).

[0031]

[Equation 3]

[0032]

[Equation 4]

W is a window function. If W = 1, number 3
It is, I; represents a (x ^* X ^*) of the subscript x ^* regarding two-dimensional discrete values Fourier transform. In order to eliminate the discontinuity at the boundary of the small area, a window function that becomes zero at the boundary, for example, Gaussian or Hamming window may be used as W.

For the calculation of the equation 3, it is possible to use the fast Fourier transform algorithm. Number 4 P
1 (k ^* 1, X ^* ) is the amplitude of the Fourier transform. It is the same even if the power which is the square of the amplitude is calculated instead of the amplitude for convenience of calculation. The pattern P1 (k ^* 1, X ^* ) obtained by the above process is approximately unchanged if the magnitude of the positional deviation is equal to or smaller than the size of the small area even if the position of the input image is shifted. The phase of the Fourier transform calculated here is stored in the memory for use in the subsequent processing.

Next, the conversion unit A24 outputs P1 (k ^* 1, X
^* ) Is again subjected to Fourier transform with respect to the position variable X ^* , and its amplitude is calculated.

[0036]

[Equation 5]

[0037]

[Equation 6]

The phase of the Fourier transform obtained here is also stored in the memory for use in the subsequent processing. The processing described so far is a two-step processing, but if necessary, 3
It is also possible to carry out multi-stage processing of more than steps. The conversion unit A24 further performs the following variable conversion.

[0039]

[Equation 7]

The pattern P (k ^* ) thus obtained has a property similar to the amplitude of the Fourier transform of the entire image calculated in the embodiment of the first invention, and does not change even when the input image is translated. is there.

Next, P is sent to the conversion unit B25, where the first
The process of the conversion B in the embodiment of the invention is performed, and the pattern is converted into a pattern that is invariable even when the pattern is rotated or enlarged or reduced. The function of the matching unit 26 is similar to that of the first embodiment of the invention. The inverse conversion unit is also divided into multiple stages corresponding to the division of the conversion unit into multiple stages. First, the inverse transform unit 27 performs the inverse Fourier transform as in the first embodiment of the invention.
The inverse coordinate transformation is performed to calculate the pattern P '(k ^* ). Next, the inverse conversion unit A28 performs the inverse conversion of the conversion of the expression 7 to obtain P
After calculating 2 ′ (k ^* 1, k ^* 2), an inverse Fourier transform is performed using this P2 ′ and the phase calculated by the conversion unit A24 to obtain P1 ′ (k ^* 1, X ^* ). calculate. Further, the inverse transform unit A28 performs an inverse transform using this P1 'and the phase calculated by the first local Fourier transform in the transform unit A24 to calculate the feedback image I'.

The image memory unit 26 corrects the input image based on this feedback image as in the first embodiment of the present invention, and sends it again to the conversion unit A24. This loop is repeated until the similarity by matching exceeds a predetermined threshold value, and if sufficient similarity is not obtained even after reaching the predetermined number of repetitions, the next rank The point that the candidate pattern is changed to one and the same operation is continued is also the same as the first embodiment of the present invention.

In this embodiment, the inverse transformation up to the feedback image I'is performed to correct the input. However, only the inverse transformation up to P1 '(k ^* 1, X ^* ) in the inverse transformation unit A28 is performed. It is also possible to perform the correction processing on P1 (k ^* 1, X ^* ) calculated from the input image without using it and to loop the matching processing and the correction processing. Further, a variation is possible in which the loop up to P1 is rotated several times to stabilize P1 ', and then the loop up to the feedback image I'is rotated.

In the first embodiment of the present invention, since the Fourier transform of the entire image is first performed, noise suppression may be difficult when background noise has the same spatial frequency component as the target pattern. is there. In the second embodiment of the present invention, since the spatial frequency distribution P1 (k ^* 1, X ^* ) for each location is calculated as an intermediate representation of the transformation, even if noise has the same spatial frequency component as the pattern. Even if it is, it can be suppressed if it is located away from the pattern.

FIG. 3 is a block diagram showing an embodiment of the third invention. The embodiment shown in FIG. 3 includes an image input unit 31 using an image scanner or a television camera,
A preprocessing unit 32 that digitizes the data and performs necessary preprocessing such as edge enhancement, a conversion unit C33 that converts the input image into a pattern on the coordinate axis that has undergone complex logarithmic conversion, and an output pattern of the conversion unit C33. Image memory 34 for storing
And a conversion unit D35 that performs a multi-stage hierarchical Fourier transform on this pattern and converts the input into a pattern that is approximately invariant to small displacements, rotations, and scaling changes.
And a matching unit 36 that performs pattern determination and candidate pattern determination by performing template matching between the output of the conversion unit D35 and the standard pattern, and the conversion unit D35 from the candidate pattern determined by the matching unit 36 and the data of the conversion unit D35.
And an inverse conversion unit D37 for performing the inverse conversion of

Next, the operation of the embodiment of the third invention will be described. Similar to the first embodiment of the invention, the image input unit 3
The image information input from 1 undergoes necessary preprocessing in the preprocessing unit 32. Next, the conversion unit C33, with respect to the pattern I (x, y) sent from the preprocessing unit, in the case of w = Log (z + a) x> 0 w = Log (z-a) x <0 Case; z = x + iy, w = u + iv, a: positive constant,
Coordinate transformation represented by complex logarithmic transformation of i = √−1 (x, y) → (u,
v), and change to a pattern on the uv axis. Here, a is a positive constant for preventing the logarithmic function from diverging at the origin. Due to the well-known nature of logarithmic transformation, the pattern obtained by this process will approximately translate when the original image is scaled or rotated. The pattern processed by the conversion unit C33 is sent to and stored in the image memory 34.

Next, the conversion unit D35 divides this pattern into a large number of small areas, and first calculates the Fourier transform of the image and its amplitude for each small area. Here, if the position of the original image is shifted, the image in each small area also moves, but if the size of the positional deviation in each small area is smaller than the size of the small area, such processing is performed. The pattern obtained by is approximately invariant.

Further, in the transforming section D35, the amplitudes of the local Fourier transforms thus obtained at a large number of positions in the image are newly regarded as a pattern, and the Fourier transform on the position variable is performed again. Calculate the amplitude. Mathematical Expression 3 in the embodiment of the second invention except that this processing is performed on the pattern subjected to the complex logarithmic transformation.
~ It is exactly the same as the processing of Expression 6.

The pattern obtained by such processing has approximately the same property as the amplitude of the Fourier transform of the input image represented by the coordinate axis subjected to the complex logarithmic transformation, and therefore the pattern is translated on this coordinate axis. Even if the original input image is enlarged or reduced or rotated, it is approximately unchanged. Therefore, by the above processing, a pattern that is almost invariable with respect to a small displacement, scaling conversion, and rotation can be obtained.

Next, in the matching section 36, the first
Similar to the invention described above, the template matching is performed with the standard pattern which has been subjected to the same processing as the processing so far. The inverse conversion unit D37 is also the matching unit 36.
The feedback image is calculated based on the candidate pattern determined in. The process of the inverse transform unit D37 is exactly the same as the process of calculating the feedback image from P2 'in the inverse transform unit A28 in the embodiment of the second invention.

The image memory unit 34 corrects the input image based on this feedback image as in the first embodiment of the invention, and sends it again to the conversion unit D35. This loop is repeated until the similarity due to matching exceeds a predetermined threshold value, and if sufficient similarity is not obtained even after reaching the predetermined number of repetitions, the next rank The point that the candidate pattern is changed to one and the same operation is continued is the same as the embodiment of the first invention.

As described in the variation of the second embodiment of the present invention, the inverse transform of all the multi-stage Fourier transforms is not performed, but the correction process is performed by using the result of the inverse transform process up to the middle, and the matching process is performed. It is also possible to rotate the correction processing loop. Further, the point that the process of combining the loop of the correction process by returning to the feedback image can be performed is also the same as the embodiment of the second invention.

[0053]

As described above, according to the present invention, it is possible to perform pattern recognition processing that is strong against enlargement / reduction of patterns, rotation, and positional displacement, and background noise is superimposed on the presented pattern. Also in this case, matching with the standard pattern is significantly improved, and it is possible to perform highly accurate pattern recognition processing.

[Brief description of drawings]

FIG. 1 is a configuration diagram showing an embodiment of a first invention.

FIG. 2 is a configuration diagram showing an embodiment of a second invention.

FIG. 3 is a configuration diagram showing an embodiment of a third invention.

[Explanation of symbols]

 11, 21, 31 image input unit 12, 22, 32 pre-processing unit 13, 23, 34 image memory 14 conversion unit 15, 26, 36 matching unit 16 inverse conversion unit 24 conversion unit A 25 conversion unit B 27 inverse conversion unit B 28 Inverse Transform Unit A 33 Transform Unit C 35 Transform Unit D 37 Inverse Transform Unit D

Claims (3)

[Claims] 1. A first conversion unit for performing a Fourier-Melan transform on an input image to convert the image into a pattern invariant to displacement, enlargement / reduction, and rotation operations; and a first conversion unit. A matching unit that determines a candidate pattern by comparing the obtained pattern with a standard pattern; a first inverse transform unit that performs an inverse transform of the transform performed by the first transform unit on the candidate pattern; An image signal processing apparatus, comprising: an image memory for removing background noise unrelated to a pattern from an input image based on the image, and performing pattern determination by repeating loop processing. 2. A second method in which an input image is divided into small regions, multi-stage hierarchical Fourier transform is performed, and the image is converted into a pattern invariant to displacement by a process of obtaining the amplitude of the Fourier transform of the image. A conversion unit; a third conversion unit that performs extended Melan transform on the pattern obtained by the second conversion unit to convert it into a pattern that is invariant to scaling and rotation operations; The inverse conversion of the third conversion unit is calculated from the matching unit that determines the candidate pattern by comparing the pattern obtained from the conversion unit with the standard pattern, and the candidate pattern determined by the matching unit and the data from the third conversion unit. A second inverse transform unit, a third inverse transform unit that calculates a feedback image from the data from the second inverse transform unit and the data from the second transform unit, and regardless of the pattern from the input image based on the feedback image Background And a image memory to remove noise, the processor of the image signal and performs pattern judgment by repeating the loop process. 3. A fourth transforming unit for subjecting an input image to coordinate transformation represented by complex logarithmic transform, and a pattern obtained by the fourth transforming unit divided into small regions to form a multi-stage hierarchical Fourier transform. And a pattern obtained by the fifth transforming unit, which transforms the image into a pattern that is invariant to small displacement, enlargement / reduction, and rotation operation by the process of obtaining the amplitude of the Fourier transform of the image. And a matching unit for determining a candidate pattern by comparing with a standard pattern, and a fourth inverse transform for computing a feedback image by subjecting the candidate pattern to the inverse transform of the multistage hierarchical Fourier transform and the amplitude calculation of the Fourier transform. And an image memory that removes background noise irrelevant to the pattern from the input image based on the feedback image, and an image signal characterized by performing pattern determination by repeating loop processing. Processing apparatus.