JPH04207790A - Motion vector detector - Google Patents

Abstract

PURPOSE:To detect a motion vector based on easy judgement and by a simple operation by detecting a vector with the minimum degree of matching detected by block matching, arranging peripheral vectors sequentially, performing a correlative arithmetic operation with a reference wave, and generating output in accordance with correlation. CONSTITUTION:Only vector quantity (x,y) with the minimum degree of matching is obtained after performing the block matching. Thereby, the correlation detection of the peripheral vectors is performed. For example, at a remainder comparator 31, a position of deviation with the minimum remainder per block is detected, and it is outputted as the motion vector of a plate, and is supplied to a delay circuit group 32. An orthogonal transform circuit 33 for Fourier transformation, etc., and a peak detection circuit 34 are provided, and a decision signal when a peak exists at a spatial frequency(sf)=1, 2, and 3 of axis of abscissa is taken out to the output 38 of an OR gate 35. The decision signal is supplied to a sub blocking circuit, and also, the motion vector is outputted to the output of an AND gate 36.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a motion vector detection device, and particularly to a device for determining a motion vector from the result of block matching. The present invention also relates to a device that performs motion compensation using detected motion vectors.

[Summary of the invention]

The present invention provides a motion vector detection device that detects a motion vector by block matching that detects the degree of matching at corresponding positions between image signals that differ by one frame in time. Detects the minimum vector of A motion vector detection device is characterized in that it detects a vector as a motion vector and/or determines that a moving object contour exists within a block. The present invention can detect motion vectors with easier judgment and simpler operations than in the past. In particular, according to the present invention, motion compensation can be performed even at the edge of a moving object, and this block can be prevented from becoming blurred.

[Prior Art and Problems to be Solved] First, motion vector detection will be explained.

Here, it is assumed that one image (referred to as a frame; interlacing is not considered here) has H pixels horizontally and ■ lines vertically, as shown in FIG. Then, assume that the screen is subdivided into blocks of P pixels×Q lines. FIG. 27 shows one block. In this case, P=5 and Q=5. C indicates the center pixel position of this block.

Figure 28 shows the positional relationship between a block whose center pixel is C and a block whose center is C'. When the former is a block of interest in the current frame, the block that matches that image is the It is assumed that the block is located at the center of C' in the frame. FIG. 28A shows the case of a motion vector of ±1 pixel in the horizontal direction and +1 line in the vertical direction. FIG. 28B shows the case of a motion vector of ±3 pixels in the horizontal direction and +3 lines in the vertical direction. Figure 28C shows ±2 pixels in the horizontal direction and 11 pixels in the vertical direction.
This is the case for line motion vectors. - Generally, a motion vector is determined for each block of the current frame.

When the motion vector detection range is ±8 pixels horizontally and ±T lines vertically, a block in the current frame is shifted horizontally by ±S and vertically by ±T from its center C. must be compared with blocks of various previous frames with C'. FIG. 29 shows that when the position of the center C of a block in the current frame is R, it is necessary to compare it with (2S+1)×(2T+l) blocks of the previous frame to be compared. That is, all blocks of the previous frame having C' in the square position in FIG. 29 are to be compared. FIG. 29 shows an example in which S=4 and T=3.

An example of P=5, Q=5, S=4, and T=3 will be explained below.

Looking at it from a different perspective, if we consider a certain block of interest in the current frame to be the 5 pixels x 5 line position in the inner center of Figure 30, then horizontally ±3 pixels and vertically ±T centering on the same position in the previous frame. Outside of Figure 30 including the line (P+23
) Pixel

Motion vector detection is often
Sometimes I think about thinning things out. That is, in the above example, horizontal ±
Although it is assumed that all positional relationships are compared within the range of 3 pixels and vertical ±T lines, some combinations of comparisons may be omitted. However, in the following explanation, for the sake of simplicity, all combinations will be compared.

FIG. 31 is a conceptual diagram of the configuration of motion vector detection. Data a supplied from the input terminal 101 is the current frame image data, and data b is 1
This is the image data of the previous frame that has been delayed by a frame.

The data b of the previous frame is converted into data with various positional relationships as shown in FIG. 28 by the delay unit 103, and each of the data is compared with the data a. The block comparing unit 104 outputs the degree of matching (degree of matching) of images for each shift position, and the determining unit 105 compares the matching degrees and outputs the most appropriate amount of shift, that is, a motion vector, to the output terminal 106. Output.

The motion vector detection hardware shown in FIG. 31 requires a large capacity memory in addition to the frame memory 102, the delay unit 103 and the block comparison unit 104, and if the comparison is not thinned out, the block comparison unit 104 The arithmetic unit 105 also becomes large-scale, and the determination unit 105 also becomes large-scale depending on the method used. Therefore, there is a problem in that a system in which motion vector detection is used requires a considerable amount of hardware.

It has also been proposed to detect motion vectors using software (see Japanese Patent Laid-Open No. 105178/1983). However, due to constraints on software processing speed, a considerably simplified comparison and determination method must be used, resulting in the problem of low accuracy of detected motion vectors.

The present invention relates to reducing the amount of hardware for the delay section 103 and block comparison section 104 excluding the determination section 105 in FIG. 31, particularly for the delay section 103.

FIGS. 32 and 33 show the first conventional example. FIG. 32 corresponds to the frame memory 102 and delay unit 1030 portions in FIG. 31.

In FIG. 32, F delays one frame of input data in the frame memory 102. H is a delay circuit for one horizontal scan (line). Moreover, a square without a letter is a delay circuit for one pixel, that is, a register. Furthermore, (H-
9) is a delay circuit that has 9 pixels less than the number of pixels for one line, and together with the registers for the other 9 pixels, it becomes a delay for one line.

The input data a of the current frame is the delay circuit 10 surrounded by a broken line.
7, the signal is delayed by 3 lines and 5 pixels, reaches the center position R of the block, and is output to the output terminal 108.

Data b has a difference of one frame from data a, so data a and b are exactly the same pixel order! This is the data.

Then, as shown in Fig. 32, nine stages of registers and (H-9
) delay circuit is connected in series, so the tap of C(0,0) delayed by 3 lines and 5 pixels from data b is
There is a time difference of exactly one frame from R.

Assume that a horizontally scanned data string comes to the input terminal 101 in FIG. 32, similar to a normal image signal.

Then, when I stopped at a certain timing, (C(i.

j)) (i=-3n+S, j=-Tn+T), the pixels in Figure 29 around that position in the previous frame are just obtained for the R pixel in the current frame. There is.

For example, C(1,0) is older by one pixel horizontally than C(0°0)(R), and if Figure 29 is a display screen, the data of the pixel to the left of R can be obtained. .

To obtain a motion vector, the following processing is performed using each tap in FIG. That is, (C(i,j))(
For each tap of i=-3n+S, j=-Tn+T),
Since data for each shifted position in Figure 29 can be obtained,
The block comparison unit 104 compares each tap with R, calculates the degree of match (degree of matching of images), and calculates each (i.

Obtain the result for each j). Then, the determination unit 105 determines the result and determines a valid motion vector (i).

Select j). In some cases, j and j may be determined as real numbers rather than integers.

So, next is how to find the degree of match for each (i, j).
This is done with the configuration shown in FIG. This circuit of FIG. 33 is required for each C(t,j). In other words, since block comparison is not thinned out, the circuit in FIG. 33 is (2S+1)X
(2T+2) pieces are required.

In FIG. 33, for each C(i, j), we want to perform a block comparison as shown in FIG.
, j), each pixel of the block shifted by one pixel horizontally and j lines vertically in the previous frame is obtained, and the difference is calculated by the subtraction circuit 109 for each corresponding position in both blocks.
The absolute value conversion circuit 110 takes the absolute value and converts it into the total number of pixels constituting one bronc, that is, PXQ, and the addition circuit 11
1 and memory 112, and this cumulative output is expressed as (
i, j).

Note that although an example is shown in which the absolute value of the difference (ABS) is taken, the square of the difference, etc. may also be used.

The explanation so far is still incomplete.

This is because the input in Fig. 32 is horizontally scanned, and basically the same horizontally scanned data is output to the taps of R and each C (i, j) with only a temporal shift. The problem is that it is not possible to compare each block as shown in the figure. However, this problem is solved by having the memory 112 in the accumulator of FIG.

If R and C(i, j) in FIG. 33 are re-inputted, each corresponding pixel of the block like each combination in FIG. 28 can be obtained in order, but since horizontal scanning is performed, C(
After each pixel of the first line of the block is input to the input of i, j), the data is used for comparison of the adjacent block without moving to the next line, and the horizontal After proceeding by H/P blocks lined up in , it is finally possible to compare each pixel in the second line of the original block comparison. Therefore, the accumulation for obtaining the degree of matching can be suspended midway through the calculation for each line, and can be changed to the accumulation for comparing adjacent blocks. For this purpose, a memory 112 capable of storing 872 or more words is provided as an accumulation memory.

FIG. 34 explains this memory 112.
= H/P, and bk (k=0~n-
1) is the content of each address for storage during the calculation.

Further, in FIG. 33, 113 is an address ordering circuit, and this address generation circuit 113 increments the address every P cycles corresponding to the number of horizontal pixels constituting one block, and sets the address to "0" after completing one horizontal scan. ”, and the memory address allocation circuit as shown in FIG. 34 can be controlled.

Furthermore, the accumulation for block matching needs to be reset for each H/P number of blocks lined up horizontally on the screen. This reset is from b0 to memo IJ 112.
b, 1-1 may all be set to 0'', or the address generation circuit 113 may be configured to select an address Z in which 0 is stored in advance in the memory when accumulating the first pixel in each block comparison. You may do so.

In order to obtain the data of the position in Figure 32 (3H+5)
However, this can be omitted by replacing the frame delay (F) with a delay circuit of (F=-3H-5).

As shown in FIG. 34, as a memory 112 for accumulation,
The number of words n = H/P is prepared because a motion vector is to be determined for each block in FIG. If you want to obtain one motion vector per frame, one word is sufficient.
The number of words required varies depending on the policy.

Although the circuits shown in FIGS. 32 and 33 are feasible and relatively easy to control, the problem is that they require large-scale hardware.

Among the hardware, the frame memory 102 and the arithmetic circuit (109, 110, 1
11) is essential in principle. (H-9) The large number of delay circuits and the memory 112 for each H/P word for each accumulation are a problem in implementation, that is, they are only necessary for the structure and are not essentially necessary. It is desirable to eliminate this by using a different configuration method.

These delay circuits and memories are required each time the input data is horizontally scanned.

That is, a memory is required in order to compare blocks as shown in FIG. 28 with input that is horizontally scanned as in the normal case of FIG. 35.

Therefore, in order to avoid this, it is sufficient to use a scanning method adapted to each block, that is, to input a data string that is horizontally scanned block by block, as shown in FIG. 35B.

Therefore, a second conventional example that can be considered is to use data scanned block by block as shown in FIG. This is a method of creating a data string that scans within a wide area outside the block shown in the center of FIG. 30 and comparing the blocks.

In this case, since both data strings have different numbers of data, the following measures are required. In other words, one block comparison unit for data string a is (PXQ) pixels, while for data string A, it is ((2S+p)X (2T+Q)) pixels, so the clock is increased by that ratio. The operation is performed according to the time chart shown in FIG. 36. That is, when scanning as shown in FIG. 35B, the data string a should have a timing as shown in FIG. 36A, but by increasing the rate of this data string a, the timing as shown in FIG. Timing. The frequency ratio of this clock is (2S+P) ・ (2
T+Q) 10P-Q. As a result, there will be time slots where no data exists. On the other hand, the data string side is as shown in FIG. 36C.

In FIG. 37, a scan conversion circuit 115 is a circuit that converts an input data string a as shown in FIG. 36B, and a scan conversion circuit 116 converts an input data string a as shown in FIG. 36C. This is a circuit that converts As shown in FIG. 37, this scan-converted data string b' is supplied to a delay section consisting of a (2S+1)×(2T+1) stage shift register similar to that in FIG. 32, and is configured as shown in FIG. A block comparison is made. This second conventional example
The delay circuit (H-9) in the conventional example (FIG. 32) becomes unnecessary, and the memory of the accumulator in FIG. 33 only needs to be one word at a time.

However, as can be seen from FIG. 36C, data row b' has (P+23) pixels horizontally, with an extra 2S pixels.
Further, in the vertical direction, there are similarly 2T extra lines, so control is required to stop the accumulation operation at the extra pixels.

In Figure 37, P-32 with 102' means a delay that is 32 pixels less than the frame delay, and R and C (0
, 0) to have a one frame difference in timing.

The block-by-block scan conversion from television scan (PXQ) is to convert the scan as shown in Figure 35A to the block-by-block scan as shown in Figure 35B. It requires a minimum of line memory and is not very small in terms of hardware (although it is not very small in terms of hardware).
Scan conversion of PXQ as in FIG. 36A is not necessarily burdensome since the scan of FIG. 5B is required. However, as shown in FIG. 36B, if the data string is a', or as shown in FIG. 36C, (P+23)/(Q+2
The scan conversion in 7) is not used for other processing and is only for this purpose and is a burden.

To be precise, this (P+23)/(Q+27) scan conversion is performed by scanning as shown in FIG. 35B while overlapping 23 pixels horizontally and 2T lines vertically.

In the second conventional example mentioned above, the (P+23) x (Q+2T) stage shift register is simpler in terms of hardware than the (H-9) delay circuit group, so compared to the first conventional example, the delay section is It is simple, and the accumulator memory can be just a register, which has a considerable effect. On the other hand, PXQ scan conversion and (P+23) ・ (Q+2T
) scan conversion is required, and for that purpose (H-9
) It requires more memory than the delay circuit group, and after scan conversion, the data rate increases by a factor of (P+23) There is a problem.

Furthermore, motion vector compensation is performed using the detected motion vector in various image processes.

This is because in the case of processing that cannot be performed only on one frame or field, processing between frames or fields cannot be performed unless the movement of the moving image on the screen is taken into account. Specifically, it is often needed for image band compression, but it is also needed for Y/C separation and non-interlacing.

Motion vector compensation first compares two frames (or fields) to obtain the motion vector, that is, the deviation between the two images, and then adjusts either frame (field) by that amount.
The idea is to shift the two so that they overlap.

Originally, in a video, a moving object not only simply moves in parallel, but also rotates and scales, but current image processing technology detects only the bones in translation in the form of motion vectors and attempts to take this into account. There is. As described above, the general method for detecting motion vectors is to divide the screen into small square blocks and obtain motion vectors for each block. The problems that arise when detecting motion vectors and performing motion compensation are (1) how to judge when the function of the degree of matching has many minimum values as a result of block matching; and (2) how to judge the block. Since only one vector is obtained for each movement, it is out of the question if there are multiple movements, but even simple movements cannot save blocks that form the outline of a moving object or blocks that contain small isolated moving objects.

For example, when a car is moving to the right as shown in Figure 39, on this screen: (1) In blocks C-c and C-d, the movement of the car can be captured by block matching.

(2) B-b, B-c, B-d, B-e, C-
The blocks b, C-e, D-b, D-cSD-d, and D-e correspond to the outline of a moving object, and since there are two motion vectors in each block, it is troublesome to determine the motion vector.

(3) Blocks other than those in (1) and (2) above are stationary and can capture motion vectors.

Conventionally, the determination unit 105 in FIG. 31 detects a motion vector using the procedure shown in FIG. 41, and performs motion compensation using this motion vector. This determination unit 105
As shown in the figure, it consists of a residual comparison circuit 120 to which a residual error for each position of each shift after block comparison is supplied, and a comparison and noise removal circuit 121. The comparison and noise removal circuit 121 receives from the residual comparison circuit 120 the position and amount of the minimum value with the smallest residual, the position and amount of the minimum value with the second smallest residual, and the third smallest residual. The location of the local minimum value and its amount are provided. Comparison and noise removal circuit 12
In step 1, what seems to be noise is removed and the residuals are compared in size. As a result, a determination signal indicating whether or not motion compensation is possible and a motion vector passed through the AND gate 123 are output.

As shown in FIG. 41, this conventional determining section performs block matching (step 124) and detects the minimum degree of matching (step 125). The vectors of these minimum points are sorted in order of degree of matching (step 126).
), a significant minimum point (it is determined that a moving object with that vector exists) is determined (step 127). The peak number of this minimum point is checked (step 128), and if the number of peaks is one, motion compensation is performed (step 12).
9), in other cases, motion compensation is not performed.

In such a determination procedure, various algorithms can be used to detect local minima, sort local minimum point vectors in order of degree of matching, and determine significant local minimum points.
It is generally a complex operation.

In other words, it is not a uniform, simple, repeated operation, but an operation that requires a lot of judgment (the next operation is determined by the judgment).

When the range in which motion vectors are searched for in block matching is ±S pixels horizontally (X axis) and 10T pixels vertically, the amount of motion vectors (X) for each motion vector.

y) (x=±s, y=±T), the frame differences (field differences) of the image blocks are accumulated, and three-dimensional data as shown in contour lines is obtained as shown in FIG. The minimum value in this three-dimensional data is the minimum value of the degree of matching.

For example, in FIG. 40A, a small value exists at a vector that is in the background during camera panning or inside a moving object. In FIG. 40B, two minimums (two vectors) exist, such as in the case of a moving body contour.

Note that the moving body contour section is, for example, the fourth section D-b in FIG. 39.
This refers to a block like the one shown in Figure 2.

Disadvantages of the conventional motion vector determination unit and motion compensation are as follows: (1) The motion vector detection procedure, that is, the procedure for determining a significant minimum point from the distribution of the degree of matching as a result of block matching is complicated.

(2) Since only one vector can be obtained for a block,
The outline of a moving object cannot be compensated for, and therefore a moving image is likely to have a blurred outline.

That's true.

This invention aims to eliminate such drawbacks.

Furthermore, interframe coding using motion compensation is known. Interframe encoding is used for band compression for transmitting image signals, and an encoding system that combines motion compensation and interframe difference is simply realized with a configuration as shown in FIG. 43. For the operating principle of FIG. 43, etc., please refer to the following literature, for example.

(Multi-dimensional signal processing of rTV images, written by Fukinuki, Nikkan Kogyo Shimbun, pages 266-280, for example, Figure 7 on page 274 of this page)
゜ In Fig. 43, 132 denoted as DCT is a discrete cosine transform circuit, 133 denoted as IDC, T is its inverse transform circuit,
135 is a frame memory that creates a frame delay. The input data is supplied to the motion vector detection circuit 136 and also to the subtraction circuit 131, where it is subtracted from the local decoded output of the previous frame from the frame memory 135.

Local decoding is formed by an inverse transform circuit 133 and an adder circuit 134. The encoded data generated from the discrete cosine transform circuit 132 is encoded by a code assignment circuit 137 using a Huffman code or the like to reduce the amount of data.

In such a situation, frame memory (FM) motion compensation is performed by using the motion vector detection circuit 1 outside the loop in the center of Fig. 43.
36, and is realized by the motion vector that is the output. As mentioned above, the motion vector detection circuit 136
Frame memory 13 required by the band compression device (Figure 43)
In addition to 5, there is a problem in that a frame memory is required.

Therefore, it is a problem that a frame memory is required for each of motion vector detection and band compression.

The present invention aims to provide a band compression device that combines both of these functions and can be configured with a single frame memory.

[Means for Solving the Problems] The present invention provides a scan conversion means for generating a first data series in which image data is converted from normal horizontal scanning to scanning within a block of P pixels x Q lines; a frame memory that generates a frame time difference and generates a second data series, a shift register is connected to one of the first and second data series, and a certain tap of the shift register is used as a reference; Set an image block that constitutes the (PXQ) pixels back from the tap and a block that is shifted horizontally by i and vertically by 1 pixel, and when the block does not shift from each pixel of that block, it becomes the delay difference to the reference tap. This motion vector detection device is characterized in that a tap delayed by the amount of delay is selected and compared with the other end of the frame memory to check the degree of match.

[Effect]

The memory capacity required to generate the delay required for block matching can be significantly reduced.

〔Example〕

Embodiments of the present invention will be described below with reference to the drawings. The following description describes a preferred embodiment for block matching, a preferred embodiment of a determining unit when detecting a motion vector, and a preferred embodiment of an interframe coding combined with frame compensation, that is, a preferred embodiment of a band compression device. are described in order.

First, an embodiment of the present invention relating to block matching will be described. In this embodiment, it is assumed that the input data has already been subjected to PXQ scan conversion. That is, it is assumed that the data strings shown in FIG. 35B and FIG. 35A are input. Note that, as mentioned earlier, one PXQ scan conversion is not burdensome.

If the circuit shown in FIG. 1 is used, the periphery of the output pixel of R in the previous frame will be obtained for each tap in a form where the screen is turned upside down, left and right.

Here, the frame memory 102 waits for three outputs, and the output e is delayed by 5H and 50 pixel layers from one frame (F-
5) (-50), the output f is delayed by 50 pixels less than one frame (F-50), and the output g is delayed by 5H more than one frame by 50 pixels (F+5H-50).

Further, the shift registers connected to each of e, f, and g are connected in a scanning manner within a block, and each of them has a tap. Note that 50 pixels is twice as many as PXQ. A tap delayed by 50 pixels from f, indicated as d in FIG. 1, is a tap with a time difference of exactly one frame from R.

Note that it may be considered that the same circuits as before are continued for e', f', and g', but as will be described later, this is not necessary. Although f and g can be obtained by adding a delay circuit to e' and f', here it is assumed that they are obtained from a frame memory.

First, considering the timing at which the first data of a block in the current frame appears in R, the next 25 pixels are each pixel in that block, and at the same time, in d, the previous data at the same position as that block on the screen appears. The first data of the block of the frame appears, and for the next 25 pixels, data whose positions on the screen correspond to the R side are sequentially obtained.

Therefore, if this 25 pixel period is compared in blocks using the circuit shown in FIG. 33 in which the memory 112 is simply a register, the same degree of matching as in the case of the C(0,0) tap will be obtained.

The second step is to abbreviate each tap in Figure 1 according to its arrangement.
It is a diagram. Here, the separation by two lines is for niblocks (
PXQ=25 pixels) is a shift register stage separation. Even after FIG. 2, keep in mind that the connections are always as shown in FIG. 1. In a diagram like FIG. 2, compared to the screen, the arrangement of pixels is reversed vertically and horizontally.

Now, if there is a shift of 12 pixels in the horizontal direction, that is, C(
-2,0) and R.

Even in this case, the R side is considered from the timing when the first data of a certain block of the current frame appears on R.

The 25 pixels of the previous frame corresponding to that block are then in the register in the dotted line area of FIG.

Now, the block shifted by 12 pixels in the horizontal direction is in the register in the solid line area at the same time. This area straddles two lines, and since the shift register is not connected horizontally between these two lines, it is necessary to consider the right side and left side of these two lines separately.

First, on the right side, e', which is horizontally shifted by 12 pixels from d in FIG. 1, may be used as the tap for C(-2,0).

However, when tapping e', the three pixels following the first pixel of the block are good, but the next two pixels are outside the area, and the next three pixels are valid, and the next two pixels are outside the area. .

Next, regarding the part on the left, if we consider the pixel on the top of the diagram and the bottom part of the figure, we can see that the first data of the block is e'
This is actually a C# tap because it should appear on the tap of e' three cycles after it is in the current state. In other words, it can be seen that C# can be used as a tap for this left part. Then, at the timing when tapping e' leaves the solid line area, select C''.
' and C'' can be selected.

In the end, block matching with a horizontal shift of -i pixels (0<i≦S) uses two taps in the shift register, one tap i taps before d and one PXQ-i taps before d. By selecting every i cycle and every i cycle, necessary data can be supplied to the comparator circuit.

Similarly, if we consider a case where there is a vertical shift of, for example, -2 pixels,
- Since the area indicated by the dotted chain line must be considered as a comparison block of the previous frame, it will now straddle another shift register column above and below. However, it is immediately clear that in this case, C'''' and C'''' should be selected based on the same considerations.

In other words, block matching vertically shifted by -j pixels (0<j≦T) is performed using j×Q taps and d
Tap before (HXQ-(Q-j)XP) tap (
In other words, the latter is delayed from e by the same delay as from f to d, from d' (tap after (Q-j)XP tap) in the first half j × P cycle and the second half (Q-j) All you have to do is switch.

The problem is the general case of a shift of 11 pixels horizontally and -J pixels vertically, but let's consider, for example, a case where there is a shift of 13 pixels horizontally and 12 pixels vertically. In this case, this means the area indicated by the solid line in FIG. 3, and this must be block-compared with the area indicated by the dotted line on the R side.

The solid line area in FIG. 3 now straddles two vertical and horizontal lines and is divided into four parts. That is, if we consider the shift registers in the shift register shown in FIG. 1, they are divided into various parts. The time chart in FIG. 4 shows this situation.

FIG. 4A is a clock. R in FIG. 4B indicates the timing at which the data of the current frame in the dotted line area in FIG. 3 is collected as "H". This lasts PXQ=25 cycles. In conjunction with this, the data in the solid line area in FIG. 3 is shown in d and d' in FIG. 4, C and D.

The numbers 00 to 44 used in Figure 4 indicate where each pixel of the block compared on Figure 3 is located when the first pixel of the block is placed in R, i.e. in the first cycle of Figure 4. It shows that. This OO~25 of 44
If the pixels are not lined up like y in Figure 4, it will not be possible to compare them with R using an arithmetic unit as shown in Figure 33.

The solid line area in FIG. 3 is divided into four parts by two lines: lower right, lower left, upper right, and upper left. 0O101 in the lower right part,
There are pixels 10, 11, 20, and 21, which are C'
When viewed from the tap, the timing matches the data string of R.

In the lower left part is 02.03.04.12.13.14.
There is 22.23.24. This idea is similar to the previous example when there is only a horizontal shift, and the tap of C'' matches the timing of R.

The upper right part has 30, 31, 40, 41, and this is the same idea as the previous vertical slip, and the timing matches R with the tap of C''''.

The upper left part has 32, 33, 34, 42, 43, 44. For example, 32 is shifted by 12 pixels horizontally and 13 pixels vertically from the position 00 on the screen, and if the block is the dotted line area If it does not protrude from one shift register column like this, it is at a delay amount of 17 stages before the 00 position, so C "" which is 17 stages behind from the 32 position in the actual shift register connection direction. You can see that tapping can be used instead.

These things can be clearly seen by looking at the time chart in Figure 4. C', C", C"', C"' for d and d'
Since this is simply a time lag relationship, the difference in FIG. 4F depends on the number of stages of the shift register.

C′, C#, C″′, and C#″ of G and H are obtained. The data string of y is C' to C at the timing shown below in Figure 4.
#” should be selected.

-When formed into a film, -i pixels horizontally (0<i≦S), vertically-
Comparison with a block shifted by j pixels (0<j≦T) is (1) C' which is (j X P + i) steps earlier than d.
(2) From d ((PXQ+jxP)-(P-i)) Tap shown by C" in front of the pitcher (3) From d'(jXF' (P-i)) Finally tap shown by Crt' (4) From d', PXQ-((Q-j)XP0(P-
j)) Using the four taps indicated by C'''' in front of the pitcher, you can select their output appropriately and compare it with the R output. The selection procedure is (1) In the range of (Q-j) XP cycles from the beginning of the block, (P-i) cycle is C' and the next i cycle is C
Repeat the selection.

(2) In the remaining jXP cycles, repeat the selection of (P-i) cycle as C''' and the next i cycle as C#''.

The above is summarized below.

For block comparison with the block of the previous frame shifted by -i pixels horizontally and -j pixels vertically, (0≦i≦S), (0≦j≦
T) The required data strings are C' (-i, -j) and C in Figure 5.
'' (-i, -j), C''' (-i, -j), C'' (
-i, -j) by switching a maximum of four taps.

Also, in Figure 5, the previous position of d is C″(0,0)
It is. Further, although all the embodiments described so far have been explained only in the case of negative horizontal and vertical deviations, the case of positive deviations can also be considered basically in the same way. In FIG. 5, the necessary taps are shown for the case where there is a positive shift both horizontally and vertically. Taps required for combinations such as horizontal positive and vertical negative are omitted because they overlap with taps required for other purposes.

In FIG. 5, shift registers for three blocks from the e input, four blocks from the f input, and three blocks from the g input are required. However, this can be devised to reduce the number of shift registers.

For example, in Fig. 5, even in the case of a positive deviation, the same place as in the case of a negative deviation, that is, the position C'' (0, 0), is used as the reference, but this reference is moved to the position d# in Fig. 5. As shown in FIG. 6, the taps required in the case of a positive deviation can also serve as taps required in the case of a negative deviation.

If the tap is taken out in the case of positive deviation as shown in Figure 6,
Considering this in conjunction with Fig. 5, at least a shift register derived from g is unnecessary, so - It may look nice, but actually changing the reference from the position of d to the position of d'' means R
The side must be changed depending on whether there is a negative shift or a positive shift, and in the case of a positive shift, a tap with a delay amount of this side must be created from R by the delay difference between d and d'. This is not much of an advantage compared to the fact that a shift register for g is no longer required.

For example, in Fig. 6, even when there is a positive deviation both horizontally and vertically, the same place as when there is a negative deviation both horizontally and vertically, that is, C'' (
0, 0), but if we move this reference to the position d'' in Figure 7, the taps required in the case of positive displacement both horizontally and vertically are as shown in Figure 7. It can serve as all the taps needed when there is a negative vertical deviation.

The position of d'' in Fig. 7 is the maximum deviation from the position of d, which is 1=
If the position is shifted by s=4, j=T=3, and this is taken as a reference, the same combinations as in the case of negative shift both horizontally and vertically can be used for all tap selection procedures.

The same thing can be said in the case of horizontally positive and vertically negative deviations, or horizontally negative and vertically positive deviations. If the tap extraction is shown in Figure 7 in the case of a positive shift both horizontally and vertically, the shift register derived from g becomes unnecessary when considered in conjunction with Figure 5, and the shift register from f only needs to be for two blocks. . However, if there is a positive shift both horizontally and vertically, the reference changes from d to d#, so if there is a negative shift both horizontally and vertically on the R side, the position will be moved closer to R by the delay difference between d and d'. You must create a tap with a delay amount of and use it.

However, the number of delay circuits required for this purpose is small.

Note that since there is no connection for the first block of shift registers from e, this can be saved.

To summarize the above, this embodiment consists of frame memories 13 and 14 and a delay section as shown in FIG. 8, and a block comparison section as shown in FIG. 9.

As a result, as shown in FIG. 8, R can be divided into four types, R+ soil, R-+, R+-1R--1, depending on the positive and negative of the horizontal and vertical deviations.

The reference point for each of them is d in Figure 8.
++, d-1, and d1-1d--, so the delay difference between the two is compensated. That is, R+1 and R-1 and R0- and R-1 are exactly 5 (=4 pixels), and R+1 and R-1 are TP+S.
(-3X5+4=19 pixels). Output e in Figure 1
, f, and g correspond to the shift register connected to e.
' and f' become a shift register.

In e', as shown in FIG. 7, there is no connected tap for the first block, so this is omitted and a delay is imposed on the frame memory accordingly.

Also, in the next block, the first two lines of 10 pixels are still not used. This is because the 10 pixels of the first two lines of the first block are not used for the shift register column connected to f', so the delay is transferred to the frame memory all at once. Also, since there is no need for something corresponding to g, it is not shown in FIG. The squares indicated by dotted lines in FIG. 8 indicate the savings made in unnecessary shift register parts.

The delay difference between e' and f' is originally 9942 minutes, which is Q x H pixels, but since the first block of the shift register from e' is omitted, QXH - PXQ = 5 H - 2
There is a delay of 5 pixels.

The frame memory must be set so that there is a difference of one frame (F) between R-- and d-1, and F-2 (PQ) +
(TP+S)-(QH-PQ)+10=F-5H+4
becomes.

Block comparison uses a circuit like the one shown in FIG.

That is, a subtraction circuit 9 that subtracts the data selected by the selector 1 from R, an absolute value conversion circuit 10, an addition circuit 11,
An accumulator configuration consisting of register 12 is used.

For this R, when the amount of deviation is 1 pixel in the horizontal direction and 1 pixel in the vertical direction, when i is positive and j is positive, R + ten, and when i is negative and j is positive, R - ten When i is positive and j is negative, R-1 is used, and when i is negative and j is negative, R-1 is used.

and C' (i, j), C'' (i, j), C'''
(i, j), C'''' (i, j) are guided to selector 1, selected using the method shown in the explanation of Figure 4, and the difference between frames is taken and the absolute value is taken. , (or square, etc.) and accumulate for one block.

The above embodiments have been described for the case where S≦P and T≦Q, and cases other than these conditions are more complicated, but can be realized by simply extending the same idea.

The block comparison calculation circuit as shown in FIG.
, j), but as shown above, this may be thinned out. Furthermore, if, for example, n-times arithmetic operations can be made faster between the data rate and the arithmetic speed of this arithmetic unit, then n arithmetic operations for (i, j) can be time-division multiplexed with one arithmetic circuit. If this is done, the number of arithmetic units will be 1/n.

The explanation here was about non-interlaced images, but it can also be applied to interlaced images. However, when calculating the motion vector between interlaced fields, all j of the deviations (i, j) are 0.
It will have a 5 pixel offset.

Next, an embodiment will be described in which the determination procedure when detecting a motion vector can be simplified based on the result of block matching obtained by the above-described embodiment.

The procedure of this embodiment is as shown in FIG.

Unlike the conventional Fig. 41, the distribution of the degree of matching, that is, various minimums are not investigated.

After block matching step 21, only the vector quantities (x, y) with the minimum degree of matching are first obtained. (Step 22) Correlation between surrounding vectors is then detected (Step 23). This is the minimum matching vector (xi) of the eight blocks surrounding the block of interest.

yi) (i is the number assigned to the block), for example, performs Fourier transform. Let X be the left and right direction, and y be the up and down direction. That is, for example, if we are looking at block D-b in Figure 39, the eight blocks around it have the first
As shown in Figure 1, number the blocks in order as you go around them (you can start from any block in either direction) and let that number be i. Then, in the case of this FIG. 11, a waveform as shown in FIG. 12, for example, is obtained. In FIG. 12, the horizontal axis is the spatial position (SP), specifically the number i. The vertical axis is the X component of the motion vector. 12th
Figure A shows the X component of the motion vector, and Figure 12B shows the X component of the motion vector. In Figure 39, if the car is moving to the right (
Assuming that X is positive when moving to the right and y is positive when moving upward), the waveform will be as shown in Figure 12.

In the case of Fig. 11, for i=1 and i=3, c
The blocks -b and Dc are contour parts, and it is unclear whether the motion vectors are from the background or the car, but let us assume that they are as shown in FIG. 12, for example. Then, the waveform in Figure 12 is frequency analyzed by Fourier transform,
After conversion, a result like that shown in FIG. 13A should be obtained. FIG. 13A is a power spectrum after Fourier transformation. In FIG. 13A, there is one wave as the fundamental wave, so in FIG. 13A, the peak appears at the spatial frequency (sf) of 1 on the horizontal axis.

Figure 13B shows a case where the minimum matching degree vectors of the blocks around the block of interest are all close to the same vector in a large moving object, such as a background part, for example, where 5f=o
, that is, a peak appears in the DC component.

After Fourier transformation, we search for frequency components above a certain level in the data shown in Figure 13 A and B, and classify them according to their spatial frequencies (step 2).
4).

When there is a peak in the DC component as shown in Figure 13B, the motion vector is almost constant around the block of interest, so it is thought that the block of interest is also the same as the surroundings, and the motion vector is the minimum matching vector of the block of interest. It seems good (step 25).

Note that if this block of interest is an isolated moving object, the vector may be different from the surroundings.

However, it is not easy to save an isolated moving object, so whether the minimum matching vector is due to the moving object or the same as its surroundings can only be determined by the matching degree, so I don't think there is any particular problem.

As shown in Figure 13A, when the peak appears at 5f=1, the motion vectors of the blocks surrounding the block of interest change by one wave in one round, which means that the block of interest is located at the contour of a moving object. means that it exists in In this case, there should be two motion vectors within this block, so
Only at this time, the block is adaptively further divided into sub-blocks, for example, into four, and motion vector detection is performed for the sub-blocks (steps 26 and 27).

However, at this time, there is no need to perform block matching again. This is because it is clear that the candidate vector is near a group of two types of vectors existing in the surrounding blocks, and block matching is performed within a fairly limited range, simplifying the operation.

After steps 25 and 27 motion compensation is done (
Step 28). However, when the Fourier transform output does not become as shown in FIG. 13, it is impossible to detect a motion vector and motion compensation is abandoned, as in the conventional case.

FIG. 14 explains adaptive blocking.
In the case where a moving object such as a car is running as shown in Figure A, when a peak is detected at 5f = 1 to 3 in the judgment at step 24, as shown in Figure 14B, the peak is detected in the vicinity of that block. Perform subblocking locally. As shown in FIG. 14C, motion vector detection is performed recursively in units of this fine subblock.

An example of the determination section of the present invention described above is shown in FIG. The residual comparison circuit 31 detects (1) the position of the deviation with the minimum residual per block, and outputs this as a temporary motion vector. This temporary motion vector is
2. This consists of a register that stores one of the temporary motion vectors denoted by P, and a shift register that stores temporary motion vectors that are smaller in number by three blocks of pixels arranged in the horizontal direction and denoted by Q. An orthogonal transform circuit 33 such as Fourier transform and a peak detection circuit 34 to which the output thereof is supplied are provided, and when there is a peak at 5f=1.2.3 at the output terminal 3 of the OR gate 35, A determination signal is extracted, and this determination signal is supplied to the sub-blocking circuit, and a motion vector is output to the output of the AND gate 36.

For the parts explained using Fourier transform in the above explanation, you can also use Hadamard transform, Hartley transform, etc., which are not limited to Fourier transform. In short, it is sufficient to calculate the correlation with reference waves of various wave numbers and find out which wave number has the strongest correlation.

To add more information about the case of Fourier transform, it may seem difficult to calculate because there are many -multiplications, etc., but this transformation is always an 8th-order transformation, and using the FFT method, in the case of 8th-order, multiplication is often done by only addition. It is well known that there are very few.

Therefore, there is no problem. Hadamard transform etc. have coefficients of 1 or 0 and are easier to calculate.

The procedure shown in FIG. 10 involves simple operations compared to conventional processing, and does not require complicated judgments.

Finally, further embodiments of the invention combining motion hector compensation and frame differential coding will be described below. FIG. 16 shows an example of this.

In FIG. 16, 41 is a delay unit for block matching as described in the first embodiment, 42 is a subtraction circuit group, and 43, indicated by PI, is a power integration circuit group, which is used for squaring. It has a configuration in which a multiplication circuit and its multiplication output are accumulated. The output of this power integration circuit group 43 is determined by the determination unit 44.
supplied to

For the above-mentioned block comparison, the subtraction circuit group 42 and the power integration circuit group 43 calculate (i, j) (i=-8 to +S, j
= -T to 10T), a total of (2S, 1,001) (2T+1) are provided. The determination unit 44 selects the frame difference with the amount of shift that has the lowest difference power, and outputs the frame difference to the output terminal 45.
The detected motion vector is obtained.

A selector 46 connected to the subtraction circuit group 42 is provided. The output of this selector 46 is a DCT conversion circuit 47
supplied to The output d of the conversion circuit 47 is taken out as an encoded output via the code assignment circuit 51, and is also supplied to the addition circuit 49 via the inverse conversion circuit 48. The output signal C of the adder circuit 49 is supplied to the frame memory 50,
The output of the frame memory 50 is supplied to a delay diagram 41 and an adder circuit 49.

The difference from the conventional configuration is that the motion vector detection configuration is moved within the loop of the encoding circuit.

In addition, when using the method of reducing the number of motion vector detection points, and as shown in Fig. 16, (2S+1) (2T+1")
Even if multiple calculation circuits are prepared, if the motion vector is to be obtained as a real number instead of an integer unit of a pixel grid, the selector 46 is not only a simple selector, but also a variable delay circuit for the selector 46. may be connected to control the delay amount of this variable delay circuit by the amount of the detection point. Further, a delay circuit is connected to the selector 46,
An interpolation calculation circuit, such as a linear interpolation circuit, may be connected to this delay circuit to obtain a frame difference at a desired shifted point. It is also possible to combine both.

As mentioned above, the band compression device to which the present invention is applied,
Only one frame memory is required and the device is small. Also, what about the frame memory address system? 1ll (delay amount control) is not required. The only increase in hardware compared to the past is the selector, which is not a huge increase.

Although only the DCT conversion circuit and the inverse conversion circuit were inserted in the loop of Fig. 16, various processing may be inserted here, but since it is independent from this invention, it will not be discussed here. However, anything is possible. In addition, the block comparison circuit (determining the degree of match) has power (
(squared), but it can also be an absolute value or other nonlinear operation. Furthermore, the conventionally known method of reducing the number of motion vector detection points can be similarly applied to the present invention.

Another embodiment of a band compression device using a frame differential encoding method that combines motion compensation will be described. This embodiment is as shown in FIG. 17, and basically has a configuration in which the conversion circuit 47 of the embodiment shown in FIG. 61). Therefore, the details of FIG. 17 are almost the same as those of FIG. 16, and only the different points will be supplementarily explained here.

The data amount detection circuit shown as II in FIG. 17 may be similar to the power integration circuit PI in FIG.
In order to perform better band compression, it is necessary to detect the amount of band-compressed data after DCT. Examples include the integral of absolute values, the number of non-zero data, and the number of "1"s in each bit of binary representation. The output of this data amount detection circuit group 62 is supplied to the determination section 44.

In addition, the conversion circuit shown here as DCT is at least DCT
, but may include band compression techniques such as zigzag transforms used in conjunction with it. Of course, in that case, it is necessary to perform the reverse procedure using an inverse conversion circuit.

FIG. 17 allows motion vector compensation more suitable for band compression than in FIG. 16.

Figure 17 also shows that since DCT is a linear operation, that is, an inner product operation (matrix operation), the same result is obtained whether after or before taking the frame difference, so it is okay to think like the 18th example. It's obvious. FIG. 18 shows the conversion circuit group 61 in FIG. 17 moved further forward.

When the conversion circuit group 61 is moved into the motion vector detection circuit as shown in FIGS. 17 and 18, the problem is that the amount of DCT calculation increases. However, it can be considered as follows.

First, as a premise, the image is divided into blocks of horizontal P pixels×vertical Q pixels, as shown in FIG. 19, which overlaps with what has already been explained. Here, an example where P=5 and Q=5 is shown. It is also assumed that the motion vector is within a range of ±S pixels horizontally and ±T pixels vertically. Hereinafter, S=4. Let T=3.

Therefore, when a certain PXQ pixel block in the current frame is shown as the inner thick line in FIG. 20, the outer thick line in FIG. 20 is the range of pixels in the previous frame that is used for block comparison. In Fig. 20, the center is at the coordinate 0. O(x=o
, y-0) for each pixel. however,
It is omitted and only some pixels are written.

(1) (Vertical) For example, first, a block comparison corresponding to a case where one frame is shifted horizontally by 4 pixels from left to right and vertically by 3 pixels from top to bottom is to compare a certain block in the current frame to the 20th If the thick line inside the figure is used, the block of P pixels×Q pixels at the position of the thick line α in FIG. 21 will be compared in the previous frame. That is, blocks with coordinates as shown in FIG. 22A are used.

Next, to compare blocks corresponding to a shift of 4 pixels horizontally from left to right and 2 pixels vertically from top to bottom in one frame, compare the blocks at the bold line β in Figure 21. set to target. That is, a block of coordinates as shown in FIG. 22B is used.

Hereinafter, blocks that have the same horizontal shift but differ only in vertical shift will be compared in the same way.

If we compare the DCT of α and the DCT of β, we will find the following. DCT is performed in two dimensions, but if we first transform horizontally and then vertically, each group of pixels in 5 lines in Figure 22, where the horizontal transformation of α is calculated, is , β are calculated in the horizontal direction.

As shown by 68 in each group of pixels of 5 lines in FIG. 22B, 4 of the lines actually overlap, and it can be seen that if calculation has been performed for α, no calculation is necessary for β.

After all, when comparing blocks that have the same horizontal shift but differ only in vertical shift, a considerable number of calculations can be omitted during horizontal conversion.

(2) (Horizontal) Next, the block comparison corresponding to the case where the block is shifted horizontally by 3 pixels from left to right and vertically by 3 pixels from top to bottom is to compare a block in the current frame to the inside of Figure 20. If the bold line is used, then for the previous frame, a block shifted by one position to the right from α in FIG. 21 will be compared, that is, a block with coordinates as shown in FIG. 22 C will be used. When comparing the data block of FIG. 22C with the data block of FIG. 22A, the leftmost data in FIG. 22A is not in FIG. Not in Figure A.

If we transform the two-dimensional DCT in the horizontal direction first,
Regarding the Fourier transform, for example, a method such as ("Fast solution method for overlapping data" in Sanpo Publishing "'FFT usage type" by Takeshi Yasuoin et al., pages 132-133) is known, so α as shown in Figure 22A is used. If a conversion operation has been performed on the data in the block, move it to the right by 1, as shown at 69.
Conversion calculations for blocks with shifted pixels can be determined by simple calculations without the need for new calculations. This reference material talks about Fourier transform, but the same applies to cosine transform.

Although the above explanation was only about the initial horizontal one-dimensional DCT transformation of the two-dimensional DCT, the same idea can be extended to two dimensions to significantly reduce the amount of calculation. Although Figure 17 requires a significant increase in calculation hardware compared to Figure 16,
When FIG. 17 is changed to FIG. 18, the amount of calculation is significantly reduced, and although it is more than that in FIG. 16, it is possible to take advantage of the advantages of including DCT and the like in motion vector control.

Here, we will compare the amount of calculation.

In FIG. 16, if PI is a square integral, its configuration becomes the circuit shown in FIG. 23. That is, in FIG.
The output of the subtraction circuit 71 is squared by a multiplier 72, and the squared output is accumulated by an addition circuit 73 and a memory 74.

Moreover, in the case of absolute value integration, it is shown in FIG. That is, in FIG. 24, the output of the subtraction circuit indicated by 71 is squared by the multiplier 72, and the squared output is sent to the sign inverting circuit 75 and the selector 7.
6. The output of selector 76 is accumulated.

The integral for one block of PXQ pixels in the current frame is P
Accumulate XQ operations, which is (2S+1) x (2T+1)
pcs are required. Total P-Q・(2S+1)X (
2T+1) operations, per 1m element (2S+1)
This means (2T+1) times.

DCT can be calculated using a product-sum calculation circuit, and the PX of the current frame
For one block of Q pixels, it is necessary to perform the calculation PXQXpZQ times, which is P times per pixel.

In Figure 18 The part now indicated as DCT is really only DCT.

Further, it is assumed that the data amount detection circuit II is also the same as the power integration circuit PI. At that time, the calculation amount of II is the same as that of PI in FIG. 16. This is because the number of data does not change before and after DCT.

Regarding the DCT operation, if you think about it simply, there are (2S+1)(2T+1)+1 DCTs in Figure 18, so there are more DCTs than in Figure 16, and each pixel has ((2S+1)
(2T+1)+1) xP calculations, which is huge. However, for the reason mentioned above, it is not necessary to calculate all of this, but at least one DC on the a side and one on the b side.
A T calculation circuit is necessary, and it is definitely more than twice as large as shown in FIG. 16, but it is not unrealistically large.

What is also noteworthy is that the data is subjected to frequency analysis in the DCT output, as if it had been passed through a filter bank, so it is possible to think of things that were previously unthinkable.

First, for example, in band compression, a certain frequency component of the DCT output may be excluded, but if this is the case, it is not necessary to calculate II, that is, PI, for that frequency component, and the number of calculations is reduced here.

Next, for example, since each frequency component of the input data is known by the DCT input at input a, we can judge this DCT output and find the parts that do not need to be calculated in II, and perform the calculation. Can be omitted.

Alternatively, the calculation of II can be limited to limited frequency components in the low range, assuming that high frequency components within a block do not have much influence on motion vector detection.

Regarding motion vector detection, please refer to the reference cited earlier (Nikkan Kogyo Shimbun Publishing Co., Ltd., "TV Duo-Image Multidimensional Signal Processing", written by Fukueki, P2O3-204) to reduce the amount of calculation required for motion vector detection.
There are methods such as the "three-step search method"("three-step search method"), but such techniques can also be applied to this case.

FIG. 25 shows an example of this. In Figure 25, the shifted 9
This shows an example in which the amount of deviation is localized by repeating the investigation three times for each position.

FIG. 25 shows only the parts that should be replaced from inputs a and b to d in FIG. 18. In FIG. 25, the delay section 41A is the same as the one explained so far, but it has only nine taps, and the delay section 41B is basically the same, but
The amount of delay for each tap is determined based on the roughest area minimum result obtained by the determination unit 41A. Similarly, the delay unit 41C provides the amount of delay for each tap in the area limited by the determination unit 44A and determination unit 44B. In addition, 65A, 65B, 66A shown in EDI~4,
66B is an equivalent delay circuit that guarantees the delay of the arithmetic unit.

Although not shown in FIG. 25, the previously described method of reducing DCT operations can be applied here as well.

[Effects of the Invention] The present invention can detect motion vectors with simpler judgment and simpler operations than in the past. In particular, according to the present invention, motion compensation can be performed even on the outline of a moving object, and this block can be prevented from becoming blurred.

7/ 7/' / / /″

[Effect]

The present invention can detect motion vectors with easier judgment and simpler operations than in the past. ”

Claims (1)

[Claims] A motion vector detection device that detects a motion vector by block matching that detects the degree of matching of corresponding positions between image signals that differ by one frame in time, the motion vector being detected by the block matching. The vector with the minimum matching degree is detected, the vectors spatially surrounding the detected minimum vector are arranged in order, the correlation calculation with the reference wave is performed, and an output according to the correlation is generated. A motion vector detection device characterized in that the minimum vector is detected as a motion vector from the output of the calculation, and/or it is determined that a moving object contour exists within a block.