JP2000222578A - Pattern matching method and detection of movement vector - Google Patents

Abstract

PROBLEM TO BE SOLVED: To make executable a pattern matching and movement vector detection, which use Fourier transformation, with a small operation volume. SOLUTION: An input block f(k, l) is extended to the size of reference search range data g'(k, l) to obtain an input extension data f'(k, l) (step S1), and this input extension data and reference search range data are subjected to Fourier transformation (steps S2 and S3). An evaluation function e(m, n) is obtained based on F(K, L) and G(K, L) as the result of Fourier transformation and the square sum (step S6) of plural candidate block data (step S7). This evaluation function is used to discriminate the degree of similarity of a reference block (steps S8 and S9), thus reducing the operation volume.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to pattern matching of audio data and image data, and more particularly, to motion vector detection and method in image coding.

[0002]

2. Description of the Related Art When a moving image is communicated or recorded, the image is conventionally compressed (encoded) in order to reduce the bandwidth of a communication means and the capacity of a recording medium. H.264 is an international standard video coding method. 261, MPEG
1, MPEG2, and the like. These conventional coding schemes include motion vector detection and motion compensation, DCT (Dis
create Cosine Transform, quantization, zigzag scan, VLC (Variable
Length Coding).

In the above-mentioned coding method, an input image is
The image data is divided into units called macroblocks each consisting of 6 pixels × 16 pixels, and compression processing is sequentially performed in this unit. That is, a macroblock consisting of 16 pixels in the vertical direction and 16 pixels in the horizontal direction is used as a unit for motion vector detection and motion compensation.

Here, the motion vector detection will be described. At the time of motion vector detection, reference is made to a target macroblock 202 from a reference search range 205 in a reference image 203 shown in FIG. 2B in the input image 201 shown in FIG. Block 206 is determined.

[0005] As the reference search range 205, for example, in the reference image 203, an area around a block 204 at a position corresponding to the macroblock 202 of the input image 201 is assigned. When the reference block 206 is determined, a method of calculating the similarity for each of the reference candidate blocks extracted from the reference search range 205 and selecting the most similar reference candidate block as the reference block 206 is used. Used.

In order to calculate the similarity, a sum of absolute values of differences between pixel data and a sum of squares of the differences are used.
The motion vector is a vector indicating a relative position between the input block and the reference block. The difference data is constructed by subtracting the data of the reference block from the processing target block,
Data compression is realized by performing a two-dimensional DCT or the like by dividing the data into a unit of pixel × eight pixels.

Here, as an example of a conventional motion vector detecting method, a technique described in Japanese Patent Laid-Open No. 7-327226 will be briefly described.

First, a two-dimensional fast Fourier transform (hereinafter referred to as F) is performed on the reference search range data (reference block) and the input extension data (array T of target blocks).
FT: Fast Fourier Transform
), A first multiplication of these results is performed by a multiplier, and a first calculation result is obtained by performing an inverse two-dimensional FFT on the multiplication result.

Next, a two-dimensional FFT is performed on the reference search range square data (reference block R1) and coefficient data (target block array T1) obtained by squaring the reference search range data, and the data is multiplied by a multiplier. The second of the results of
To obtain a second calculation result.

Then, the subtractor subtracts 2 of the first calculation result.
Calculate the difference between the double and the second calculation result to calculate the square error,
A similar block is selected by searching for the minimum value of the square error, and a motion vector is obtained from its relative coordinates.

General one-dimensional FF with N = 2 ^M data number
The number of complex multiplications required for T is

[0012]

(N / 2) log ₂ N = MN / 2, where one complex multiplication corresponds to four real multiplications, and N × N two-dimensional FFT is N times one-dimensional FFT in the x direction. And N one-dimensional FFTs in the y direction.

Therefore, the actual number of multiplications required for an N × N two-dimensional FFT is

[0014]

## EQU2 ## ((MN / 2) .N + (MN / 2) .N) .4
= 4MN ² .

Therefore, in the case of N × N reference search range data, two-dimensional FFT of reference search range data: 4MN ² two-dimensional FFT of input extended data: 4MN ² first multiplication: 4N ² inverse two-dimensional FFT: 4MN ² the square of the reference search area data: N ² reference search range squared data FFT: 4 mN ² 2 dimensional FFT coefficient data: 4 mN ² second multiplication: a 4N ^2.

Here, twice the first calculation result is ignored because it can be realized by a shift operation. Among them, the two-dimensional FFT of the coefficient data may be executed once with the same data, and if it is ignored, the actual number of multiplications becomes

[0017]

Than [number 3] ^{4MN 2 · 4 + 4N 2 ·} 2 + N 2, the (16MN ² + 9N ²⁾ times. As an example, N =
32 and M = 5, the actual number of multiplications is 91136
Times.

[0018]

As described above, in the conventional motion vector detection, the fast Fourier transform of the reference search range square data and coefficient data and the multiplication of these results are performed. Is required.

For this reason, a sufficient motion vector search range cannot be obtained under a limited arithmetic unit, and the coding efficiency is deteriorated. Therefore, the coding bit rate increases and the image quality deteriorates. In order to obtain a sufficient motion vector range, the scale of the arithmetic unit becomes large.

An object of the present invention is to provide a pattern matching method and a motion vector detecting method which require a small amount of calculation in order to solve the above problems.

[0021]

In order to achieve the above object, a first pattern matching method according to the present invention uses a P-dimensional template data to define P as a natural number.
A pattern matching method for performing pattern matching of dimensional target data, wherein the template data is
Constructing template extension data expanded to the size of the target data using zero data, performing a P-dimensional fast Fourier transform between the template extension data and the target data, and obtaining a template Fourier transform result and a target Fourier transform result. And performing a complex multiplication of one of the complex conjugate and the other of the template Fourier transform result and the target Fourier transform result to obtain a multiplication result, and performing an inverse fast Fourier transform of the multiplication result. Obtaining a P-dimensional Fourier transform result, calculating a sum of squares of the plurality of candidate data in the target data, and using the inverse P-dimensional Fourier transform result and the sum of squares of the plurality of candidate data And performing a similarity determination using the same.

According to this method, two or more candidate data
Since the similarity is calculated using the sum of squares and the similarity is determined using the calculation result, the fast Fourier transform of the reference search range squared data and the coefficient data, and the multiplication of these results, which were conventionally required. Can be omitted. Thus, a pattern matching method with a small amount of calculation can be provided.

In order to achieve the above object, the second pattern matching method of the present invention is as follows.
A pattern matching method for performing pattern matching of P-dimensional target data using P-dimensional template data, comprising: forming template extension data obtained by expanding the template data to the size of the target data by using 0 data; , Q is a natural number smaller than the P, the template extension data and the target data are scanned to perform Q-dimensional fast Fourier transform on the result as Q-dimensional data, and the template Fourier transform result and the target Fourier transform result are Obtaining, performing one complex conjugate and the other complex multiplication of the template Fourier transform result and the target Fourier transform result, and obtaining a multiplication result; performing an inverse fast Fourier transform of the multiplication result; Obtaining a Fourier transform result; Calculating a sum of squares of a plurality of candidate data in the above, and performing a similarity determination using the inverse Q-dimensional Fourier transform result and the sum of squares of the plurality of candidate data. And

According to this method, two or more candidate data
Since the similarity is calculated using the sum of squares and the similarity is determined using the calculation result, the fast Fourier transform of the reference search range squared data and the coefficient data, and the multiplication of these results, which were conventionally required. Can be omitted. Thus, a pattern matching method with a small amount of calculation can be provided.

In order to achieve the above object, a third pattern matching method according to the present invention provides that if P is a natural number,
A pattern matching method for performing pattern matching of P-dimensional target data using P-dimensional template data, wherein the target data is divided, and one of the real data and the other is imaginary to form target complex data. Using the result of dividing the template data and 0 data to expand to the field-divided size of the target data, forming template complex data with one as a real part and the other as an imaginary part; Performing a P-dimensional fast Fourier transform of the target complex number data with the target complex number data to obtain a template Fourier transform result and a target Fourier transform result; and converting one of the template Fourier transform result and the target Fourier transform result into a first Fourier transform result And the other as the result of the second Fourier transform, Separating a first divided Fourier transform result and a second divided Fourier transform result corresponding to each division of the Rier transform result, and one complex conjugate of the second Fourier transform result and the first divided Fourier transform result And the other complex multiplication to obtain a first multiplication result. The second multiplication result is obtained by performing one complex conjugate and the other complex multiplication on the second Fourier transform result and the second divided Fourier transform result. And performing an inverse P-dimensional fast Fourier transform of the first and second multiplication results;
Obtaining first and second inverse Fourier transform results; separating the first and second inverse Fourier transform results into a real part and an imaginary part; Obtaining data, second real part data, and second imaginary part data; calculating a sum of squares of a plurality of candidate data in the target data; Performing a similarity determination based on the first imaginary part data, the second real part data, the second imaginary part data, and the sum of squares of the plurality of candidate data. It is characterized by the following.

According to this method, a plurality of candidate data 2
Since the similarity is calculated using the sum of squares and the similarity is determined using the calculation result, the fast Fourier transform of the reference search range squared data and the coefficient data, and the multiplication of these results, which were conventionally required. Can be omitted. Furthermore, dividing the target data to form the target complex number data, performing the fast Fourier transform, separating one fast Fourier transform result, and multiplying the result by the other fast Fourier transform result, and performing an inverse fast Fourier transform. Accordingly, the number of data of the fast Fourier transform and the inverse fast Fourier transform can be reduced. Thus, a pattern matching method with a small amount of calculation can be provided.

In order to achieve the above object, the fourth pattern matching method of the present invention is as follows.
A pattern matching method for performing pattern matching of P-dimensional target data using P-dimensional template data, wherein the target data is divided to form target complex number data using one as a real part and the other as an imaginary part; Using the result of dividing the template data and 0 data to expand the target data to the field-divided size to form template complex data with one as a real part and the other as an imaginary part; Q is smaller than P Assuming that the number is a natural number, the template complex data and the target complex data are scanned to perform Q-dimensional fast Fourier transform on a result obtained as Q-dimensional data to obtain a template Fourier transform result and a target Fourier transform result; One of the Fourier transform result and the target Fourier transform result is assigned to the first file. Separating the first Fourier transform result into a first divided Fourier transform result and a second divided Fourier transform result corresponding to each division of the first Fourier transform result; and the second Fourier transform result. One complex conjugate and the other complex multiplication are performed on the conversion result and the first divided Fourier transform result to obtain a first multiplied result, and one of the second Fourier transform result and the second divided Fourier transform result is obtained. Obtaining a second multiplication result by performing complex conjugate and the other complex multiplication, and performing an inverse Q-dimensional fast Fourier transform of the first and second multiplication results to obtain first and second inverse Fourier transform results And separating the first and second inverse Fourier transform results into a real part and an imaginary part. The first real part data, the first imaginary part data, the second real part data, Two Obtaining partial data; calculating a sum of squares of a plurality of candidate data in the target data; the first real part data; the first imaginary part data; Performing a similarity determination based on the real part data, the second imaginary part data, and the sum of squares of the plurality of candidate data.

According to this method, two or more candidate data
Since the similarity is calculated using the sum of squares and the similarity is determined using the calculation result, the fast Fourier transform of the reference search range squared data and the coefficient data, and the multiplication of these results, which were conventionally required. Can be omitted. Furthermore, dividing the target data to form the target complex number data, performing the fast Fourier transform, separating one fast Fourier transform result, and multiplying the result by the other fast Fourier transform result, and performing an inverse fast Fourier transform. Accordingly, the number of data of the fast Fourier transform and the inverse fast Fourier transform can be reduced. Thus, a pattern matching method with a small amount of calculation can be provided.

In order to achieve the above object, a first motion vector detecting method according to the present invention is a motion vector detecting method for detecting a relative position of a block similar to an input block of an input image in reference search range data of a reference image. A vector detection method, comprising the steps of: forming input extension data obtained by expanding the input block to the size of the reference search range data using 0 data; and two steps of the input extension data and the reference search range data. Performing a two-dimensional fast Fourier transform to obtain an input Fourier transform result and a reference Fourier transform result, and performing one complex conjugate and the other complex multiplication on the input Fourier transform result and the reference Fourier transform result, and obtaining a multiplication result. And performing an inverse two-dimensional fast Fourier transform of the multiplication result; Calculating a sum of squares of the candidate block data, characterized by comprising the step of performing a similarity determination by the square sum of the results and the plurality of candidate block data of said inverse two-dimensional fast Fourier transform.

According to this method, the similarity is calculated by using the sum of squares of the plurality of candidate block data, and the similarity is determined by using the calculation result. Fast Fourier transform of the multiplication data and coefficient data and multiplication of these results can be omitted. Thus, a motion vector detection method with a small amount of calculation can be provided.

In the first motion vector detection method, in the two-dimensional fast Fourier transform of the reference search range data, a result of the one-dimensional fast Fourier transform in a certain direction of the reference search range data corresponding to a certain input block is used. Is preferably used in two-dimensional fast Fourier transform of reference search range data corresponding to the input block of.

According to this method, the number of one-dimensional Fourier transforms can be limited to one for a portion where reference search range data corresponding to motion vector detection of a plurality of input blocks overlaps. Can be further reduced.

To achieve the above object, a second motion vector detecting method according to the present invention is a motion vector detecting method for detecting a relative position of a block similar to an input block of an input image in reference search range data of a reference image. A vector detection method, comprising: constructing input extended data obtained by extending the input block to the size of the reference search range data using 0 data; and scanning the input extended data and the reference search range data. Performing a one-dimensional fast Fourier transform of the result obtained as one-dimensional data to obtain an input Fourier transform result and a reference Fourier transform result; and obtaining one complex conjugate of the input Fourier transform result and the reference Fourier transform result. Performing the other complex multiplication to obtain a multiplication result; and performing an inverse one-dimensional fast Fourier transform of the multiplication result; Calculating a square sum of a plurality of candidate blocks data in the serial reference search range data,
Performing a similarity determination based on a result of the inverse one-dimensional fast Fourier transform and a sum of squares of the plurality of candidate block data.

According to this method, the similarity is calculated using the sum of squares of a plurality of candidate block data, and the similarity is determined using the calculation result. Fast Fourier transform of the multiplication data and coefficient data and multiplication of these results can be omitted. Thus, a motion vector detection method with a small amount of calculation can be provided.

To achieve the above object, a third motion vector detecting method according to the present invention is a motion vector detecting method for detecting a relative position of a block similar to an input block of an input image in reference search range data of a reference image. A vector detection method, wherein the reference search range data is divided into fields, one of which is a real part and the other is an imaginary part to form reference complex data, and the result of field division of the input block and 0 data are used. Extending the reference search range data to a field-divided size, and constructing input complex data with one as a real part and the other as an imaginary part; and a two-dimensional high-speed processing of the input complex data and the reference complex data. Performing a Fourier transform to obtain an input Fourier transform result and a reference Fourier transform result; and One of the Fourier transform result first
Separating the first Fourier transform result into a first field Fourier transform result and a second field Fourier transform result corresponding to each field of the first Fourier transform result; One complex conjugate and another complex multiplication are performed on the Fourier transform result and the first field Fourier transform result to obtain a first result.
Obtaining a second multiplication result by performing one complex conjugate and another complex multiplication on the second Fourier transform result and the second field Fourier transform result, and obtaining the first and second multiplication results. Performing an inverse two-dimensional fast Fourier transform of the multiplication result to obtain first and second inverse Fourier transform results; separating the first and second inverse Fourier transform results into a real part and an imaginary part; 1, real part data, first imaginary part data, second real part data, and second real part data.
Obtaining imaginary part data, calculating a sum of squares of a plurality of candidate block data in the reference search range data, the first real part data, and the first imaginary part data. Determining a similarity based on the second real part data, the second imaginary part data, and the sum of squares of the plurality of candidate block data.

According to this method, the similarity is calculated using the sum of squares of a plurality of candidate block data, and the similarity is determined using the calculation result. Fast Fourier transform of the multiplication data and coefficient data and multiplication of these results can be omitted. Further, the reference search range data is divided to form reference complex number data, and the result is subjected to fast Fourier transform.The result of one fast Fourier transform is separated and the result obtained by multiplying the result of the other fast Fourier transform is inverse fast Fourier transformed. By doing
The number of data between the fast Fourier transform and the inverse fast Fourier transform can be reduced. Thus, a pattern matching method with a small amount of calculation can be provided.

The third motion vector detection method is characterized in that in the two-dimensional fast Fourier transform of the reference complex data,
It is preferable to use the result of the one-dimensional fast Fourier transform in a certain direction of the reference complex data corresponding to a certain input block in the two-dimensional fast Fourier transform of the reference complex data corresponding to another input block.

According to this method, the number of one-dimensional fast Fourier transforms can be limited to one for portions where reference search range data corresponding to motion vector detection of a plurality of input blocks overlap. This makes it possible to further reduce the amount of calculation.

To achieve the above object, a fourth motion vector detecting method according to the present invention is a motion vector detecting method for detecting a relative position of a block similar to an input block of an input image in reference search range data of a reference image. A vector detection method, wherein the reference search range data is field-divided, one of which is a real part, and the other is an imaginary part to form reference complex data, and the result of field-dividing the input block and 0 data is used. Expanding the reference search range data to a field-divided size by using one as a real part and the other as an imaginary part to form input complex data, and scanning the input complex data and the reference complex data to obtain one. Performing a one-dimensional fast Fourier transform of the result as the dimensional data to obtain an input Fourier transform result and a reference Fourier transform result; One of the input Fourier transform result and the reference Fourier transform result is defined as a first Fourier transform result, the other is defined as a second Fourier transform result, and the first field Fourier transform result and the second Fourier transform result corresponding to each field of the first Fourier transform result. Separating the two-field Fourier transform result into two-field Fourier transform results;
One complex conjugate and another complex multiplication are performed on the field Fourier transform result and the other complex multiplication to obtain a first multiplication result, and one complex conjugate and the other complex conjugate are obtained on the second Fourier transform result and the second field Fourier transform result. Performing a multiplication to obtain a second multiplication result; performing an inverse one-dimensional fast Fourier transform of the first and second multiplication results to obtain first and second inverse Fourier transform results; And the second inverse Fourier transform result is separated into a real part and an imaginary part, and the first real part data, the first imaginary part data, the second real part data, the second imaginary part data , Calculating a sum of squares of a plurality of candidate block data in the reference search range data, the first real part data, the first imaginary part data, and the second imaginary part data. Real part data and previous A second imaginary-part data, characterized by comprising the step of performing a similarity determination by the sum of squares of the plurality of candidate block data.

According to this method, the similarity is calculated by using the sum of squares of the plurality of candidate block data, and the similarity is determined using the calculation result. Fast Fourier transform of the multiplication data and coefficient data and multiplication of these results can be omitted. Further, the reference search range data is divided to form reference complex number data, and the result is subjected to fast Fourier transform.The result of one fast Fourier transform is separated and the result obtained by multiplying the result of the other fast Fourier transform is inverse fast Fourier transformed. By doing
The number of data between the fast Fourier transform and the inverse fast Fourier transform can be reduced. Thus, a pattern matching method with a small amount of calculation can be provided.

In order to achieve the above object, a fifth motion vector detecting method according to the present invention is characterized in that an input block of an input image and a result obtained by field-dividing the input block in the reference search range data of the reference image are respectively used. A motion vector detection method for detecting a relative position of a similar block, wherein the reference search range data is divided into fields, one of which is a real part, and the other is an imaginary part, to constitute reference complex number data; and Using the field-divided result and 0 data to extend the reference search range data to a field-divided size, and constructing input complex data with one as a real part and the other as an imaginary part; And two-dimensional fast Fourier transform of the reference complex data and the input Fourier transform result and the reference Fourier transform result Obtaining a, one of the reference Fourier transform result and the input Fourier transform result first
Separating a first Fourier transform result into a first field Fourier transform result and a second field Fourier transform result corresponding to each field of the first Fourier transform result; One complex conjugate and the other complex multiplication are performed on the conversion result and the first field Fourier transform result to obtain a first multiplication result, and one of the second Fourier transform result and the second field Fourier transform result is obtained. Performing a complex conjugate and the other complex multiplication to obtain a second multiplication result, performing an inverse two-dimensional fast Fourier transform of the first and second multiplication results, and calculating the first and second inverse Fourier transform results Obtaining the first and second inverse Fourier transform results into a real part and an imaginary part;
Obtaining the first imaginary part data, the second real part data, and the second imaginary part data; and performing a plurality of non-field-divided candidate block data or field division in the reference search range data. Calculating a sum of squares of the plurality of candidate block data, the first real part data, the first imaginary part data, the second real part data, and the second imaginary part data; Performing a similarity determination based on the sum of squares of the plurality of candidate block data.

According to this method, the similarity is calculated by using the sum of squares of the plurality of candidate block data, and the similarity is determined by using the calculation result. Fast Fourier transform of the multiplication data and coefficient data and multiplication of these results can be omitted. Further, the reference search range data is divided to form reference complex number data, and the result is subjected to fast Fourier transform.The result of one fast Fourier transform is separated and the result obtained by multiplying the result of the other fast Fourier transform is inverse fast Fourier transformed. By doing
The number of data between the fast Fourier transform and the inverse fast Fourier transform can be reduced. Thus, a pattern matching method with a small amount of calculation can be provided.

In the fifth motion vector detecting method, the two-dimensional fast Fourier transform of the reference complex data may include:
It is preferable to use the result of the one-dimensional fast Fourier transform in a certain direction of the reference complex data corresponding to a certain input block in the two-dimensional fast Fourier transform of the reference complex data corresponding to another input block.

According to this method, the number of one-dimensional fast Fourier transforms can be limited to one for portions where reference search range data corresponding to motion vector detection of a plurality of input blocks overlap. This makes it possible to further reduce the amount of calculation.

In order to achieve the above object, a sixth motion vector detecting method according to the present invention is characterized in that an input block of an input image and a result of field division of the input block in the reference search range data of the reference image are respectively used. A motion vector detection method for detecting a relative position of a similar block, wherein the reference search range data is divided into fields, one of which is a real part, and the other is an imaginary part, to constitute reference complex number data; and Expanding the reference search range data to the field-divided size using the field-divided result and 0 data, and configuring input complex data as one real part and the other as an imaginary part; A one-dimensional fast Fourier transform of the result obtained by scanning the reference complex number data into one-dimensional data is performed. Obtaining a conversion result and a reference Fourier transform result; and setting one of the input Fourier transform result and the reference Fourier transform result as a first Fourier transform result, the other as a second Fourier transform result, and calculating the first Fourier transform result. Separating each of the fields into a first field Fourier transform result and a second field Fourier transform result; one complex conjugate and the other complex conjugate of the second Fourier transform result and the first field Fourier transform result; Multiplication to obtain a first multiplication result,
Subjecting the second Fourier transform result and the second field Fourier transform result to one complex conjugate and the other complex multiplication to obtain a second multiplication result, and the inverse of the first and second multiplication results Performing a two-dimensional fast Fourier transform to obtain first and second inverse Fourier transform results; separating the first and second inverse Fourier transform results into a real part and an imaginary part; Obtaining first imaginary part data, second real part data, and second imaginary part data; and calculating a sum of squares of a plurality of candidate block data in the reference search range data. Step, the first real part data, the first imaginary part data, the second real part data, the second imaginary part data, and the plurality of candidate block data not divided into fields. Or field split By the square sum of the number of candidate block data, characterized by comprising the step of performing a similarity determination.

According to this method, the similarity is calculated using the sum of squares of a plurality of candidate block data, and the similarity is determined using the calculation result. Fast Fourier transform of the multiplication data and coefficient data and multiplication of these results can be omitted. Further, the input extended data and the reference search range data are divided to form input complex number data and reference complex number data, and a fast Fourier transform is performed.The result of one of the fast Fourier transforms is separated to obtain a result of the other fast Fourier transform. By performing the inverse fast Fourier transform on the result multiplied by the result, the number of data of the fast Fourier transform and the inverse fast Fourier transform can be reduced. Thus, a pattern matching method with a small amount of calculation can be provided.

In order to achieve the above object, a seventh motion vector detecting method according to the present invention is characterized in that an input block of an input image and a result obtained by vertically dividing the input block in the reference search range data of the reference image are respectively used. A motion vector detection method for detecting a relative position of a similar block, wherein the reference search range data is divided into fields, one of which is a real part, and the other is an imaginary part, to constitute reference complex number data; and Expanding the reference search range data to a field-divided size using the vertically divided result and 0 data, and configuring input complex number data with one as a real part and the other as an imaginary part; Performing a two-dimensional fast Fourier transform with the reference complex number data to obtain an input Fourier transform result and a reference Fourier transform result; And one of the input Fourier transform result and the reference Fourier transform result as a first Fourier transform result, and the other as a second Fourier transform result, and a first field Fourier transform corresponding to each field of the first Fourier transform result. Separating a result and a second field Fourier transform result, and performing one complex conjugate and the other complex multiplication on the second Fourier transform result and the first field Fourier transform result to obtain a first multiplied result Simultaneously performing one complex conjugate and the other complex multiplication on the second Fourier transform result and the second field Fourier transform result to obtain a second multiplication result, and inverting the first and second multiplication results. Performing a two-dimensional fast Fourier transform to obtain first and second inverse Fourier transform results; Separating the result of the Rier transform into a real part and an imaginary part, and obtaining first real part data, first imaginary part data, second real part data, and second imaginary part data; Calculating a sum of squares of a plurality of non-field-divided candidate block data or a plurality of field-divided candidate block data in the reference search range data; the first real part data; The imaginary part data, the second real part data, the second imaginary part data, and 2 of the plurality of candidate block data
Performing a similarity determination based on the sum of squares.

According to this method, the similarity is calculated by using the sum of squares of the plurality of candidate block data, and the similarity is determined by using the calculation result. Fast Fourier transform of the multiplication data and coefficient data and multiplication of these results can be omitted. Further, the input extended data and the reference search range data are divided to form input complex number data and reference complex number data, and a fast Fourier transform is performed.The result of one of the fast Fourier transforms is separated to obtain a result of the other fast Fourier transform. By performing the inverse fast Fourier transform on the result multiplied by the result, the number of data of the fast Fourier transform and the inverse fast Fourier transform can be reduced. Thus, a pattern matching method with a small amount of calculation can be provided.

The seventh motion vector detecting method is characterized in that in the two-dimensional fast Fourier transform of the reference complex data,
It is preferable to use the result of the one-dimensional fast Fourier transform in a certain direction of the reference complex data corresponding to a certain input block in the two-dimensional fast Fourier transform of the reference complex data corresponding to another input block.

According to this method, the number of one-dimensional fast Fourier transforms can be limited to one for portions where reference search range data corresponding to motion vector detection of a plurality of input blocks overlap. This makes it possible to further reduce the amount of calculation.

In order to achieve the above object, an eighth motion vector detecting method according to the present invention provides a method for detecting an input block of an input image and a result obtained by dividing the input block into upper and lower parts in reference search range data of the reference image. A motion vector detection method for detecting a relative position of a similar block, wherein the reference search range data is divided into fields, one of which is a real part, and the other is an imaginary part, to constitute reference complex number data; and Expanding the reference search range data to a field-divided size using the vertically divided result and 0 data, and configuring input complex number data with one as a real part and the other as an imaginary part; The reference complex number data is scanned to perform one-dimensional fast Fourier transform on the result as one-dimensional data, and the result is referred to as the input Fourier transform result. Obtaining a Fourier transform result, wherein one of the input Fourier transform result and the reference Fourier transform result is a first Fourier transform result, and the other is a second Fourier transform result, which corresponds to each field of the first Fourier transform result. The first field Fourier transform result and the second
Separating into a field Fourier transform result;
One complex conjugate and another complex multiplication are performed on the second Fourier transform result and the first field Fourier transform result to obtain a first multiplication result, and the second Fourier transform result and the second field Fourier transform are obtained. Performing one complex conjugate and the other complex multiplication on the result to obtain a second multiplication result; performing an inverse one-dimensional fast Fourier transform on the first and second multiplication results;
Obtaining an inverse Fourier transform result of: and separating the first and second inverse Fourier transform results into a real part and an imaginary part;
Obtaining a first real part data, a first imaginary part data, a second real part data, and a second imaginary part data; Calculating the sum of squares of the candidate block data and the plurality of field-divided candidate block data;
The first real part data, the first imaginary part data, the second real part data, the second imaginary part data, and the sum of squares of the plurality of candidate block data are similar. Performing a degree determination.

According to this method, the similarity is calculated by using the sum of squares of the plurality of candidate block data, and the similarity is determined by using the calculation result. Fast Fourier transform of the multiplication data and coefficient data and multiplication of these results can be omitted. Further, the input extended data and the reference search range data are divided to form input complex number data and reference complex number data, and a fast Fourier transform is performed.The result of one of the fast Fourier transforms is separated to obtain a result of the other fast Fourier transform. By performing the inverse fast Fourier transform on the result multiplied by the result, the number of data of the fast Fourier transform and the inverse fast Fourier transform can be reduced. Thus, a pattern matching method with a small amount of calculation can be provided.

[0053]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS (Embodiment 1) Hereinafter, a motion vector detecting method according to an embodiment of the present invention will be described.
This will be described with reference to the drawings.

Here, in order to simplify the explanation, the input block f (k, l) (k = 0 to 15, l =
0 to 15), the reference search range data g of the reference image
{Candidate block g in (k, l) (k = 0 to 31, l = 0 to 31)} (k + m, l + n) (k = 0 to 15, l
= 0 to 15, m = 0 to 16, n = 0 to 16), and the difference of 2
Consider a case in which m and n corresponding to data having the smallest sum of squares are searched.

FIG. 3 shows the relationship between the reference search range data and the candidate blocks. In FIG. 3A, reference numeral 31 denotes an input block f (k, l) (k = 0 to 15, l = 0 to 15).
And 33 is the reference search range data g ＾ (k, l)
(K = 0 to 31, 1 = 0 to 31). 34
Is a candidate block g ＾ (k + m, l +
n).

G ＾ (k + m, l + n) is k = 0 to 1
5, l = 0 to 15, m = 0 to 16, n = 0 to 16, but for convenience, g ＾ (k + m, l + n)
To extend m = 0-31 and n = 0-31, k +
When m and l + n are other than 0 to 31,

[0057]

## EQU4 ## k + m mod 32

[0058]

## EQU5 ## It is assumed that l + n mod 32 is obtained. Note that mod represents a remainder.

Input block 31 and reference candidate block 34
The sum of squares Err (m, n) of the difference

[0060]

(Equation 6)

Is obtained. Item 1

[0062]

(Equation 7)

Since is a fixed value for (m, n), the search for m and n that minimizes the sum of squares of the difference is performed by using the evaluation function e (m, n) of the following equation (Equation 8). What is necessary is just to search for m and n to be the minimum.

[0064]

(Equation 8)

Further, the input block 31 is expanded to the size of the reference search range data 33 as follows, and FIG.
The input extended data 32 is configured as shown in FIG. The input extended data 32 is f ＾ (k, l) (k = 0 to 31 and l
= 0 to 31), and are represented as follows.

[0066]

F ＾ (k, l) = f (k, l) (k = 0 to 15 and 1 = 0 to 15) = 0 (k = 16 to 31 or 1 = 16 to 31) Equation 10 is derived.

[0067]

(Equation 10)

This is because f ＾ (k, l) and g ＾ (k, l)
And the two-dimensional DF
T (Discrete Fourier Transf
orm), F (K, L) and G (K, L) can be used to represent the above equation (10). That is, a two-dimensional DFT is represented by DFT (•),

[0069]

F (K, L) = DFT (f ＾ (k, l))

[0070]

G (K, L) = DFT (g ＾ (k, l)), and the two-dimensional DFT result of the variable (m, n) is set to the variable (K,
L), the following equation is obtained.

[0071]

(Equation 13)

Here, ^* represents a complex conjugate. Therefore, the following Expression 14 is obtained.

[0073]

[Equation 14]

Here, IDFT (•) is an inverse two-dimensional DFT
And both sides are functions of (m, n).

From Expression 8, Expression 10, Expression 11, Expression 12, and Expression 14, Expression 15 is obtained.

[0076]

(Equation 15)

According to Equation 15, the input extension data f ＾ which has the size of the reference search range data g ＾ is filled with zeros in the portion having a difference in the size of the processing target block f.
And the input extended data f デ ー タ and the reference search range data g
２ is performed, and the complex conjugate of the result of the two-dimensional DFT of the input extended data f ＾ is multiplied by the result of the two-dimensional DFT of the reference search range data g ＾, and the inverse two-dimensional DFT of the multiplication result is obtained. Then, a sum of squares of the candidate block data of the reference search range data is calculated, and a result obtained by subtracting twice the result of the inverse two-dimensional DFT from the sum of squares of the candidate block data is obtained.
By searching for m and n corresponding to the smallest value,
M and n corresponding to the data having the smallest sum of squares of the difference from the processing target block f in the reference search range data g # can be searched.

Here, an embodiment according to Equation 15 will be described below.

FIG. 1 is a processing flowchart of the motion vector detecting method according to the present embodiment. Again, for simplicity, the input block f (k, l) (k = 0 to 15, l = 0) of the input image
To 15), the reference search range data g ＾ of the reference image
Candidate block g （(k + m, l +) in (m, n) in (k, l) (k = 0-31, l = 0-31)
n) (k = 0 to 15, l = 0 to 15, m = 0 to 16, n
= 0 to 16), and m and n corresponding to data having the smallest sum of squares of the difference are searched.

In FIG. 1, the target of the motion vector detection is the input block f (k, l) (k =
0 to 15, l = 0 to 15), and the reference search range data of the reference image is g ＾ (k, l) (k = 0 to 31, l = 0 to 0).
31).

In FIG. 1, step S1 is an input block expansion process, step S2 is a first two-dimensional FFT process,
Step S3 is a second two-dimensional FFT process, step S4
Is a multiplication process, step S5 is an inverse two-dimensional FFT process, step S6 is a square sum process, step S7 is a calculation process, step S8 is a minimum value search process, and step S9 is a motion vector calculation process.

In the input block expansion process in step S1, the following equation 1 is applied to the input block f (k, l).
As shown in FIG. 6, a process of expanding the data to the same size as the reference search range data g ＾ (k, l) is performed, and the input expanded data f ＾
(K, l).

[0083]

F ＾ (k, l) = f (k, l) (k = 0-15 and l = 0-15) = 0 (k = 16-31 or l = 16-31) In the two-dimensional FFT processing of No. 1, a two-dimensional FFT of the input extended data f デ ー タ (k, l) obtained in the input block expanding processing of step S1 is performed, and F (K,
L).

In the second two-dimensional FFT processing in step S3, the two-dimensional FFT of the reference search range data g ＾ (k, l) is performed.
Perform FT to obtain G (K, L).

In the multiplication process of step S4, the F obtained by the first two-dimensional FFT process of step S2 is obtained.
A complex multiplication of the complex conjugate of (K, L) and G (K, L) obtained by the second two-dimensional FFT processing in step S3 is performed.

[0086]

## EQU17 ## F (K, L) ^* G (K, L) is obtained.

In the inverse two-dimensional FFT processing in step S5, an inverse two-dimensional FFT of the multiplication result (Equation 17) obtained by the multiplication processing in step S4 is performed.
As a result of T,

[0088]

## EQU18 ## IFFT (F (K, L) ^*. G (K, L)) is obtained. The variables resulting from the inverse two-dimensional FFT are (m,
n).

In the square sum processing of step S6, m =
Reference search range data g for 0 to 16 and n = 0 to 16
By calculating the sum of squares of the candidate blocks corresponding to the position of (m, n) in ＾ (k, l),

[0090]

[Equation 19]

Is obtained.

The square value obtained here is different from (m, n)
Is used also in the calculation of the sum of squares of the candidate blocks corresponding to the position of the input block, and also in the calculation of the sum of squares of different input blocks.

In the arithmetic processing of step S7, the sum of squares obtained by the sum of squares processing of step S6, that is,

[0094]

(Equation 20)

Then, twice the result of the two-dimensional inverse FFT (Equation 18) obtained by the inverse two-dimensional FFT processing in step S5 is subtracted, and the following equation is obtained for m = 0-16 and n = 0-16. An evaluation function e (m, n) represented by 21 is obtained.

[0096]

(Equation 21)

In the minimum search processing in step S8, the evaluation function e (m,
In (n), m = 0-16 and n = 0-16 are searched for the minimum (m, n), and a motion vector is calculated in the motion vector calculation process in step S9 based on the search result.

As described above, a motion vector according to Equation 15 is obtained.

The inverse two-dimensional FFT processing in step S5, the square sum processing in step S6, the arithmetic processing in step S7, and the (m, n) in the minimum search processing in step S8
Is independent for each processing, and may be a method that is repeated for each set of (m, n), a method that is performed for all (m, n) for each processing, or a method that combines these. Further, the complex multiplication may be performed on the result of the complex conjugate of only the other.

FIG. 4 shows an example of a system configuration for executing the motion vector detection according to the present embodiment. In FIG. 4, reference numeral 81 denotes a CPU, 82 denotes a first storage unit, and 83 denotes a second storage unit.
Is an image input means, and 85 is an internal bus. The image input unit 84 captures a pair of images corresponding to the motion vector detection, and writes the pair of images into the input image area and the reference image area of the second storage unit 83.

Next, the CPU 81 executes the above-described input block expansion processing, two-dimensional FFT processing, multiplication processing, inverse two-dimensional FFT processing, square sum processing, arithmetic processing in accordance with the program stored in the second storage means 83 stored in advance. And the like to obtain a search result.

The first storage means 82 is constituted by, for example, a RAM, and is used for temporarily storing a part of a program necessary for the CPU processing, an input block, reference search range data, various intermediate data, variables, and the like. Used. CP
U may be a DSP, and the configuration is not limited to these.

For one input block (macro block), the number of multiplications in each process is simply calculated as N × N
For reference search range data, two-dimensional FFT of the reference search area data: 4 mN ² 2 dimensional FFT of the input extension data: 4 mN ² complex multiplication: 4N ² inverse 2D FFT: 4 mN ² 2 square of the candidate block (refer to the search range the square of the data): a N ^2.

Here, twice the result of the inverse two-dimensional FFT can be realized by the shift operation and is ignored, and the square of the reference search range data is the average number of times of multiplication by sharing with a plurality of input blocks. The actual number of multiplications is

[0105]

## EQU22 ## 4MN ² 3 + 4N ² + N ² = 12MN ² + 5N ² times. By substituting N = 32 and M = 5 as an example, 6
This is 6560 actual multiplications. Since the actual number of multiplications by the conventional method is 91136, the actual number of multiplications in the present embodiment is 73% of the conventional number.

When performing pattern matching of P-dimensional target data using P-dimensional template data, the input block f (k, l) in FIG. 1 is converted into P-dimensional (P is a natural number) template data and reference search range data g ＾.
The pattern matching can be performed by processing each step as P-dimensional data, such as (k, l) as P-dimensional target data and 2-dimensional FFT as P-dimensional FFT.

(Embodiment 2) Hereinafter, a motion vector detecting method according to another embodiment of the present invention will be described with reference to the drawings.

As in the first embodiment, for the input block f (k, l) (k = 0 to 15, 1 = 0 to 15) of the input image, the reference search range data g ＾ (k , L)
Candidate block g in (k = 0-31, l = 0-31)
＾ (k + m, l + n) (k = 0 to 15, l = 0 to 15,
A case is considered in which m and n corresponding to data in which m = 0 to 16 and n = 0 to 16) have the smallest sum of squares of the difference.

G ＾ (k + m, l + n) is k = 0 to 1
5, l = 0 to 15, m = 0 to 16, n = 0 to 16, but for convenience, g ＾ (k + m, l + n)
To extend m = 0-31 and n = 0-31, k +
When m and l + n are other than 0 to 31,

[0110]

(K + m) mod 32,

[0111]

It is assumed that (1 + n) mod 32 is obtained. mod represents a remainder.

The difference between the input block and the reference candidate block is 2
The search for m and n that minimizes the sum of the products Err (m, n) may be performed by searching for m and n that minimize the evaluation function e (m, n) in Expression 8.

Further, the input block is extended to the size of the reference search range as follows, and the input extended data f ＾ (k,
1) (k = 0 to 31 and l = 0 to 31).

[0114]

F２５ (k, l) = f (k, l) (k = 0 to 15 and l = 0 to 15) = 0 (k = 16 to 31 or 1 = 16 to 31) It looks like 10.

Further, g ＾ (k, l) (k = 0 to 31, l
= 0-31) and f ＾ (k, l) (k = 0-31, l = 0
To 31) to convert the one-dimensional data to gs (k) (k
= 0 to 1023) and fs (k) (k = 0 to 1023) are configured as in the following Expressions 26 and 27.

[0116]

Gs (k + 32 · l) = g ＾ (k, l) (k
= 0-31, l = 0-31)

[0117]

Fs (k + 32 · l) = f ＾ (k, l) (k
= 0-31, l = 0-31) Here, from Expression 26, Expression 10 is

[0118]

[Equation 28]

Is obtained. Note that i = m + 32 · n.
For convenience, the range of i of gs (k + i) is defined as i = 0 to 1023.
When k + i becomes 1024 or more in order to expand to

[0120]

It is assumed that (k + i) mod 1024 is taken. mod represents a remainder.

Since the above equation shows the cross-correlation between fs (k) and gs (k), the result F of each one-dimensional DFT
The above equation can be expressed using (K) and G (K).
That is, a one-dimensional DFT is represented by DFT (•),

[0122]

F (K) = DFT (fs (k))

[0123]

When G (K) = DFT (gs (k)), the result of the one-dimensional DFT by the variable i is represented by the variable K, and the following is obtained.

[0124]

(Equation 32)

Here, ^* represents a complex conjugate. Therefore, the following Expression 33 is obtained.

[0126]

[Equation 33]

Here, IDFT (•) represents a one-dimensional inverse DFT, and both sides are functions of i.

Equation 8, Equation 10, Equation 26, Equation 30, Equation 31,
From Expression 33 and Expression 33, Expression 34 below is obtained.

[0129]

(Equation 34)

According to Expression 34, the input extended data f ＾ in which the size of the reference search range data is made by filling the portion to be processed f with a portion having a difference in size with 0,
The input extension data is scanned to form the input scan data fs, the reference search range data g # is scanned to form the reference scan data gs, and the input scan data fs and the reference scan data g
s, and multiplies the complex conjugate of the one-dimensional DFT result of the input scan data fs by the one-dimensional DFT result of the reference scan data gs.
FT is performed, the sum of squares of the candidate block data of the reference search range data is calculated, and 1 is calculated from the sum of squares of the candidate block data.
A result obtained by subtracting twice the result of the dimensional inverse DFT is obtained, and i corresponding to the minimum value, i.e., m and n are searched. It can be seen that m and n corresponding to the data with the smallest sum of squares can be searched.

Here, the motion vector detecting method according to the present embodiment according to Expression 34 will be described below.

FIG. 5 is a processing flowchart showing an example of the procedure of the motion vector detecting method according to the present embodiment. Here again, for simplicity, the input block f (k, l) (k =
For 0 to 15, l = 0 to 15), reference search range data g ＾ (k, l) (k = 0 to 31, l = 0 to 0) of the reference image
31), the candidate block g） (k) at (m, n)
+ M, l + n) (k = 0 to 15, l = 0 to 15, m = 0
１６16, n = 0ｎ16) and m and n corresponding to the data having the smallest sum of squares of the difference are searched.

In FIG. 5, the input block of the input image to be subjected to the motion vector detection is f (k, l) (k = 0 to 0).
15, l = 0 to 15), and the reference search range data of the reference image is g ＾ (k, l) (k = 0 to 31, l = 0 to 3).
1).

In FIG. 5, step S91 is an input block extended scanning process, and step S92 is a first one-dimensional FF.
T processing, step S93 is scanning processing of reference search range data, step S94 is second one-dimensional FFT processing, step S95 is multiplication processing, and step S96 is inverse one-dimensional FFT.
Processing, step S97 is a sum of squares processing, step S98 is an arithmetic processing, step S99 is a minimum value search processing, and step S100 is a motion vector calculation processing.

In the input block extended scanning process at step S91, the reference search range data g ＾ (k, l) is obtained for the input block f (k, l) as shown in the following Expression 35.
Is expanded and scanned to the same size as that described above to form input scan data fs (k).

[0136]

Fs (k + 32 · l) = f (k, l) (k = 0 to 15 and 1 = 0 to 15) = 0 (k = 16 to 31 or 1 = 16 to 31) First of step S92 In the one-dimensional FFT processing of, a one-dimensional FFT of the input extended data fs (k) obtained in the input block extended scanning processing in step S91 is performed, and
(K) is obtained.

In the reference search range data scanning process in step S93, the reference search range data g ＾ (k, l) is scanned, and reference scan data gs is formed as follows.

[0138]

Gs (k + 32 · l) = f (k, l) (k
= 0-31 or 1 = 0-31) In the second one-dimensional FFT processing in step S94, one-dimensional FFT of the reference scan data gs (k) is performed, and G
(K) is obtained.

In the multiplication processing in step S95, the F obtained by the first one-dimensional FFT processing in step S92 is obtained.
(K) and the second one-dimensional F in step S94
Performs complex multiplication with G (K) obtained by the FT processing,
As a result of the multiplication,

[0140]

(37) F (K) ^* · G (K) is obtained.

In the inverse one-dimensional FFT processing in step S96, an inverse one-dimensional FFT of the multiplication result (Equation 37) obtained by the multiplication processing in step S95 is performed.
As a result of

[0142]

## EQU38 ## IFFT (F (K) ^*. G (K)) is obtained. The variable resulting from the inverse one-dimensional FFT is i.

In the square sum process of step S97, m
= 0 to 16 and n = 0 to 16, the sum of squares of the candidate blocks corresponding to the position of (m, n) in the reference search range data g ＾ (k, l) is calculated, and is expressed by the following Expression 39. Is obtained.

[0144]

[Equation 39]

The square value obtained here is another (m, n)
Is used also in the calculation of the sum of squares of the candidate blocks corresponding to the position of the input block, and also in the calculation of the sum of squares of different input blocks.

In the arithmetic processing of step S98, the sum of squares obtained by the sum of squares processing of step S97, that is,

[0147]

(Equation 40)

From the result of the inverse one-dimensional FFT obtained by the inverse one-dimensional FFT processing in step S96 (Equation 38), 2
By subtracting the times, for m = 0-16 and n = 0-16,
An evaluation function e (m, n) represented by the following Expression 41 is obtained.
Here, i = m + 32 · n.

[0149]

[Equation 41]

In the minimum search processing in step S99,
Evaluation function e obtained by the arithmetic processing in step S98
A minimum (m, n) is searched for m = 0 to 16 and n = 0 to 16 of (m, n), and based on the search result,
In the motion vector calculation process in step S100, a motion vector is calculated.

As described above, a motion vector is obtained according to the equation (34).

Here, the inverse one-dimensional FFT of step S96
The variable i = m + 32 · n of the processing, (m, n) in the square sum processing in step S97, the arithmetic processing in step S98, and the minimum search processing in step S99 are independent for each processing, and one set of ( The method may be repeated for each (m, n), may be performed for all (m, n) for each process, or a combination of these methods. Also, the complex multiplication may be performed on the result of taking the complex conjugate of only the other.

For one input block (macro block), the number of multiplications in each process is simply calculated as N × N
, One-dimensional FFT of reference search range data: 4MN ² one-dimensional FFT of input extended data: 4MN ² complex multiplication: 4N ² inverse one-dimensional FFT: square of candidate block 4MN ² (reference search range the square of the data): a N ^2.

Here, twice the result of the inverse one-dimensional FFT can be realized by the shift operation and is ignored, and the square of the reference search range data is the average number of times of multiplication by sharing with a plurality of input blocks. The actual number of multiplications is

[0155]

[Formula 42] 4MN ² · 3 + 4N ² + N ² = 12MN ² + 5N ² times. By substituting N = 32 and M = 5 as an example, 6
This is 6560 actual multiplications. Since the actual number of multiplications by the conventional method is 91136, the actual number of multiplications in the present embodiment is 73% of the conventional number.

When performing pattern matching of P-dimensional target data using P-dimensional template data, the input block f (k, l) shown in FIG. 5 is converted to P-dimensional (P is a natural number) template data and reference search range data g ＾.
(K, l) is replaced with P-dimensional target data, and step S9 is performed.
Instead of the input block extension scanning process of 1 to convert P-dimensional data to Q (Q is a natural number smaller than P) dimension data by scanning and to perform data extension by using 0, the reference search range data of step S93 The scanning process is replaced with a process in which P-dimensional data is converted to Q (Q is a natural number smaller than P) -dimensional data by scanning, and the two-dimensional data in each step is converted to P
Process as one-dimensional data and convert one-dimensional FFT to Q-dimensional FFT
For example, pattern matching can be performed by processing the one-dimensional data in each step as Q-dimensional data.

(Embodiment 3) Hereinafter, a motion vector detecting method according to still another embodiment of the present invention will be described with reference to the drawings.

As shown in the first embodiment, the two-dimensional FF
A two-dimensional FFT method that further reduces the amount of computation when performing motion vector detection using T will be described as the present embodiment.

As shown in FIG. 6A, in the input image 41, two input blocks are divided into the first input block 4
When the two reference search range data are adjacent in the y direction (vertical direction in the figure) as in the second and second input blocks 43, the two reference search range data are divided into the first reference search range 46 shown in FIG. 2
Is used, as in the reference search range 48 of FIG.

FIG. 7 illustrates the configuration of the calculation of the reference search range data. In FIG. 7, reference numeral 51 denotes first reference search range data g ＾ ₁ (k, l), and 52 denotes second reference search range data g ＾ ₂ (k, l), where k = 0 to 31,
l = 0 to 31, and the input block has a size of 16 × 16.

A two-dimensional FFT is converted to a one-dimensional F in the x and y directions.
When processing by FT, the first reference search range 51 (g ＾ ₁
(K, l)) is accessed in the x direction, and the one-dimensional FFT in the x direction is repeated 32 times to obtain the first intermediate data 53 (gi).
₁ (K, l)), and the first intermediate data 53
The first two-dimensional FFT result 55 (G ₁ (K, L)) is obtained by accessing the direction and repeating the one-dimensional FFT in the y direction 32 times.
Can be obtained.

The second reference search range data 52 (g
＾ ₂ (k, l)) is accessed in the x direction, and the one-dimensional FFT in the x direction is repeated 32 times to obtain the second intermediate data 54 (g
i ₂ (K, l)), the second intermediate data 54 is accessed in the y direction, and the one-dimensional FFT in the y direction is repeated 32 times, and the second two-dimensional FFT result 56 (G ₂ (K, l,
L)) can be obtained.

In the following, based on FIG.
The motion vector detection method according to the present embodiment will be described using a two-dimensional FFT of the reference search range data as an example. 8, step S61 is 1-dimensional FFT processing of x-direction, the step S62 is storage process to the storage means, the step S63 is a one-dimensional FFT processing in the y direction, these processes, the reference search range data g ^ ₁ (K, l) two-dimensional FFT result G ₁ (K, L) and reference search range data g
The two-dimensional FFT result G ₂ (K, L) of ＾ ₂ (k, l) is obtained.

In step S61, one-dimensional F in the x direction
By performing the FT processing, the reference search range data g ＾ ₁
One-dimensional FFT of (k, l) in the x direction is performed 32 times. Thus, the first intermediate data gi ₁ (K, l) is obtained and stored in the storage unit in step S62.

The first intermediate data gi
₁ (K, l) is read, and in step S63, one-dimensional FFT processing in the y direction is performed 32 times to obtain a two-dimensional FFT result G ₁ (K, L). The lower half data (K = 0 to 15, l = 16 to 31) of the first intermediate data gi ₁ (K, l) is held by the storage processing of step S62.

Next, the one-dimensional F in the x direction in step S61
By the FT processing, the reference search range data g ＾ ₂ (k, l)
One-dimensional FFT in the x direction is performed 16 times on the lower half data (k = 0 to 15, l = 16 to 31), and the lower half data (K = 1) of the second intermediate data gi ₂ (K, l) 0-1
5, l = 16 to 31) and stores it in the storage means in step S62.

Next, the first intermediate data gi
₁ Lower half data of (K, l) (K = 0 to 15, l = 1
6 to 31) and the lower half of the second intermediate data gi ₂ (K, l) (K = 0 to 15, l = 16 to 31), and K = 0 to 15, l = 1 to 15 At

[0168]

## EQU43 ## Assuming that gi ₂ (K, 1) = gi ₁ (K, 1 + 16), the second intermediate data gi ₂ (K, 1) (K = 0 to ₂ )
31, l = 0 to 31).

The one-dimensional F in the y direction in step S63
By the FT process, gi ₂ (K, l) (K = 0 to 31, l
= 0 to 31), the one-dimensional FFT in the y direction is performed 32 times, and a two-dimensional FFT result G ₂ (K, L) 67 is obtained.

First reference search range data g ＾ ₁ (k,
l) Lower half data (k = 0-31, l = 16-3)
1) and the upper half data (k = 0 to 31, l = 0 to 15) of the second reference search range data g ＾ ₂ (k, l) are common, so that their one-dimensional FFT in the x direction is used. Of the lower half of the first intermediate data gi ₁ (K, l) (K = 0 to 31, l = 16 to 31) and the second intermediate data gi ₂ (K, l) Half of the data (K = 0 to 31,
l = 0 to 15) are also common, so that
A two-dimensional FFT result G ₂ (K, L) can be obtained.

A method of calculating reference search range data according to the present embodiment will be described with reference to FIG. 9, reference numeral 71 denotes first reference search range data g 範 囲₁ (k,
1) and 72 are the second reference search range data g ＾ ₂ (k,
l), where k = 0 to 31, l = 0 to 31, and the input block has a size of 16 × 16.

The first reference search range data 71 (g ＾
₁ (k, l)) is accessed in the x direction, and the one-dimensional FFT in the x direction is repeated 32 times to obtain the first intermediate data 53 (gi).
₁ (K, l)), and the first intermediate data 53
By accessing the direction and repeating the one-dimensional FFT in the y-direction 32 times, the first two-dimensional FFT result 75 (G ₁ (K,
L)).

The first intermediate data 53 (gi
₁(K, l)) lower half data (K = 0 to 31, l =
16 to 31), and the second reference search range data 72
(G ＾_Two(K, l)) lower half data (k = 0 to 3)
1, l = 16 to 31) in the x direction and
The one-dimensional FFT is repeated 16 times, and the second intermediate data 74
(Gi _Two(K, l)) lower half data (K = 0 to 3)
1, 1 = 16 to 31) and the first intermediate data 53 (g
i₁(K, l)) lower half data (K = 0 to 31, l
= 16 to 31) to the second intermediate data 74 (gi)_Two(K,
l)) Upper half data (K = 0 to 31, l = 0 to 1)
5), and the second intermediate data 74 (gi_Two(K, l)
(K = 0-31, l = 0-31)) in the y direction
Then, the one-dimensional FFT in the y direction is repeated 32 times to obtain the second 2
Dimensional FFT result 76 (G_Two(K, L)).

In the present embodiment, two input blocks (macroblocks) are considered. However, the same processing can be performed on input blocks above the first input block. One-dimensional FFT
May be performed with 32 × 16 data. If the number of multiplications in each process is simply calculated for one input block (macroblock), in the case of N × N reference search range data, the FFT of the reference search range data in the x direction: 16 · (MN /
2) · 4 = 32 MN FFT of reference search range data in y direction: ² MN ² FFT of input extended data: 4 MN ² Complex multiplication: 4N ² Inverse FFT: square of 4MN ² candidate block (square of reference search range data) : the N ^2.

Here, since twice the result of the inverse two-dimensional FFT can be realized by shifting, it is ignored, and 2 times of the reference search range data is ignored.
The power is the average number of times of multiplication by sharing a plurality of input blocks. The actual number of multiplications is

[0176]

[Formula 44] 32MN + 2MN ² + 4MN ² · 2 + 4N ² + N ² = 32MN + 10MN ² + 5N ² times. By substituting N = 32 and M = 5 as an example, 6
This is 1440 actual multiplications. Since the actual number of multiplications by the conventional method is 91136, the actual number of multiplications in the present embodiment is 67% of the conventional number.

(Embodiment 4) Hereinafter, a motion vector detecting method according to still another embodiment of the present invention will be described with reference to the drawings.

As in the first embodiment, for the input block f (k, l) (k = 0 to 15, 1 = 0 to 15) of the input image, the reference search range data g ＾ (k , L)
Candidate block g in (k = 0-31, l = 0-31)
＾ (k + m, l + n) (k = 0 to 15, l = 0 to 15,
A case is considered in which m and n corresponding to data in which m = 0 to 16 and n = 0 to 16) have the smallest sum of squares of the difference.

G ＾ (k + m, l + n) is k = 0 to 1
5, l = 0 to 15, m = 0 to 16, n = 0 to 16, but for convenience, g ＾ (k + m, l + n) is extended to m = 0 to 31, n = 0 to 31 K +
When m and l + n are other than 0 to 31,

[0180]

(K + m) mod 32,

[0181]

It is assumed that (l + n) mod 32 is obtained. mod represents a remainder.

The difference between the input block and the reference candidate block is 2
The search for m and n that minimizes the sum of the products Err (m, n) may be performed by searching for m and n that minimize the evaluation function e (m, n) in Expression 8.

Further, the input block is extended to the size of the reference search range as shown in the following Expression 47, and the input extended data f ＾ (k, l) (k = 0-31 and l = 0-31) is obtained. Constitute.

[0184]

F４７ (k, l) = f (k, l) (k = 0 to 15 and l = 0 to 15) = 0 (k = 16 to 31 or 1 = 16 to 31) It looks like 10.

Further, g ＾ (k, l) (k = 0 to 31, l
= 0-31) and f ＾ (k, l) (k = 0-31, l =
0 to 31) is divided for each line, and the reference top data gt (k, l) (k = 0 to 31, l = 0 to 1)
5), reference bottom data gb (k, l) (k = 0 to 3)
1, l = 0 to 15), input top data ft (k, l)
(K = 0 to 31, l = 0 to 15), input bottom data f
b (k, l) (k = 0 to 31, l = 0 to 15) is configured as in the following Expressions 48 to 51.

[0186]

Ft (k, l) = f ＾ (k, 2 · l)

[0187]

Fb (k, l) = f ＾ (k, 2 · l + 1)

[0188]

Gt (k, l) = g ＾ (k, 2 · l)

[0189]

Gb (k, l) = g ＾ (k, 2 · l + 1) Here, for convenience, ft (k, l), fb (k, l), g
t (k, l), gb (k, l), k, l of (k, l)
Is outside the range of 0 to 31,

[0190]

[Equation 52] k mod 32

[0191]

It is assumed that l mod 16 is taken.

From Expressions 48 to 51, Expression 10 is

[0193]

(Equation 54)

Are expressed as follows.

When n is an even number, the above equation (54) is obtained.
Is set as n ＾ = n / 2, and since n０〜 = 0 to 8, it is expressed by the following Expression 55.

[0196]

[Equation 55]

On the other hand, when n is an odd number,
Is set as n ₁ = (n + 1) / 2, and n ₁ = 1 to 8,
Since n ₂ = (n−1) / 2 and n ₂ = 0 to 7, it is represented by the following equation 56.

[0198]

[Equation 56]

FIG. 10 shows that in Equation 55, m = 0 and n = 2.
FIG. 11 shows a data relationship when (even number) is set, and FIG. 11 shows a data relationship when m = 0 and n = 3 (odd number) in Equation 56.

The input extended data 121 shown in FIG.
Of (f ＾ (k, l)), f (k, l) (k = 0 to 1)
5, 1 = 0 to 15) corresponds to FIG.
An area 122a of the input top data 122 shown in FIG.
And the area 123a of the input bottom data 123 shown in FIG.

Reference search range data 1 shown in FIG.
24 (g ＾ (k, l)), the input extended data 121
(K, l) corresponds to the area 121a of (f ＾ (k, l)) because the reference search range data 124 (g ＾ (k, l)
1)) area 124a. The area 124a has k = 0.
１５15, l = 2１７17.

The reference search range data 124 (g ＾ (k,
The area 124a of l)) corresponds to the area 125a of the reference top data 125 shown in FIG. 10E and the area 126a of the reference bottom data 126 shown in FIG. 10F, respectively. Note that the region 125a and the region 126a
k = 0 to 15, and l = 1 to 8.

According to the correspondence between these areas,

[0204]

[Equation 57]

Similarly, when n is an even number, the expression is the same as Expression 55.

In FIG. 11A, the input extended data 1
F (k, l) (k = 0 to 31) in 31 (f ＾ (k, l))
15, l = 0 to 15) corresponds to FIG.
An area 132a of the input top data 132 shown in FIG.
And the area 133a of the input bottom data 133 shown in FIG.

Reference search range data 1 shown in FIG.
34 (g ＾ (k, l)), the input extended data 131
(K, l) corresponds to the area 131a of (f ＾ (k, l)) because the reference search range data 134 (g ＾ (k, l)
1)) area 134a. The area 134a has k = 0.
１５15, 1 = 3〜18.

Reference search range data 134 (g ＾ (k,
1)) corresponds to an area 135a of the reference bottom data 135 shown in FIG. 11E and an area 136a of the reference top data 136 shown in FIG. 11F, respectively. The region 135a has k = 0 to 15, l = 1 to
8 and the area 136a is k = 0 to 15, l = 2 to 9
It is.

According to the correspondence between these areas,

[0210]

[Equation 58]

In the same manner, when n is an odd number, the equation becomes the same as Equation 56.

Equation 54 represents ft (k, l) and gt (k,
l), fb (k, l) and gb (k, l), ft (k,
l) and gb (k, l), fb (k, l) and gt (k,
1) represents the cross-correlation, so that Ft (K, L), Fb
Equation (54) can be expressed using (K, L), Gt (K, L), and Gb (K, L). That is, a two-dimensional DFT is represented by DFT (•),

[0213]

Ft (K, L) = DFT (ft (k, l))

[0214]

Fb (K, L) = DFT (fb (k, l))

[0215]

Gt (K, L) = DFT (gt (k, l))

[0216]

When the result of the two-dimensional DFT using the variable (m, n) is expressed by the variable (K, L) by Gb (K, L) = DFT (gb (k, l)), Equation 66 is obtained.

[0219]

[Equation 63]

[0218]

[Equation 64]

[0219]

[Equation 65]

[0220]

[Equation 66]

Here, ^* represents a complex conjugate. Therefore, the following Expressions 67 to 70 are obtained.

[0222]

[Equation 67]

[0223]

[Equation 68]

[0224]

[Equation 69]

[0225]

[Equation 70]

Here, IDFT (•) is the inverse two-dimensional DFT
And both sides are functions of (m, n).

Equation 8, Equation 10, Equation 54, Equation 59 to Equation 62,
And according to Equations 67 to 70, when n is an even number, the following Equation 71 is obtained. If n is an odd number, Equation 72 is obtained.

[0228]

[Equation 71]

[0229]

[Equation 72]

According to the equations (71) and (72), the input extension data f which is the size of the reference search range data is obtained by filling the processing target block f with a portion having a difference in size with zeros.
＾ (k, l) and the input extended data f ＾ (k, l)
And the reference search range data g ＾ (k, l) are divided into fields, respectively, and ft (k, l), fb (k, l), gt
(K, l) and gb (k, l), and ft (k,
l), fb (k, l), gt (k, l), gb (k,
l) Two-dimensional DFT results Ft (K, L), Fb (K,
L), Gt (K, L), and Gb (K, L), as in the first embodiment, the reference search range data g ＾
It can be seen that m and n corresponding to data having the smallest sum of squares of the difference with the processing target block f (k, l) can be searched for in (k, l).

Next, a specific calculation method will be described. ft
(K, l), fb (k, l), gt (k, l), gb
(K, l) is a real number, and the following equations 73 and 74 form complex number data fc (k, l) and gc (k, l). Note that k = 0 to 15, l = 0 to 15, and j are imaginary units.

[0232]

Fc (k, l) = ft (k, l) + j · fb (k, l)

[0233]

Gc (k, l) = gt (k, l) + j · gb (k, l) FIG. 12 shows the structure of the data of Equations 73 and 74. In FIG. 12A, the input extended data 141 (f ＾
(K, l)) is divided into fields, and the input top data 142 (ft (k, l)) and the input bottom data 143
(Fb (k, l)), and the input complex number data 144 (fc (k, l,
1)) is constituted.

Also, as shown in FIG. 12B, the reference search range data 145 (g ＾ (k, l)) is divided into fields, and the reference top data 146 (gt (k, l)) and the reference bottom data 147 (gb (k, l)) is obtained, and reference complex number data 148 (gc (k, l)) is constructed from these based on Equations 73 and 74.

The DFTs of fc and gc are as shown in the following equations (75) and (76), respectively.

[0236]

Fc (K, L) = DFT (fc (k, l)) = DFT (ft (k, l) + j · fb (k, l)) = Ft (K, L) + j · Fb (K , L)

[0237]

Gc (K, L) = DFT (gc (k, l)) = DFT (gt (k, l) + j · gb (k, l)) = Gt (K, L) + j · Gb (K , L) Here, since Gt and Gb are real DFT results, if the data arrangement is reversed, it is equal to the complex conjugate, ie,

[0238]

Gt (31−K, 15−L) ^* = Gt (K, L)

[0239]

(78) The property of Gb (31−K, 15−L) ^* = Gb (K, L) is obtained. When the complex conjugate is obtained by changing the variables of equations (75) and (76), the following equation (79) is obtained. Become.

[0240]

Gc (31−K, 15−L) ^* = Gt (31−K, 15−L) ^{* −} j · Gb (31−K, 15−L) ^* = Gt (K, L) −j Gb (K, L) From Equations 75, 76, and 79, as shown in the following Equations 80 and 81, 2Gt (K, L) and 2Gb
(K, L).

[0241]

[Expression 80] 2 · Gt (K, L) = Gc (K, L) + Gc
(31-K, 15-L) ^*

[0242]

81 · Gb (K, L) = 1 / j · (Gc (K,
L) -Gc (31-K, 15-L) ^* ) Here, 1 / j may be set to the imaginary part by changing the sign of the real part to the imaginary part. When inverse two-dimensional DFT of the separated result is performed,

[0243]

IDFT (Fc (K, L) ^* · 2 · Gt (K, L)) = IDFT (2 · (Ft (K, L) ^* + j · Fb (K, L) ^* ) · Gt (K , L)) = 2 · IDFT (Ft (K, L) ^* · Gt (K, L)) + j · 2 · (Fb (K, L) ^* · Gt (K, L))

[0244]

IDFT (Fc (K, L) ^* · 2 · Gb (K, L)) = IDFT (2 · (Ft (K, L) ^* + j · Fb (K, L) ^* ) · Gb (K , L)) = 2 · IDFT (Ft (K, L) ^* · Gb (K, L)) + j · 2 · (Fb (K, L) ^* · Gb (K, L)), and ft (k, l), fb (k, l), gt (k,
l) and gb (k, l) are real numbers, so the real part of Equation 82 is

[0245]

84 IDFT (Ft (K, L) ^* Gt (K, L)), where the imaginary part is

[0246]

[Equation 85] 2 · (Fb (K, L) ^* · Gt (K, L))

[0247]

86 IDFT (Ft (K, L) ^* Gb (K, L)), where the imaginary part is

[0248]

87 (Fb (K, L) ^* Gb (K, L))
Becomes

Therefore, terms other than the sum of squares of Equations 71 and 72 are obtained by converting fc (k, l) and gc (k, l) to DFT
It can be determined from the results.

Here, an embodiment according to equations 71 and 72, 82 and 83 will be described below.

FIG. 13 is a processing flowchart showing an example of the procedure of the motion vector detecting method according to the present embodiment. Again, for simplicity, the input block f (k, l) of the input image
(K = 0 to 15, l = 0 to 15), the reference search range data g ＾ (k, l) of the reference image (k = 0 to 31, l
= 0 to 31), the candidate block g in (m, n)
＾ (k + m, l + n) (k = 0 to 15, l = 0 to 15,
A case is considered in which m and n corresponding to data in which m = 0 to 16 and n = 0 to 16) have the smallest sum of squares of the difference.

In FIG. 13, the input block of an input image to be subjected to motion vector detection is f (k, l) (k = 0
１５15, l = 0１５15), and the reference search range data of the reference image is g ＾ (k, l) (k = 0０〜31, l = 0０〜3).
1).

In FIG. 13, step S101 is an input block extension process, step S102 is a first complex number construction process, step S103 is a first two-dimensional FFT process,
Step S104 is a second complex number construction process, step S104.
105 is a second two-dimensional FFT processing, step S106 is a frequency separation processing, step S107 is a multiplication processing, step S108 is an inverse two-dimensional FFT processing, step S109 is a real part imaginary part separation processing, step S110 is a square sum processing, Step S111 is a calculation process, step S112 is a minimum value search process, and step S113 is a motion vector calculation process.

In the input block expansion processing in step S101, expansion is performed on the input block f (k, l) so that the input search block has the same size as the reference search range data g ＾ (k, l) as shown in the following equation. And input extension data f デ ー タ (k,
1).

[0255]

F８８ (k, l) = f (k, l) (k = 0 to 15 and l = 0 to 15) = 0 (k = 16 to 31 or l = 16 to 31) 1, the input block f (k, l) is divided into fields as shown in the following Expressions 89 to 91, so that the input top data f
t (k, l) and input bottom data fb (k, l) are formed, and input complex number data fc (k, l) is formed. In each case, k = 0 to 31, and l = 0 to 15.

[0256]

Ft (k, l) = f ＾ (k, 2 · l)

[0257]

Fb (k, l) = f ＾ (k, 2 · l + 1)

[0258]

Fc (k, l) = ft (k, l) + j · fb (k, l)
In the first two-dimensional FFT processing in step S103,
A two-dimensional FFT is performed on the input complex data fc (k, l) obtained in the first complex number configuration processing in step S102, and an input Fourier transform result Fc (K, L) is obtained.

By the second complex number construction processing in step S104, the reference search range data g ＾ (k, l) is divided into fields as shown in the following Expressions 92 to 94, and gt (k,
l) and gb (k, l) to form reference complex data gc
(K, l).

[0260]

Gt (k, l) = g ＾ (k, 2 · l)

[0261]

Gb (k, l) = g ＾ (k, 2 · l + 1)

[0262]

Gc (k, l) = gt (k, l) + j · gb (k, l) FIG. 10 shows the configuration of input complex number data fc (k, l) and reference complex number data gc (k, l). It shows how to do. FIG.
The input block 121 (f (k, l)) shown in FIG. 10A is divided into fields, and the input top data 122 (ft (k, l)) shown in FIG. 10B and the input shown in FIG. Bottom data 123 (fb (k, l)), and from these, input complex number data fc (k, l).

The reference search range data 124 (g （(k, l)) shown in FIG. 10D is divided into fields, and the reference top data 125 (gt (k,
1)) and the reference bottom data 126 shown in FIG.
(Gb (k, l)) and the reference complex data gc
(K, l). Second two-dimensional FFT processing 100
At 7, the two-dimensional FFT of the reference complex data gc (k, l) obtained in the second complex number construction processing 1006 is performed,
The reference Fourier transform result Gc (K, L) is obtained.

In the frequency separation processing of step S106, 2 · Gt (K, L) and 2 · Gt (K, L)
Calculate Gb (K, L).

In the multiplication process of step S107, the complex conjugate of Fc (K, L) obtained by the first two-dimensional FFT process of step S103 and the 2 · Gt (K) obtained by the frequency separation process of step S106. , L), and as a result of the multiplication,

[0266]

Fc (K, L) ^* · 2 · Gt (K, L) is obtained, the complex conjugate of Fc (K, L) obtained by the first two-dimensional FFT processing in step S103, and step S
2 · Gb (K,
L) and a complex multiplication with

[0267]

The following equation is obtained: Fc (K, L) ^* · 2 · Gb (K, L)

In the inverse two-dimensional FFT processing in step S108, an inverse two-dimensional FFT of the multiplication result (equations 95 and 96) obtained by the multiplication processing in step S107 is performed. 97 and 98 are obtained. The variables resulting from the inverse two-dimensional FFT are (m,
n).

[0269]

97 IFFT (F (K, L) ^* · 2 · Gt (K, L))

[0270]

IFFT (F (K, L) ^* · 2 · Gb (K, L)) In the real part imaginary part separation processing in step S109, the inverse two-dimensional FFT obtained by the inverse two-dimensional FFT processing in step S108 Separating the real part and imaginary part of each result (Equation 97 and Equation 98)

[0271]

[Equation 99] Re (IFFT (F (K, L) ^* · 2 · Gt)
(K, L))),

[0272]

[Equation 100] Im (IFFT (F (K, L) ^* · 2 · G
t (K, L))),

[0273]

Re (IFFT (F (K, L) ^* . 2.G)
b (K, L))),

[0274]

Im (IFFT (F (K, L) ^* · 2 · G)
b (K, L))). Note that Re (•) represents a real part, and Im (•) represents an imaginary part.

In the square sum processing of step S110,
For m = 0 to 16 and n = 0 to 16, the sum of squares of the candidate blocks corresponding to the position of (m, n) in the reference search range data g ＾ (k, l) is calculated, and To obtain the sum of squares represented by

[0276]

[Equation 103]

The square value obtained here is another (m, n)
Is used also in the calculation of the sum of squares of the candidate blocks corresponding to the position of the input block, and also in the calculation of the sum of squares of different input blocks.

In the arithmetic processing of step S111, the sum of squares represented by the following equation 104 obtained by the square sum processing of step S110 and the result obtained by the real part and imaginary part separation processing of step S109 Is subtracted.

[0279]

[Equation 104]

As a result, m = 0 to 16, n = 0, 2,
, 16 (even numbers from 0 to 16), an evaluation function e (m, n) represented by the following Expression 105 is obtained.

[0281]

[Equation 105]

Also, m = 0 to 16, n = 1, 3,..., 1
For 5 (odd number from 0 to 16),
The evaluation function e (m, n) expressed by

[0283]

[Equation 106]

In the minimum search processing of step S112, the minimum (m, n) of the evaluation function e (m, n) obtained by the arithmetic processing of step S111 with respect to m = 0 to 16 and n = 0 to 16 is determined. Search, and by this search result,
In the motion vector calculation process of step S113, a motion vector is calculated.

As described above, a motion vector according to Equations 71 and 72 is obtained.

Here, the inverse two-dimensional FF of step S108
T processing, real part imaginary part separation processing in step S109, square sum processing in step S110, arithmetic processing in step S111, and (m,
n) is independent for each process, and may be a method that is repeated for each set of (m, n), a method that is performed for all (m, n) for each process, or a combination of these methods. .

In the input block extension processing and the first complex number construction processing, the extension processing may be performed after the input block is divided into fields, or the extension and the complex number construction may be performed simultaneously. Further, the frequency separation processing may be performed on the result of the input Fourier transform, the complex multiplication may be performed on the result obtained by taking the complex conjugate of only the other, and the field division defines a field for each vertical line of the image. May be divided.

When the number of multiplications in each process is simply calculated for one input block (macro block), in the case of N × N reference search range data, two-dimensional FFT of reference search range data: ((N / 2)・ (MN / 2) + N ・ (M-1) ・ (N /
2) / 2) · 4 = (2M−1) N ² -dimensional FFT of input extended data: (2M−1) N ² complex multiplication: 4N ² inverse 2D FFT: (2M−1) N ² · 2 candidate (the square of the reference search range data) square blocks: a N ^2.

Here, since the result of the inverse two-dimensional FFT can be realized by a shift operation, it is ignored, and the square of the reference search range data is the average number of times of multiplication by sharing a plurality of input blocks.

The number of actual multiplications is

[0291]

(2M−1) N ² · 4 + 4N ² + N ² = 8MN ² + N ² times. By substituting N = 32 and M = 5 as an example, 4
This is the number of actual multiplications of 1984. Since the actual number of multiplications by the conventional method is 91136, the actual number of multiplications in the present embodiment is 46% of the conventional number.

The method of the present embodiment and the method of the second embodiment can be used together, and the ft and fb obtained by the first complex number construction processing in step S102 and the input complex number data fc are scanned one-dimensionally. A process for converting the gt and gb and the reference complex data gc obtained by the second complex number configuration process in step S104 into one-dimensional data by scanning is performed. It can be realized by processing as one-dimensional data.

When the method of the third embodiment is used in combination with the method of the present embodiment, the number of times of multiplication is further reduced. When the number of multiplications in each process is simply calculated for one input block (macro block), in the case of N × N reference search range data, two-dimensional FFT of reference search range data: ((16/2) · (MN / 2) + N ・ (M-1) ・ (N
/ 2) / 2) · 4 = 16MN + (M-1) N 2 2 dimensional FFT of the input extension data: (2M-1) N ² complex multiplication: 4N ² inverse 2D FFT: (2M-1) N 2 · 2 The square of the candidate block (the square of the reference search range data): N ² . Here, twice the result of the inverse two-dimensional FFT is ignored because it can be realized by a shift operation. The actual number of multiplications is

[0294]

Equation 108] 16MN + (M-1) N 2 + (2M-1)
The ^{N 2 · 3 + 4N 2 +} N 2 = 16MN + 7MN 2 + N 2 times. By substituting N = 32 and M = 5 as an example, 3
This is 9424 actual multiplications. Since the actual number of multiplications by the conventional method is 91136, the actual number of multiplications in this embodiment is 43% of the conventional number.

When performing pattern matching of P-dimensional target data using P-dimensional template data, the input block in FIG. 1 is defined as P-dimensional (P is a natural number) template data, and the reference search range data is defined as P-dimensional target data. Pattern matching can be performed by processing each step as P-dimensional data, such as converting a dimensional FFT to a P-dimensional FFT.

In this pattern matching, the method of the second embodiment can be used together. In this case, a process of converting the ft and fb obtained by the first complex number configuration processing in step S102 and the input complex number data fc into Q (Q is a natural number smaller than P) dimension data by scanning is inserted, and the second step in step S104 is performed. Can be realized by inserting a process of scanning the gt and gb obtained by the complex number construction process and the reference complex number data gc into Q-dimensional data, and processing the P-dimensional data in the subsequent steps as Q-dimensional data.

(Embodiment 5) Hereinafter, a motion vector detecting method according to still another embodiment of the present invention will be described with reference to the drawings.

The method for detecting a motion vector according to the present embodiment employs M
This is a method of searching for a field motion vector used in PEG2 or the like.

In the MPEG2 frame structure, a motion vector is also searched for in the reference search range data for each of the upper (top) field data and the lower (bottom) field data in the input block. These correspond to the search of the input top data ft (k, l) and the input bottom data fb (k, l) in the reference search range data g ＾ (k, l) according to the fourth embodiment.

Reference top data gt obtained by dividing the reference search range data into fields as in the fourth embodiment
Using (k, l) and reference bottom data gb (k, l), evaluation functions ett (m, n), etb (m, n), ebt (m ,
n), ebb (m, n) and e (m, n) described later are calculated, and a field motion vector and a frame motion vector are searched for.

Ett (m, n) is an evaluation function corresponding to the sum of squares of the difference between the input top data and the reference top data. etb (m, n) is an evaluation function corresponding to the sum of squares of the difference between the input top data and the reference bottom data.
ebt (m, n) is an evaluation function corresponding to the sum of squares of the difference between the input bottom data and the reference top data. ebb
(M, n) is an evaluation function corresponding to the sum of squares of the difference between the input bottom data and the reference bottom data. Also, e
(M, n) is an evaluation function corresponding to the sum of squares of the difference between the input extension data and the reference search range data.

Note that m = 0 to 16 and n = 0 to 8 in all of Expressions 109 to 112.

[0303]

(Equation 109)

[0304]

[Equation 110]

[0305]

(Equation 111)

[0306]

[Equation 112]

In addition, in accordance with Equations 54 to 56,
From 9 to 112, m = 0 to 16, n = 0, 2,.
For 6 (even number from 0 to 16), the following expression 113
, And m = 0 to 16, n = 1, 3,..., 15 (0
(Odd numbers up to 16), the following Expression 114 is configured.

[0308]

E (m, n) = et (m, n / 2) + e
bb (m, n / 2)

[0309]

E (m, n) = etb (m, (n-1) /
2) + ebt (m, (n + 1) / 2) Similarly to the fourth embodiment, Expressions 109 to 112 can be expressed as Expressions 115 to 118 below.

[0310]

[Equation 115]

[0311]

[Equation 116]

[0312]

[Formula 117]

[0313]

[Equation 118]

Expressions 109 to 112 and 115 to 11
8, the following equations 119 to 122 are obtained.

[0315]

[Equation 119]

[0316]

[Equation 120]

[0317]

[Equation 121]

[0318]

[Equation 122]

According to the formulas 119 to 122, the fourth embodiment
In the same manner as in the above, by the specific calculation method of Expressions 82 and 83, ett (m, n), etb (m, n), ebt
(M, n) and ebb (m, n) can be calculated.

FIG. 14 is a processing flowchart showing an example of the procedure of the motion vector detecting method of the present embodiment. The data is the same as in the fourth embodiment. In FIG. 14, the input block of the input image to be subjected to the motion vector detection is f
(K, l) (k = 0 to 15, l = 0 to 15), and the reference search range data of the reference image is g 、 (k, l) (k =
0 to 31, 1 = 0 to 31).

In FIG. 14, step S121 is an input block expansion process, step S122 is a first complex number construction process, step S123 is a first two-dimensional FFT process,
Step S124 is a second complex number construction process, step S124
125 is a second two-dimensional FFT process, step S126 is a frequency separation process, step S127 is a multiplication process, step S128 is an inverse two-dimensional FFT process, step S129 is a real part imaginary part separation process, step S130 is a square sum process, Step S131 is a calculation process, step S132 is a minimum value search process, and step S133 is a motion vector calculation process.

The input block extension processing in step S121, the first complex number construction processing in step S122, the first two-dimensional FFT processing in step S123, and step S12
4, the second complex number construction processing, the second two-dimensional FFT processing in step S125, the frequency separation processing in step S126, the multiplication processing in step S127, the inverse two-dimensional FFT processing in step S128, and the real part imaginary part separation in step S129 The processing is the same as in the fourth embodiment.

In the multiply-add processing of step S130, g
Calculate the sum of squares of the candidate field blocks corresponding to the positions of (m, n) in t (k, l) and (m, n) in gb (k, l), and obtain m = 0 to 16, For n = 0 to 8, the sum of squares represented by the following Expressions 123 and 124 is obtained. The square value obtained here is used in the calculation of the sum of squares of the candidate blocks corresponding to different (m, n) positions, and also in the calculation of the sum of squares of different input blocks.

[0324]

[Equation 123]

[0325]

[Equation 124]

In the calculation processing of step S131, the sum of squares obtained by the sum of squares processing of step S130 is
By subtracting the result obtained by the real part and imaginary part separation processing in step S129, for m = 0 to 16 and n = 0 to 8,
Evaluation functions ett (m, n), etb (m, n), ebt (m,
n), and ebb (m, n).

[0327]

[Equation 125]

[0328]

[Equation 126]

[0329]

[Equation 127]

[0330]

[Equation 128]

Also, m = 0-16, n = 0, 2,..., 1
6 (even number from 0 to 16), an evaluation function e (m, n) expressed by Expression 129 is obtained, and m = 0 to 16, n
= 1, 3,..., 15 (odd numbers from 0 to 16), an evaluation function e (m, n) represented by Expression 130 is obtained.

[0332]

E (m, n) = et (m, n / 2) + e
bb (m, n / 2)

[0333]

E (m, n) = etb (m, (n-1) /
2) + ebt (m, (n + 1) / 2) In the minimum search processing in step S132, step S
The evaluation function ett (m
t, nt) and etb (mt, nt), mt = 0 to 8, n
A search is performed for a minimum value of t = 0 to 8 and information on whether it is ett or etb and (mt, nt) are set as one of the search results, and the evaluation functions ebt (mb, nb) and ebb (m
(mb, nb) is searched for a minimum of mb = 0 to 8 and nb = 0 to 8, and information as to whether it is ebt or ebb and (mb, nb) are set as one of the search results, and e (m, n)
For (m = 0-16) and n = 0-16, search for (m, n) which is minimum. Based on these search results, two field motion vectors and a frame motion vector are calculated in the motion vector calculation process of step 133.

As described above, a motion vector according to equations 119 to 122 is obtained.

Here, the inverse two-dimensional FF of step S128
T processing, real part imaginary part separation processing in step S129, square sum processing in step S130, arithmetic processing in step S131, and (m,
n) or (mt, nt) and (mb, nb) are independent for each processing, and one set of (m, n) or (mt, n)
t), (mb, nb), or (m, n) or (mt, nt), (mb, n)
The method performed in b) or a combination of these methods may be used.

In the input block extension processing and the first complex number construction processing, the extension processing may be performed after the input block is divided into fields, or the extension and the complex number construction may be performed simultaneously. Further, the frequency separation processing may be performed on the result of the input Fourier transform, and the complex multiplication may be performed on the result of taking the complex conjugate of only the other.

In the field structure of MPEG2, the position in an input block consisting of 16 × 16 in a certain field is the upper (upper) 16 × 8 data and the lower (L
A motion vector in the reference search range data is also searched for every 16 × 8 data on the lower side. In this case, the field division and expansion processing of the input block in the frame structure are performed by 16 × 8 data on the upper side and 1 × on the lower side.
It can be realized in the same manner as the frame structure by substituting the upper and lower division into 6 × 8 data.

If the number of multiplications in each process is simply calculated for one input block (macro block), the result is the same as that of the fourth embodiment. In the case of N × N reference search range data,

[0339]

[Expression 131] It becomes 8MN ² + N ² times. Substituting N = 32 and M = 5 as an example gives 41
This is the number of actual multiplications of 984.

The method of the second embodiment can be used in combination with the method of the second embodiment. The ft and fb obtained by the first complex number construction process in step S122 and the input complex number data fc are scanned into one-dimensional data. Is inserted, and a process of converting the gt and gb and the reference complex data gc obtained by the second complex number construction process in step S124 into one-dimensional data by scanning is inserted.
This can be realized by processing the dimensional data as one-dimensional data.

If the method of Embodiment 3 is used together, the number of multiplications is further reduced. One input block (macro block)
When simply calculating the number of multiplications of each process for the above is the same as in the fourth embodiment, in the case of N × N reference search range data,

[0342]

## EQU1 ## The number is 16MN + 7MN ² + N ² times. By substituting N = 32 and M = 5 as an example, 39
This is 424 actual multiplications.

[0343]

As described above, according to the present invention, the similarity is calculated by using the sum of squares of a plurality of candidate block data, and the reference block is determined by using the calculation result. It is possible to provide a pattern matching method and a motion vector detection method in which the number of actual multiplications is smaller than in the related art.

[Brief description of the drawings]

FIG. 1 is a flowchart showing a procedure of a motion vector detection method according to Embodiment 1 of the present invention.

FIG. 2A is an explanatory diagram showing an example of an input block in an input image, and FIG. 2B is an explanatory diagram showing an example of a reference search range and a reference candidate block in a reference image for the input block.

3A is an explanatory diagram illustrating an example of a relationship between an input block and input extended data, and FIG. 3B is an explanatory diagram illustrating an example of a relationship between reference search range data and a candidate block;

FIG. 4 is a block diagram showing an example of a system configuration for executing the motion vector detection method according to the first embodiment;

FIG. 5 is a flowchart showing a procedure of a motion vector detection method according to the second embodiment of the present invention.

FIG. 6A is an explanatory diagram illustrating an example of an input block in an input image applied to the motion vector detection method according to the third embodiment of the present invention, and FIG. 6B is a reference search for the input block in a reference image; Explanatory drawing showing examples of ranges and reference candidate blocks

FIG. 7 is an explanatory diagram showing an example of a procedure for calculating reference search range data in the motion vector detection method according to the third embodiment.

FIG. 8 is a flowchart showing a procedure of a motion vector detection method according to the third embodiment.

FIG. 9 is an explanatory diagram showing another example of the procedure of calculating the reference search range data in the motion vector detection method according to the third embodiment.

FIGS. 10A to 10C show a fourth embodiment of the present invention.
5A to 5F are diagrams illustrating an example of a relationship between input extended data, input top data, and input bottom data applied to the motion vector detection method in FIG. Explanatory diagram showing an example of the relationship between reference top data and reference bottom data

FIGS. 11A to 11C are explanatory diagrams showing another example of the relationship between input extension data, input top data, and input bottom data applied to the motion vector detection method according to the fourth embodiment; (D)-(f) is an explanatory view showing another example of the relationship between reference search range data, reference top data, and reference bottom data corresponding to the input extended data.

FIG. 12A is a diagram illustrating a procedure for forming input complex data from input extension data in the motion vector detection method according to the fourth embodiment, and FIG. 12B is a reference search range in the motion vector detection method. Explanatory diagram showing the procedure for constructing reference complex data from data

FIG. 13 is a flowchart showing a procedure of a motion vector detection method according to the fourth embodiment.

FIG. 14 is a flowchart showing a procedure of a motion vector detection method according to the fifth embodiment of the present invention.

[Explanation of symbols]

 Reference Signs List 31 input block 32 input extended data 33 reference search range data 34 candidate block 81 CPU 82 first storage unit 83 second storage unit 84 image input unit 85 internal bus 201 input image 202 input block 203 reference image 204 input block position 205 Reference search range 206 Selected reference block S1 Input block expansion processing S2 First two-dimensional FFT processing S3 Second two-dimensional FFT processing S4 Multiplication processing S5 Inverse two-dimensional FFT processing S6 Sum of squares processing S7 Operation processing S8 Minimum value Search processing S9 Motion vector calculation processing

Claims (16)

[Claims] 1. A pattern matching method for performing pattern matching of P-dimensional target data using P-dimensional template data, where P is a natural number, wherein the template data is a size of the target data by using 0 data. Constructing the template extended data expanded to: performing a P-dimensional fast Fourier transform between the template extended data and the target data to obtain a template Fourier transform result and a target Fourier transform result; and the template Fourier transform result. Performing one complex conjugate and the other complex multiplication on the and the target Fourier transform result to obtain a multiplication result; performing an inverse fast Fourier transform of the multiplication result to obtain an inverse P-dimensional Fourier transform result; Sum of squares of multiple candidate data in target data Step a, the pattern matching method, which comprises the steps of performing similarity determination using a sum of squares of the plurality of candidate data and the inverse P-dimensional Fourier transform result of the calculation. 2. A pattern matching method for performing pattern matching of P-dimensional target data using P-dimensional template data, where P is a natural number, wherein the template data is 0 data and the size of the target data is Constructing the template extended data extended to Q, where Q is a natural number smaller than P, scan the template extended data and the target data and perform Q-dimensional fast Fourier transform on the result as Q-dimensional data; Obtaining a template Fourier transform result and a target Fourier transform result; performing one complex conjugate and another complex multiplication on the template Fourier transform result and the target Fourier transform result to obtain a multiplication result; Inverse fast Fourier transform of Obtaining a result; calculating a sum of squares of the plurality of candidate data in the target data; determining a similarity using the inverse Q-dimensional Fourier transform result and the sum of squares of the plurality of candidate data Performing a pattern matching. 3. A pattern matching method for performing pattern matching of P-dimensional target data using P-dimensional template data, where P is a natural number, wherein the target data is divided, and one is a real part and the other is an imaginary part. Using the result of dividing the template data and 0 data to expand the target complex data to the field-divided size of the target data, with one as a real part and the other as an imaginary part. Constructing data; performing a P-dimensional fast Fourier transform of the template complex data and the target complex data to obtain a template Fourier transform result and a target Fourier transform result; and the template Fourier transform result and the target Fourier transform One of the transformation results is defined as a first Fourier transformation result, and the other As a second Fourier transform result, separating the first Fourier transform result into a first divided Fourier transform result and a second divided Fourier transform result corresponding to each division, and the second Fourier transform result and the first One complex conjugate and another complex multiplication are performed on the divided Fourier transform result to obtain a first multiplied result, and one complex conjugate and the other are obtained on the second Fourier transform result and the second divided Fourier transform result. Performing a complex multiplication to obtain a second multiplication result; performing an inverse P-dimensional fast Fourier transform of the first and second multiplication results to obtain first and second inverse Fourier transform results; The first and second inverse Fourier transform results are separated into a real part and an imaginary part, and the first real part data, the first imaginary part data, the second real part data, and the second imaginary part are obtained. Get data Calculating a sum of squares of a plurality of candidate data in the target data; the first real part data; the first imaginary part data; and the second real part data; Performing a similarity determination based on the second imaginary part data and a sum of squares of the plurality of candidate data. 4. A pattern matching method for performing pattern matching of P-dimensional target data using P-dimensional template data, where P is a natural number, wherein the target data is divided into one real part and the other imaginary part. Constructing the target complex number data as: and expanding the template data to the field-divided size of the target data by using the result of the division of the template data and 0 data, so that one is a real part and the other is an imaginary part. Where Q is a natural number smaller than P, the template complex data and the target complex data are scanned to perform Q-dimensional fast Fourier transform on the Q-dimensional data, and the template Fourier transform result and the target Obtaining a Fourier transform result; and pairing with the template Fourier transform result. One of the elephant Fourier transform results is defined as a first Fourier transform result, and the other is defined as a second Fourier transform result.
Separating a Fourier transform result into a first divided Fourier transform result and a second divided Fourier transform result corresponding to each division of the first Fourier transform result; and the second Fourier transform result and the first divided Fourier transform Performing one complex conjugate and the other complex multiplication on the result to obtain a first multiplication result, and performing one complex conjugate and the other complex multiplication on the second Fourier transform result and the second divided Fourier transform result Obtaining a second multiplication result; performing an inverse Q-dimensional fast Fourier transform of the first and second multiplication results to obtain first and second inverse Fourier transform results; Separating the inverse Fourier transform result into a real part and an imaginary part, and obtaining first real part data, first imaginary part data, second real part data, and second imaginary part data When Calculating a sum of squares of a plurality of candidate data in the target data; the first real part data; the first imaginary part data; the second real part data; Performing a similarity determination based on the imaginary part data of (i) and the sum of squares of the plurality of candidate data. 5. A motion vector detecting method for detecting a relative position of a block similar to an input block of an input image in reference search range data of a reference image, wherein the input block is referred to using 0 data. Configuring input extended data extended to the size of the search range data; performing a two-dimensional fast Fourier transform between the input extended data and the reference search range data to obtain an input Fourier transform result and a reference Fourier transform result Step, performing one complex conjugate and the other complex multiplication of the input Fourier transform result and the reference Fourier transform result,
Obtaining a multiplication result; performing an inverse two-dimensional fast Fourier transform of the multiplication result; calculating a sum of squares of a plurality of candidate block data in the reference search range data; Performing a similarity determination based on a result of Fourier transform and a sum of squares of the plurality of candidate block data. 6. In the two-dimensional fast Fourier transform of the reference search range data, a result of the one-dimensional fast Fourier transform in a certain direction of the reference search range data corresponding to a certain input block is referred to as a reference search corresponding to another input block. The motion vector detection method according to claim 5, wherein the method is used in two-dimensional fast Fourier transform of range data. 7. A motion vector detecting method for detecting a relative position of a block similar to an input block of an input image in reference search range data of a reference image, wherein the input block is referred to using 0 data. Configuring input extended data expanded to the size of the search range data; performing a one-dimensional fast Fourier transform on the result of scanning the input extended data and the reference search range data into one-dimensional data; Obtaining a transformation result and a reference Fourier transformation result, performing one complex conjugate and the other complex multiplication on the input Fourier transformation result and the reference Fourier transformation result,
Obtaining a multiplication result; performing an inverse one-dimensional fast Fourier transform of the multiplication result; calculating a sum of squares of a plurality of candidate block data in the reference search range data; Performing a similarity determination based on a result of Fourier transform and a sum of squares of the plurality of candidate block data. 8. A motion vector detecting method for detecting a relative position of a block similar to an input block of an input image in reference search range data of a reference image, wherein the reference search range data is divided into fields. Real part, the step of configuring reference complex number data with the other being an imaginary part, and using the result of field division of the input block and 0 data to expand the reference search range data to the field-divided size, Configuring input complex number data with the real part and the other as the imaginary part; performing a two-dimensional fast Fourier transform between the input complex number data and the reference complex number data to obtain an input Fourier transform result and a reference Fourier transform result; One of the input Fourier transform result and the reference Fourier transform result is defined as a first Fourier transform result, and the other is defined as a first Fourier transform result. Separating a Fourier transform result into a first field Fourier transform result and a second field Fourier transform result corresponding to each field of the first Fourier transform result; and the second Fourier transform result and the first field Fourier transform. Performing one complex conjugate and the other complex multiplication on the conversion result to obtain a first multiplication result, and performing one complex conjugate and the other complex multiplication on the second Fourier transform result and the second field Fourier transform result. Performing a second multiplication result; performing an inverse two-dimensional fast Fourier transform of the first and second multiplication results to obtain first and second inverse Fourier transform results; 2 is separated into a real part and an imaginary part, and the first real part data, the first imaginary part data, the second real part data, Obtaining second imaginary part data; calculating a sum of squares of a plurality of candidate block data in the reference search range data; the first real part data; and the first imaginary part Performing a similarity determination on the basis of data, the second real part data, the second imaginary part data, and a sum of squares of the plurality of candidate block data. Method. 9. In the two-dimensional fast Fourier transform of the reference complex number data, a result of the one-dimensional fast Fourier transform in a certain direction of the reference complex number data corresponding to a certain input block is converted into a reference complex number data corresponding to another input block. 2
9. The motion vector detecting method according to claim 8, wherein the method is used in a two-dimensional fast Fourier transform. 10. A motion vector detecting method for detecting a relative position of a block similar to an input block of an input image in reference search range data of a reference image, the method comprising: Constructing reference complex data with the real part and the other as the imaginary part, and expanding the reference search range data to the field-divided size using the result of field division of the input block and 0 data, Part, forming the input complex number data as the imaginary part, performing a one-dimensional fast Fourier transform on the result of scanning the input complex number data and the reference complex number data into one-dimensional data, Obtaining a reference Fourier transform result; and converting one of the input Fourier transform result and the reference Fourier transform result into a second Separating a Fourier transform result and a second Fourier transform result into a first field Fourier transform result and a second field Fourier transform result corresponding to each field of the first Fourier transform result; One complex conjugate and the other complex multiplication are performed on the conversion result and the first field Fourier transform result to obtain a first multiplication result, and one of the second Fourier transform result and the second field Fourier transform result is obtained. Performing a complex conjugate and the other complex multiplication to obtain a second multiplication result; performing an inverse one-dimensional fast Fourier transform of the first and second multiplication results to obtain first and second inverse Fourier transform results Step; separating the first and second inverse Fourier transform results into a real part and an imaginary part; Obtaining data, second real part data, and second imaginary part data; calculating a sum of squares of a plurality of candidate block data in the reference search range data; Of the real part data, the first imaginary part data, the second real part data, the second imaginary part data, and the sum of squares of the plurality of candidate block data. And a motion vector detecting method. 11. A motion vector detecting method for detecting a relative position of an input block of an input image and a block similar to each of the results of field division of the input block in the reference search range data of the reference image, the motion vector detecting method comprising: Constructing reference complex data by dividing the search range data into fields and setting one as a real part and the other as an imaginary part; using the result of field division of the input block and 0 data to convert the reference search range data into a field; Expanding the input complex number data into one of a real part and the other as an imaginary part, and performing a two-dimensional fast Fourier transform of the input complex number data and the reference complex number data; Obtaining the input Fourier transform result and the reference Fourier transform result. One of the transformation results is defined as a first Fourier transformation result, and the other is defined as a second Fourier transformation result, and is separated into a first field Fourier transformation result and a second field Fourier transformation result corresponding to each field of the first Fourier transformation result. Performing one complex conjugate and the other complex multiplication of the second Fourier transform result and the first field Fourier transform result to obtain a first multiplication result, and performing the second Fourier transform result and the second Performing one complex conjugate and the other complex multiplication on the two-field Fourier transform result to obtain a second multiplication result; performing an inverse two-dimensional fast Fourier transform on the first and second multiplication results; Obtaining a second inverse Fourier transform result; separating the first and second inverse Fourier transform results into a real part and an imaginary part; Obtaining first data, first imaginary part data, second real part data, and second imaginary part data; and a plurality of non-field-divided candidate block data in the reference search range data. Or calculating a sum of squares of a plurality of field-divided candidate block data; the first real part data, the first imaginary part data, the second real part data, and the second real part data; Performing a similarity determination based on the imaginary part data and the sum of squares of the plurality of candidate block data. 12. In the two-dimensional fast Fourier transform of the reference complex number data, a result of the one-dimensional fast Fourier transform in a certain direction of the reference complex number data corresponding to a certain input block is converted to a reference complex number data corresponding to another input block. The motion vector detection method according to claim 11, which is used in two-dimensional fast Fourier transform. 13. A motion vector detecting method for detecting, in a reference search range data of a reference image, a relative position of an input block of an input image and a block similar to each of a result of field division of the input block, the method comprising: Constructing reference complex number data by dividing the search range data into fields and setting one as a real part and the other as an imaginary part; and performing field division on the reference search range data using a result of field division of the input block and 0 data Constructing input complex data with one as a real part and the other as an imaginary part, and one-dimensional high-speed scanning of the input complex data and the reference complex data into one-dimensional data. Performing a Fourier transform to obtain an input Fourier transform result and a reference Fourier transform result; One of the result of the Fourier transform and the result of the reference Fourier transform is a first Fourier transform result, and the other is a second Fourier transform result. The first field Fourier transform result and the second Fourier transform result correspond to each field of the first Fourier transform result. Separating the second Fourier transform result and the first field Fourier transform result into one complex conjugate and another complex multiplication to obtain a first multiplication result, Performing one complex conjugate and the other complex multiplication on the two Fourier transform result and the second field Fourier transform result to obtain a second multiplication result; and an inverse one-dimensional high speed of the first and second multiplication results. Performing a Fourier transform to obtain first and second inverse Fourier transform results; and the first and second inverse Fourier transform results. Into a real part and an imaginary part to obtain first real part data, first imaginary part data, second real part data, and second imaginary part data; Calculating a sum of squares of a plurality of candidate block data in the range data; the first real part data; the first imaginary part data; the second real part data; Of the imaginary part data and the plurality of candidate block data that are not divided into fields or the plurality of candidate block data that are divided into fields.
Performing a similarity determination based on a sum of squares. 14. A motion vector detection method for detecting a relative position of an input block of an input image and a block similar to a result of vertically dividing the input block in the reference search range data of the reference image, the motion vector detection method comprising: Constructing reference complex number data by dividing the search range data into fields and setting one as a real part and the other as an imaginary part; and dividing the reference search range data into fields using a result of vertically dividing the input block and 0 data. Constructing input complex data with one as a real part and the other as an imaginary part, performing a two-dimensional fast Fourier transform of the input complex data and the reference complex data, and obtaining an input Fourier transform result. Obtaining a reference Fourier transform result; and calculating the input Fourier transform result and the reference Fourier transform result. Separating a first Fourier transform result into a first Fourier transform result and a second Fourier transform result into a first field Fourier transform result and a second field Fourier transform result corresponding to each field of the first Fourier transform result; One complex conjugate and another complex multiplication are performed on the second Fourier transform result and the first field Fourier transform result to obtain a first multiplication result, and the second Fourier transform result and the second field Fourier transform are obtained. Performing one complex conjugate and the other complex multiplication on the result to obtain a second multiplication result; performing an inverse two-dimensional fast Fourier transform of the first and second multiplication results; Obtaining a Fourier transform result; separating the first and second inverse Fourier transform results into a real part and an imaginary part; Obtaining the first imaginary part data, the second real part data, and the second imaginary part data. Calculating the sum of squares of the candidate block data of the first real part data, the first imaginary part data, the second real part data, and the second imaginary part data; Performing a similarity determination based on the sum of squares of the plurality of candidate block data. 15. In the two-dimensional fast Fourier transform of the reference complex number data, a result of the one-dimensional fast Fourier transform in a certain direction of the reference complex number data corresponding to a certain input block is converted to a reference complex number data corresponding to another input block. 15. The motion vector detection method according to claim 14, which is used in two-dimensional fast Fourier transform. 16. A motion vector detecting method for detecting a relative position of an input block of an input image and a block similar to a result of vertically dividing the input block in the reference search range data of the reference image, the motion vector detecting method comprising: Constructing reference complex number data by dividing the search range data into fields and setting one as a real part and the other as an imaginary part; and dividing the reference search range data into fields using a result of vertically dividing the input block and 0 data. Constructing input complex data with one as a real part and the other as an imaginary part, and one-dimensional high-speed scanning of the input complex data and the reference complex data into one-dimensional data. Performing a Fourier transform to obtain an input Fourier transform result and a reference Fourier transform result; And a reference Fourier transform result, a first Fourier transform result, and the other a second Fourier transform result. A first field Fourier transform result and a second field Fourier transform result corresponding to each field of the first Fourier transform result And performing one complex conjugate and the other complex multiplication on the second Fourier transform result and the first field Fourier transform result to obtain a first multiplication result, and the second Fourier transform result And performing one complex conjugate and the other complex multiplication on the result of the second field Fourier transform to obtain a second multiplication result; performing an inverse one-dimensional fast Fourier transform of the first and second multiplication results; Obtaining a first and a second inverse Fourier transform result; and dividing the first and second inverse Fourier transform results into a real part and an imaginary part. Separating, obtaining first real part data, first imaginary part data, second real part data, and second imaginary part data; and performing field division in the reference search range data. Calculating a sum of squares of the plurality of candidate block data and the plurality of candidate block data obtained by field division; the first real part data, the first imaginary part data, and the second real part data And a step of performing a similarity determination based on the second imaginary part data and the sum of squares of the plurality of candidate block data.