JPH06133297A - Method and device for compensating motion of picture whose resolution is converted - Google Patents

Abstract

PURPOSE:To compensate the motion of a moving picture signal by magnifying/ reducing the signal at an optional magnification. CONSTITUTION:A fraction part of a motion vector is inputted from a longitudinal lateral separation device 73 to a coefficient decision device 97 in a lateral filter 74B of a motion compensation device and a filter coefficient required for real number motion compensation is fed to multipliers 94-0-94-N. The multipliers 94-0-94-N form a filter whose tap number is N with delay devices 93-1-93-N-1, the coefficient decision device 97 calculates a filter coefficient required for the real number motion compensation with an optional tap number M, the motion is compensated by each of the delay devices 93, each of the multipliers 94 and an adder 96 and its output is fed to a longitudinal filter 75. The processing in the longitudinal filter 75 is similar to that in the lateral filter 74B to compensate the motion in the longitudinal direction. Thus, the picture quality of a decoded picture subject to prediction coding is improved.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a motion compensation method and device capable of resolution conversion in a digital moving image signal.
Specifically, predictive coding and discrete cosine transform (DCT)
To provide a moving picture decoding method and apparatus which employs a novel motion compensator capable of reproducing a picture of arbitrary resolution (number of pixels) by inputting a bit stream of a digital moving picture coded by To do.

[0002]

2. Description of the Related Art Since a digital moving image signal is represented by a huge amount of data as compared with a document or a voice, in order to economically use a communication network or a storage device, the amount of this image data must be compressed. It needs to be transmitted or stored.

FIG. 3 shows a conceptual block diagram of the entire system for transmitting or storing image data. The digital moving image input from the input terminal 50 is encoded by the encoder 6
It is encoded by 1 and compressed. The output bit stream is communicated or accumulated according to the application (communication / accumulation unit 62). On the data receiving side, the image data is reproduced by the decoding device 63 and output from the output terminal 52.

First, a signal processing method in the encoder 61 will be described. Most of the conventionally known compression methods utilize the correlation of image signals, and for example, a hybrid coding method combining predictive coding and discrete cosine transform (DCT) has been widely researched and developed. This technology is at the core of the technology as an international standardization method for digital image coding methods being promoted by the International Standardization Committee (ISO). Also, much development has been done in terms of hardware.

In the hybrid coding system in which the predictive coding and the DCT are combined, moving image data is first compressed by the predictive coding by utilizing the correlation in the time direction. Specifically, each frame (image data at a certain time) is divided into several small blocks, and the moving distance from the adjacent frame is detected for each block. Hereinafter, this distance is referred to as a "motion vector". Next, after using this motion vector to compensate for the motion of each block, the difference between frames is calculated. This process is called "motion compensation", which reduces the amplitude value of the difference data and reduces the number of bits required for communication / accumulation per pixel.

The predictively coded data is then DCT
It is processed and further compressed using spatial correlation. The DCT is performed for each block, and a block size of size 8 × 8 (vertical 8 pixels, horizontal 8 pixels) is generally used in the above international standardization method. The "coefficient value" obtained by DCT is ubiquitous in the low frequency region because the image signals have a correlation. As a result, entropy (the amount of information per pixel) is reduced, and compression by Huffman coding is possible.

The DCT coefficient values are first quantized, then Huffman coded, and output from the encoder 61 in order to compress the data amount. The data compressed by the encoder 61 is then input to the communication / accumulation unit 62 and communicated or accumulated according to the purpose.

In the decoding device 63, the input bit stream is processed in the reverse procedure of the processing in the encoder 61, the image input to the encoder 61 is reproduced and restored, and output from the output terminal 52. To be done.

FIG. 4 shows a circuit configuration of the conventional decoding device 63. The decoding process here is basically performed in the reverse order of the encoding process. For example, as an example of encoding processing, ISO (International Organization for
Standardization) ISO-IEC /
JTC1 / SC2 / WG11, MPEG90 / 176,
Rev. 2 (Dec. 1990). Since this coding system is composed of 8 × 8 point DCT, predictive coding and Huffman coding,
In the following, it is assumed that an 8 × 8 size DCT is used on the encoding side.

The image data received at the input terminal 51 is temporarily accumulated in the buffer 20, taken out by a fixed amount, inputted to the inverse Huffman decoder 21, decoded, and converted into a DCT coefficient and an inverse quantization table. , A motion vector is obtained. The procedure is the reverse of the encoding process. As a result of the processing, the obtained DCT coefficient and the quantization table are sent to the inverse quantizer 22, and the motion vector for the unenlarged or reduced full-size image obtained by the inverse Huffman decoder 21 is converted to the motion vector converter 30. To, respectively.

In the inverse quantizer 22, the DCT coefficient is multiplied by a constant using the quantization table. The multiplier at this time is the reciprocal of the value in the encoder 61 (FIG. 3).

The dequantized DCT coefficient in the dequantizer 22 is supplied to the reduced block generator 231 included in the block generator group 23 via one of the terminals a, b and c of the changeover switch 31. , 8 × 8 block generator 232
It is input to one of the enlarged block generators 233 according to the processing purpose.

When the image is not enlarged or reduced, the inversely quantized DCT coefficient output from the inverse quantizer 22 is sent to the 8 × 8 block generator 232 via the terminal b of the changeover switch 31. It is input and placed in the same block state as the DCT output in the encoder 61 (FIG. 3).
The block size at this time is 8 × 8 pixels.

When the block size reduction conversion is performed, the DCT coefficient output from the inverse quantizer 22 is reduced by the reduced block generator 231 via the terminal a of the changeover switch 31.
Entered in.

FIG. 5 shows the position where the DCT coefficient value exists for explaining the output of the reduced block generator 231. The output signal of the reduced block generator 231 is N ₁ and N
₂ is a positive number less than 8 and 8 × 8 per block
Among the DCT coefficients, there are N ₁ × N ₂ coefficients on the low frequency side (the closer to the origin 0, the lower the frequency is shown) among the DCT coefficients. Reduction ratio at this time, the vertical N _1/8-fold, lateral N _2/8
Double.

On the other hand, when the size conversion is performed, D
The CT coefficient is input to the enlarged block generator 233 via the terminal c of the changeover switch 31.

In the expanded block generator 233, zeros are added to the higher frequency side of the 8 × 8 DCT coefficients per block dequantized by the inverse quantizer 22, and the size N ₁ × Generate N ₂ blocks. However, N ₁ and N ₂ are positive numbers larger than 8. Furthermore, expansion ratio in this case is vertical N _1/8 times, a lateral N _2/8-fold.

The blocks generated by the processing in the block generator group 23 are subjected to inverse discrete cosine transform (IDCT) for each block in the inverse discrete cosine transformer 24.
The block size at this time is N ₁ × N ₂ . Inverse discrete cosine transformer 2 obtained as a result of inverse discrete cosine transform
The output signal from 4 is added with the motion-compensated signal from the changeover switch 35 by the adder 40 and output from the output terminal 52 as a reproduced image.

The motion compensation processing by the adder 40 is performed by the I picture (Intra Picture) and P picture (Predictive Pictu) depending on the encoding processing of the output image from the adder 40.
re) and B picture (Bidirectional Picture)
Any of the street image types are used. I
In a picture, all blocks included in the picture are intra-picture coded and not predictive coded.
In the P picture, one of the inter-picture predictive coding and the intra-picture coding which has the smaller prediction error data is selected for each block, and in the B picture, the forward (past) or backward (future) I or P picture is selected. Used for prediction.

When the output image from the adder 40 is an I picture, the changeover switch 35 is connected to the terminal d, and the adder 40 adds nothing.

When the output image from the adder 40 is a P picture, the output signal of the inverse discrete cosine converter 24 is added by the adder 40 to the past image used for prediction, which is the output of the changeover switch 35. And output from the output terminal 52.

When a past image is stored in the frame memory 26-1 through the terminal a of the changeover switch 33, the accumulated image signal is motion-compensated by the motion compensator 27-1, and the changeover switch 35 is used. Is input to the adder 40 via the terminal a. For motion compensation at this time, the motion vector output from the motion vector converter 30 is used, and this is transmitted via the terminal a of the changeover switch 34 to the motion compensator 27.
-1 is input.

On the other hand, when past images are accumulated in the frame memory 26-2 via the terminal c of the changeover switch 33, the accumulated image signals are motion-compensated by the motion compensator 27-2 and switched. It is input to the adder 40 via the terminal c of the switch 35. For motion compensation at this time, the motion vector output from the motion vector converter 30 is used, and this is input to the motion compensator 27-2 via the terminal b of the changeover switch 34.

When the output image from the adder 40 is a B picture, the past and future images are stored in the frame memories 26-1 and 26-2, respectively, and the motion compensator 2 respectively.
Motion compensation is performed by 7-1 and 27-2. After that, both are input to the interpolator 28, and the arithmetic mean of the both or one of them is output to the terminal b of the changeover switch 35. The output signal of the switch 35 is added to the output of the inverse discrete cosine converter 24 by the adder 40, and the reproduced image is output from the output terminal 52.

Further, the motion vector converter 30, the values of the motion vectors used for motion compensation, the vertical direction N _1/8
Times, the laterally fold N _2/8, an appropriate value according to the conversion ratio required for reproduction image (enlargement or reduction ratio of an image) is generated. The value is also sent to the switch 34 as a real value. Here, the motion compensator 27-1 or 27-
The process in 2 will be further described.

FIG. 6 shows the internal circuit of the motion compensator 27-1 (or the motion compensator 27-2 having the same circuit configuration). The motion vector value, which is the output of the motion vector converter 30, is input to this circuit through the terminal a of the changeover switch 34. Also, the frame memory 26-1 (or 26-
The past or future image which is the output of 2) is input.

The motion vector value input via the terminal a of the switch 34 is separated into an integer part and a decimal part other than that by an integer separator 72 in both the vertical and horizontal directions. It is input to the container 71 and the vertical / horizontal separator 73. However, the decimal part is represented by a real value of 0 or more and less than 1.

In the integer motion compensator 71, the pixel block input from the frame memory 26-1 is motion-compensated with integer precision and then input to the horizontal filter 74. In the vertical / horizontal separator 73, the fractional part of the motion vector input from the integer separator 72 is separated into a horizontal value and a vertical value, which are input to the horizontal filter 74 and the vertical filter 75, respectively.

The horizontal filter 74 performs motion compensation in the horizontal real number precision, and its output is the vertical filter 75.
Entered in. The vertical direction filter 75 performs vertical direction real number precision motion compensation, and the output thereof is the motion compensator 27-.
The final output of 1 is applied to the interpolator 28 and the terminal a of the switch 35.

FIG. 7 shows a circuit configuration of the horizontal filter 74, and the operation will be described using this. The operation of the vertical filter 75 is the same as that of the horizontal filter 74 except for the processing direction, and the circuit configuration and operation are the same. In these filters, pixel blocks are motion compensated by a distance represented by the fractional part of the motion vector. For simplification of description, this motion compensation will be referred to as “real-value motion compensation” below.

In the horizontal filter 74, the vertical / horizontal separator 73 is used.
The fractional part of the horizontal motion vector input from the motion vector quantizer 90 is 0, 0.
5, 0.75, 0.25 to obtain the final output of one of the following equations. x (p + 0) = x (p) x (p + 0.5) = (x (p) + x (p + 1)) 2 ⁻¹ x (p + 0.75) = (x (p) + 3x (p + 1) )) 2 ⁻² x (p + 0.25) = (3x (p) + x (p + 1)) 2 ⁻² Thus, the real-value motion-compensated pixel value is calculated as the interior division point of two neighboring pixels. It

The absolute value of the fractional part of the horizontal motion vector input from the vertical / horizontal separator 73 is either 0, 0.5, 0.75 or 0.25 in the motion vector quantizer 90. If the quantized value is represented by s, the switch 91
Changes according to four values of s.

When s = 0, the switch 91 is connected to the terminal a, and the input signal is not processed at this time.

When s = 0.5, the switch 91 has the terminal b.
Connected to. When the pixel value at the position p (p is an integer) is represented by x (p), x (p) is the output x (p + 1) delayed by the delay device 93-a and the adder 96-a. They are added and then multiplied by 2 ⁻¹ by the shifter 92-a. As a result, (x (p) + x (p + 1)) 2 ⁻¹ is output via the terminal b of the switch 91 as a value x (p + 0.5) motion-compensated by 0.5 pixel.

When s = 0.75, the switch 91 is connected to the terminal c. As a result, x (p) is added by 3x (p + 1), which is the result of passing through the delay device 93-b and the multiplier 95-b, by the adder 96-b, and then 2 by the shifter 92-b. ^-Doubled . As a result, (x (p) + 3x (p + 1)) 2 ^-2 is output via the terminal c of the switch 91 as a value x (p + 0.75) that is motion-compensated by 0.75 pixels.

When s = 0.25, the switch 91 is connected to the terminal d. As a result, x (p) becomes 3x (p) through the multiplier 95-c, and the output x of the delay device 93-c is x.
After (p + 1) is added by the adder 96-c, it is multiplied by 2 ^-2 by the shifter 92-c. As a result, (3x (p) + x (p + 1)) 2 ^-2 is output via the terminal d of the switch 91 as a value x (p + 0.25) motion-compensated by 0.25 pixel. The image whose real-valued motion is compensated in this way is changed over by the changeover switch 91.
And is applied to a vertical filter 75 that operates in the same way as the horizontal filter 74, but differs in direction.

[0037]

In the prior art, when a predicted image is generated, the data once coded by the coding device is decoded again, and this is used as the predicted image to prevent the accumulation of prediction errors. is doing. However, this measure is effective only when the image is not enlarged or reduced. Therefore, when the image is enlarged or reduced on the decoding side, an error is accumulated in the predicted image, and the image is deteriorated each time the prediction is repeated, which leaves a problem to be solved.

In the prior art, the fractional part of the motion vector is quantized to any of 0, 0.5, 0.75 and 0.25. However, if it is reduced or expanded at an arbitrary magnification, a finer real value that cannot be represented by these four types of fixed approximations is generated. Therefore, when motion compensation is performed for such a fine position, there remains a problem to be solved that the pixel of the reproduced image is displaced.

[0039]

In a horizontal and vertical filter, when a fractional part of a motion vector generated by reduction or expansion is input from a motion vector of a full-size image that is not reduced or expanded, real-valued motion is detected. A coefficient determiner that calculates and outputs a filter coefficient required for compensation under an arbitrary number of taps, and a number of delay devices, multipliers, and additions required to configure the filter having the arbitrary number of taps. And a container.

[0040]

In the coefficient determiner, the filter coefficient required for the real-valued motion compensation is calculated under the arbitrary number of taps, and the motion compensation is performed by the required number of delay devices, multipliers and adders. ing. In the conventional method, as seen in the fixed approximation processing in the horizontal filter 74, two neighboring pixels are used at the time of real value motion compensation. This process can be regarded as a low pass filter process. However, when the image is enlarged or reduced, the sampling frequency changes, so that the characteristic of the filter becomes different from that of the low-pass filter. Therefore, in the present invention, the number of pixels used for the filtering process is increased to increase the degree of freedom of the filter characteristic, and the characteristic of the filter used for the real-valued motion compensation is dynamically determined as the optimum filter coefficient according to the reduction or enlargement of the image, By using the coefficient for correction, deterioration of image quality caused by accumulation of prediction error when the reproduced image is enlarged or reduced is suppressed.

Further, in the present invention, the filter coefficient value is calculated by using a given pixel, obtaining a continuous function that minimizes the error between them, and resampling this function at a required position. There is.

[0042]

1 is a block diagram of a lateral filter 74B according to the present invention.
The circuit configuration of one embodiment is shown. This is realized by replacing the horizontal filter 74 in the conventional decoding device shown in FIG. 7 with the filter 74B shown in FIG.

The difference from the horizontal filter 74 in the conventional decoding apparatus shown in FIG. 7 is that an optimum filter coefficient according to the reduction or enlargement of an image necessary for real-valued motion compensation is obtained under an arbitrary number of taps. A coefficient determiner 97 for dynamically calculating, and N-1 delay units 93-1 and 93-2 necessary for constructing the filter having the arbitrary number of taps,
93-3, ..., 93-N-1 and multipliers 94-0, 9
, 94-2, ..., 94-N-1, 94-N, and the adders 96-1, 96-2 ,.
This is the point used. Motion compensator 27 in FIG.
The other configuration of -1 is the same as the conventional technique. The operation brought about by the configuration different from these conventional techniques will be described in detail.

Embodiment 1 In the horizontal filter 74B according to the present invention shown in FIG. 1, the image signal x input from the integer motion compensator 71 (FIG. 6) is used.
(p) is a multiplier 94-0, 94-1, 94-2, ..., 9
4-N-1, 94-N and delay devices 93-1, 93-2, 9
3-3, ..., 93-N-1 and adders 96-1, 96
Filtered by -2, ..., 96-N-1. As a result, when the number of taps N is an even number, the final output from the multiplier 94-N is y (p) = bΣa _k x ((p + kN / 2) -1) z ^-k . Here, Σ represents addition from k = 0 to N−1.

On the other hand, when the number of taps N is odd, the final output is y (p) = bΣa _k x (p + k- (N-1) / 2) z ^-k . Here, Σ represents addition from k = 0 to N−1.

Next, a method of determining the filter coefficients a _k and b in the above equation will be described. The fractional part of the motion vector is input from the vertical / horizontal separator 73 to the coefficient determiner 97, and as a result, the filter coefficient required for real number motion compensation is output. Here, when the fractional part of the motion vector is s and the pixel value at the position t is x (t), the coefficient is calculated as follows.

FIG. 2 shows an explanatory diagram of the calculation method in the horizontal filter 74B of FIG. First, the pixel value x (t) at a given number of integer positions (t is an integer)
Is used to calculate the continuous function f (t) (t is a real number). Here, for f (t), an approximate expression that minimizes the error from the given number of pixel values is used. When the Lagrange interpolation function is used as an example, the coefficient L _k (t) is given by the following equation. f (t) = Σx (k) L _k (t) where L _k (t) = Π {t−m) / (k−m)}, and Σ is the cumulative sum from k = −N to M And Π represents the accumulation from m = −N to M excluding m = k.

In the above equation, the number of taps of the filter is M + N + 1
Is. Here, for example, when the number of taps is 4,
The continuous function is given by f ₄ (t) = where _{g 41 + g 42 + g 43} + g 44, g 41 = -x (-1) t (t-1) (t-2) / 6 g 42 = x (0) (t + 1) ( t-1) (t-2 ) / 2 g 43 = x (1) (t + 1) t (t-2) / 2 g 44 = x (2) (t + 1) t (t-1) / 6 this time, In FIG. 1, the multiplication coefficient b of the multiplier 94-N
When the value of x is 1, the coefficients of x (-1), x (0), x (1), and x (2) are a ₀ , a ₁ , a ₂ , ..., A _N-1 , respectively. Also, the coefficient a _k (k = 0.1, ..., N−1) is multiplied by b,
The hardware can also be scaled to a simplified value, such as an integer value.

On the other hand, in another embodiment, the tap number N is 3
If, then the continuous function is given by where _{f 3 (t) = g 31} + g 32 + g 33, g 31 = x (-1) t (t-1) / 2 g 32 = -x (0) (t + 1) (t-1) g 33 = x (1) (t + 1) t / 2 By substituting the motion vector fractional part s for t in these equations, the pixel value x (s) motion-compensated by the fractional position is calculated as shown in FIG. f (t) is required.

Embodiment 2 When the reduction ratio is 2 to 1, the fractional part of the value of the reduced motion vector is limited to any of 0, 0.5, 0.75 and 0.25. However, the sharpness of the reproduced image can be maintained by increasing the tap number N of the filter and correcting the filter characteristic. For example, using the 4-tap continuous function f ₄ (t) shown in the first embodiment, the decimal position, 0.5,
Representing 0.75 and 0.25, f ₄ (0.50) = {− x (−1) + 9x (0) + 9x (1) −x (2)} 2 ⁻⁴ f ₄ (0.75) = { -5x (-1) + 35x (0 ) + 105x (1) -7x (2)} 2 -7 f 4 (0.25) = {- 7x (-1) + 105x (0) + 35x (1) -5x (2)} It becomes ^2-7 . The transfer function of the filter this time, _{respectively, H 1 (z) = {} - 1 + 9z -1 + 9z -2 -z -3} 2 -4 H 2 (z) = {- 5 + 35z -1 + 105z -2 -7z - _{^{3} 2 -7 H 3 (z}} ) = {- a ^{7 + 105z -1 + 35z -2 -5z} -3} 2 -7. Here, instead of the above formula H ₁ (z), use a filter having a characteristic closer to an all-pass characteristic, H ₁ '(z) = { ^-1 + 5z ^-1 + 5z ^-2 -z ^-3 } 2 ^-3 You can also In addition, the transfer function described above is calculated as H ₁ (z) = {9 (z ⁻¹ + z ⁻² ) − (1 + z ⁻³ )} 2 ⁻⁴ H ₂ (z) = [− 5 + 7 {5 (z ⁻¹ + 3z ^-2 ) -z ^-3 }] 2 ^-7 H ₃ (z) = [7 {-1 + 5 (3z ^-1 + z ^-2 } -5z ^-3 ] 2 ^-7 If expanded to the number of multipliers 94 Can be reduced and the hardware can be simplified.

Embodiment 3 When the reduction ratio is 4: 3, the part of the value of the reduced motion vector other than the integer is i / 8 (i = 0, 1, 2, ..., 7).
Takes 7 types of values. However, in order to simplify the operation of appropriately selecting the filter coefficients corresponding to the seven types of values, the types of selection can be reduced. For example, four types, i / 4 (i = 0,
The coefficient determiner 97 quantizes and approximates any one of 1, 2, 3). At this time, the accuracy of the position of motion compensation is approximated to 1/4, but the sharpness of the reproduced image can be maintained by setting the number of taps of the filter to 3 or more.

The horizontal filter 74B has been described in each of the above embodiments, but when it is used as the vertical filter 75 in FIG. 6, the horizontal filter in FIG. 74B can be used as is.

[0053]

EFFECTS OF THE INVENTION In the decoding apparatus, the motion compensation of the enlarged or reduced moving image is performed with real number accuracy by using a filter that approximates the all-pass characteristic of 3 taps or more, and thus the decoded image of the predictively encoded image is obtained. Since the image quality can be improved, the effect of the present invention is extremely large.

[Brief description of drawings]

FIG. 1 is a circuit configuration diagram showing an embodiment of a lateral filter that characterizes the present invention.

FIG. 2 is a function diagram for explaining the operation of the circuit of FIG.

FIG. 3 is a conceptual configuration diagram of a conventional image signal encoding and decoding system.

FIG. 4 is a circuit configuration diagram of a conventional decoding device which is a component of FIG.

5 is a DCT showing the existence position of a low frequency side DCT coefficient value used for inverse discrete cosine transform in the configuration of FIG. 4;
It is a coefficient position diagram.

6 is a circuit configuration diagram of a motion compensator which is a component of FIG. 4.

7 is a circuit configuration diagram of a horizontal filter included in the conventional motion compensator of FIG.

[Explanation of symbols]

20 buffer 21 inverse Huffman decoder 22 inverse quantizer 23 block generator group 24 inverse discrete cosine transformer 25 field synthesizer 26-1, 26-2 frame memory 27-1, 27-2 motion compensator 28 interpolator 30 motion vector converter 31, 33-35 changeover switch 40 adder 50, 51 input terminal 52 output terminal 61 encoder 62 communication / accumulation unit 63 decoding device 71 integer motion compensator 72 integer separator 73 vertical / horizontal separator 74, 74B Horizontal filter 75 Vertical filter 90 Motion vector quantizer 91 Switch 92-a to 92-c Shifter 93-a to 93-c, 93-1 to 93-N-1 Delay device 94-0 to 94-N , 95-a to 95-c multiplier 96-a to 96-c, 96-1 to 96-N-1 adder 97 coefficient determiner 231 reduction Lock generator 232 8 × 8 block generator 233 enhanced block generator a ₀ ~a _N-1, b the filter coefficient f (t) fractional part x (t) pixel values of the continuous function s motion vector

Claims (2)

[Claims] 1. A motion compensation method (27) for a resolution-converted image, wherein motion compensation is performed on a digital moving image reduced or enlarged for a full size image using a motion vector value for the full size image. A continuous function (f (t)) having a minimum error is obtained for an arbitrary number of pixel values (x (t)) located in the vicinity of the pixel to be compensated, and the continuous function is set to the full size. A motion compensation method for a resolution-converted image, in which a fractional part (s) of a motion vector generated as a result of reducing or enlarging an image is substituted to perform motion compensation with real-value accuracy. 2. A motion of a resolution-converted image in which a digital moving image reduced or expanded with respect to a full-size image is motion-compensated by using a filter means (74, 75) for a motion vector value for the full-size image. In the compensator (27), the filter means receives the fractional part (s) of the motion vector resulting from the reduction or enlargement of the full-size image, and performs an arbitrary number of taps (N) required for real-valued motion compensation. Coefficient determining means (97) for calculating and determining the filter coefficient for taps (a) and the filter coefficient for output (b) based on the arbitrary number of taps, and The required number of serially connected delay means (93) and the output obtained at each tap of the serially connected delay means are multiplied by the filter coefficient (a) for the tap. Tap for multiplying means (94-0~94-N-1) that tap adding means for obtaining sequential addition to adding an output of said taps for multiplication means (96-1~96-N-
1) and an output multiplication means (94-N) for multiplying the addition output by the output filter coefficient.