KR100205146B1 - Motion estimation method in digital video encoder - Google Patents

Abstract

The present invention provides a method for detecting a motion vector of a reference region for a search region of a previous frame using an array processor, comprising: a first step of inputting a pixel value of a predetermined search region for a previous frame to the array processor; Inputting a pixel value for the reference region into an array processor; Detecting a motion vector of the reference region with respect to the search region; And a fourth step of setting a new search area in a previous frame, inputting only pixel values not overlapping with the search area stored in the array processor to the new array processor, and returning to the third step. That is, the present invention has an effect that the motion vector can be detected in a short time by sequentially inputting previous frame data not overlapped with the reference block when the motion vector is detected using the processor array.

Description

Motion Prediction Method in Digital Video Encoder System

The present invention relates to a video signal compression technique in a digital video encoder system, and more particularly, to a motion prediction method in a digital video encoder system for predicting and compressing a motion of a video signal. The core technology among digital video encoder system development technology is the compression technology of video signal. The compression of the video signal is obtained by removing the redundancy, spatial redundancy, temporal redundancy, and statistical redundancy among the color signals inherent in the video signal. Among them, the most compression is achieved by eliminating temporal redundancy. For this purpose, motion estimation and compensation techniques are used. The BMA (Blocking Matching Algorithm), which is the most commonly used method of motion vectors, and is adopted by three of the four proposed all-digital methods, assumes that the motion of the screen is moved horizontally and vertically. It is a method of estimating which position of a block image of a frame most coincides with a block image of a frame before motion occurs and estimating the normal motion vector of the position. In general, the block size is usually 8 × 8 or 16 × 16. This means that the larger the block size, the lower the bit rate of the motion vector to be transmitted. The size of the block. The range of motion is theoretically from -∞ to + ∞, but in practice, the search interval is determined in consideration of computation time or hardware complexity. That is, redundancy is eliminated when motion compensation is performed on the blocks of the previous frame obtained by the motion estimation as shown in FIG. Here, FIG. 1a shows the blocks of the previous frame used to predict the next frame, and FIG. 1b shows the blocks of the previous prem that have been readjusted to the current block position for motion compensation. Mathematically, the when to find the best matching block, in the image of the two frames in a row, the previous frame f d _{1 (x, y), f} 2 (x, y) of the current frame, f ₂ (x , y) and f ₁ (x-α, y-β) are converted to f ₁ (α, β), and the difference between f ₁ (x-α, y-β) and f ₂ (x, y) is obtained. We use the method of predicting the minimum value of the difference as a motion vector. The formula for calculating the difference between frames mainly uses the mean absolute error (E _abs ).

[Equation]

Where | B | stands for block size.

The wider the search area for motion estimation, the better the performance. The higher the resolution, such as high-definition television, the wider the search area is required. 31 area of search is required.

Conventionally, the use of a two-level search method array processor has been introduced to realize such motion search at high speed.

This will be described in detail below.

First, two-level search is to find a motion vector roughly for a frame of low resolution, and then search for the high resolution frame in detail based on the vector found at low resolution.

Low-resolution images are called top-level images, and high-resolution images are called bottom levels. For example, assuming that the resolution of a bottom level image is 2048 × 1024, a resolution of a low resolution image is 256 × 128 when a high resolution image is reduced to 1/8.

If the search area is -128 to 127 pixels horizontally, and -32 to 31 pixels vertically, only 16-15 pixels horizontally and -4 to 3 pixels vertically are searched at the top level. That is, the search area is reduced to 1 / (8 × 8) than when searching directly in the high-resolution area, and at the bottom level, only horizontal -4 to 3 and vertical -4 to 3 pixels are searched based on the motion vector obtained from the top level. Or you can search a slightly wider area, usually like -16 to 15 pixels horizontally and -4 to 3 pixels vertically, just like the pixels you searched at the top level.

In the case of using the processor array, as shown in Fig. 2, finding where the corresponding block of the previous frame is located with respect to the current block (1) is to find the motion. As the size of the motion vector of the block increases as the resolution increases, the amount of computation increases. Block matching processor arrays are used to handle this increased computation.

As shown in FIG. 2, u × v processor arrays PE11 to PEuv for block matching are gathered, and each processor array PE11 to PEuv is moved from the reference blocks REF1, REF2, ... to block matching. The search area data (FD1, FD2, ...) of the frame is taken as an input, and a mean square error (MAS) or mean absolute error (MAE) is calculated for one motion vector.

Since the number of processor arrays is u × v, we can predict u movement horizontally and v movement vertically.

After the predetermined delay time d, the MAE or MSE obtained from the processor arrays PE11 to PEuv is compared at the output controller 10, and the output controller 10 is the MAE or MSE detected from each processor array PE11 to PEuv. The smallest value will output the smallest MAE / MSE 1, MAE / MSE 2, ... and the corresponding motion vectors (VEC1, VEC2, ...).

An example of this process is as follows.

First, as shown in FIG. 3, it is assumed that the motion search area FD is 6 × 6 pixels, and the reference block REF for the motion search is configured by 2 × 2 pixels. At this time, the array processors PE11 to PEuv may be configured with a number of 5 × 5 equal to the number of motion vectors. That is, since there are components of -2, -1,0, + 1, + 2 in the horizontal and vertical directions as shown in FIG. 4, 25 motion vectors may exist, and the 25 motion vectors may be detected. For this purpose, 25 array processors PE11 to PEuv should be used as shown in FIG. Each array processor PE11 to PE55 performs matching between the search area FD and the reference block REF with respect to the corresponding motion vector. Matching may employ a method by MAE or MSE as described above. That is, the MAE between the pixel values P29, P30, P35, and P36 in the search area FD of FIG. 4 and the pixel values R1, R2, R3, and R4 of the reference block REF may be detected by Equation 1 below. Can be.

[Equation 1]

MAE = (| P29-R1 | + | P30-R2 | + | P35-R3 | + | P36-R4 |)

On the contrary, MSE can be detected by Equation 2.

[Equation 2]

MSE = (P29-R1) ² + (P30-R2) ² + (P35-R3) ² + (P36-R4) ² ))

Accordingly, the array processor PE55 may calculate the MAE or MSE of the pixel values P1, P2, P7, and P8 of the search area FD and the pixel values R1, P2, R3, and R4 of the reference block REF. The calculation is performed according to 1, 2, and the processors PE11 to PE54 other than the array processor PE55 also obtain MAE or MSE for each search area. As such, the position of the processor of the array processor (one of PE11 to PE55) having the smallest MAE or MSE among the MAEs or MSEs detected by the array processors PE11 to PE55 becomes a motion vector. That is, when the array processor PE11 obtains the smallest MAE or MSE, the motion vector becomes (+ 2, + 2), and when the array processor PE55 obtains the smallest MAE or MSE, the motion vector becomes ( -2, -2).

6 shows a block diagram of an array processor for obtaining a MAE or MSE between such search region FD and reference region REF.

As shown in the drawing, the pixel values of the search area are applied and stored in the search memory 101 through the input terminal A. FIG. At this time, the pixel values stored in the search memory 101 are as follows. That is, pixel values of two rows of the pixel values of the search area FD are input to each of the search memories 101 of the array processors PE55, PE54, PE53, PE52, and PE51. That is, the pixel values P1, P2, P7, and P8 are arranged in the array processor PE55, the pixel values P2, P3, P8, and P9 are arranged in the array processor PE54, and the pixel values P3, P3, and the like are arranged in the array processor PE53. P4, P9, P10, pixel values (P4, P5, P10, P11) are input to the array processor PE52, and pixel values P5, P6, P11, P12 are input to the array processor PE51. will be.

In this manner, the pixel values of the search region FD corresponding to two rows and three rows are applied to the array processor PE41-45, and the array corresponding to three rows and four rows to the array processor PE31-35, respectively. Pixels of the search area FD corresponding to four and five rows are stored in the processor PE21-25, and five and six rows are respectively stored in the array processor PE11-15.

At this time, the pixel values of the reference block REF are input through the terminal B at the same time when the pixel values of the fifth and sixth rows of the search area FD are input to the array processors PE11-PE15. When the input of the pixel values of the search area FD is finished, the operator 102 detects and outputs the MAE or MSE of the pixel values of the search memory 101 and the pixel values of the reference block REF area. will be.

In this way, the MAE / MSE detected by each array processor P11-P55 is applied to the output controller 10, and the output controller 10 outputs the smallest value among the MAE / MSE array processors PE41-45. And the motion vector assigned to one of the two components is output as the motion vector of the reference block REF for the search area FD.

Here, the search areas FD1, FD2, ... of the previous frame data are sequentially input to the processor arrays PE11 to PEuv to match the reference blocks REF1, REF2, .... The search areas FD1, FD2, ... are sequentially input to the processor arrays PE11 through PEuv, while the reference blocks REF1, REF2, ... are simultaneously connected in parallel to the respective processor arrays PE11 through PEuv. The MSE or MAE is calculated by comparing with the reference blocks (REF1, REF2, ...) in the area corresponding to the input vector.

That is, as shown in FIG. 7, the search area FD1 of the previous frame data is input to the processor arrays PE11 to PEuv, and the reference block REF1 is input to predict the movement, and the reference block REF1 is input. After that, the motion vector VEC1 and MAE / MSE1 can be obtained after a predetermined delay time d.

Subsequently, a search region FD2 and a reference block REF2 are inputted to predict motion of another second reference block REF1 in another region, and a vector VEC2 and MAE / MSE2 after a predetermined time d are inputted. You can get it.

However, in the conventional method of detecting a motion vector in this manner, since the search areas of the previous frame data must be input to the processor arrays PE11 to PEuv respectively corresponding to the reference blocks as described above, there are many methods for detecting the motion vector. There was a problem that it took time.

SUMMARY OF THE INVENTION The present invention has been made to solve this problem, and an object of the present invention is to detect a motion vector in a short time by sequentially inputting previous frame data into a processor array that does not overlap the reference block when the motion vector is detected using the processor array. The present invention provides a motion prediction method in a digital video encoder system. A motion prediction method in a digital video encoder system according to the present invention is a method of detecting a motion vector of a reference region with respect to a search region of a previous frame by using an array processor, wherein the pixel values of a predetermined search region for a previous frame are obtained. A first step of inputting to a processor; Inputting a pixel value for the reference region into an array processor; Detecting a motion vector of the reference region with respect to the search region; And a fourth step of setting a new search area in a previous frame, inputting only pixel values not overlapping with the search area stored in the array processor to the new array processor, and returning to the third step.

1 is a diagram illustrating a motion compensation method using a block matching method.

2 is a block diagram of a conventional array processor for detecting a motion vector.

3 is a diagram illustrating an example of a search area and a reference block.

4 is a diagram illustrating a state of obtaining a motion vector of a reference block with respect to a search area.

5 is a block diagram showing a state in which data of a search area and a reference block are input into a conventional array processor.

6 is a block diagram of an array processor.

7 is an input / output diagram of an array processor for detecting a conventional motion vector.

8 is an input / output diagram of a motion processor according to the present invention.

9 is a block diagram showing a state in which data of a search area and a reference block are input into an array processor according to the present invention.

10 illustrates a search region applied to an array processor in accordance with the present invention.

11 is a diagram illustrating a relationship between a reference block and a search region according to the present invention.

12 illustrates a state in which data in a search region is continuously input to an array processor according to the present invention.

* Explanation of symbols for main parts of the drawings

10: output controller 101: search memory

102: operator

Hereinafter, an embodiment of the present invention will be described in detail with the accompanying drawings.

The inventors have found an efficient way to find a motion vector without having to repeat it as shown in FIG. 7 when it is an adjacent reference block. That is, when the search areas overlap each other, when the search area for the predetermined first reference block and the search area for the second reference block overlap, only the data of the non-overlapping area is input and the motion for the second reference block is duplicated. You can get a vector. This is shown in FIG.

In FIG. 8, the processor arrays PE11 to PEuv respectively input a search region FD1 and a reference block REF1 of previous frame data, and calculate MSE or MAE by comparing with the reference block REF1 in the region corresponding to the vector. Do it.

After a predetermined time d, the output controller 10 outputs the corresponding motion vector VEC1 having the smallest value of the smallest value of the MAE or MSE detected from each of the processor arrays PE11 to PEuv.

Here, when the pixels of the search area FD1 and the next input search area FD2 are mostly the same and partly different from each other, if only the different pixels are replaced in the processor arrays PE11 to PEuv, the processor arrays PE11 to PEuv are replaced. It becomes the same as that which inputs frame data FD2. Therefore, the processor arrays PE11 to PEuv input only the data D2 of the new search area FD2 after the processing of the search area FD1 and delete the data of the search area FD1 excluded from the search area.

Similarly, the processor arrays PE11 to PEuv can be repeated for the reference blocks 3, 4, and 5, ..., which are the previous frame data (FD2, FD3, ...) in the structure of the processor arrays PE11 to PEuv. This is because the data (D2, D3. ,,,) inputted as) is sequentially inputted as data of a new frame is inputted.

In order to implement this, the inventor further configured one row (PE01-PE05) of the array processor as shown in FIG.

When the pixel values of the search area FD shown in FIG. 10 are initially input to the array processors PE00-PE55, the first, second, third, fourth, fifth and sixth rows of the search area FD1 are shown as shown. Input a pixel value (i.e., 1, 2, 3, 4, 5, 6 rows of pixels form an initial search region) and detect the motion vector of the reference block REF1 for this search region FD1. Done.

When the detection of the motion vector of the reference block REF1 for the first, second, third, fourth, fifth, and sixth rows by the initial input ends, the new second search area FD2 for the second reference block REF2 is completed. ), And it can be seen that rows 3, 4, 5, 6, 7, and 8 are new search areas (FD2). At this time, the pixel data of 3, 4, 5 rows overlaps the initial search region FD1, and thus, the pixel values in the array processors PE31-PE35, PE21-PE25, PE11-PE15 do not need to be changed. Able to know. That is, when a new pixel value for the second search area FD2 is newly input to the array processors PE01-PE05, PE51-PE55, PE41-PE55, the motion in the second search area FD2 with respect to the second reference block REF2 is newly input. You can get a vector. In the same way, the motion vector of the third, fourth, search regions FD3, FD4, ... of the third, fourth, reference blocks REF3, REF4, ... can be obtained. . 9 sequentially shows rows of pixels applied to the array processors PE01-PE55 to newly set the search area to detect the motion vector.

11 illustrates a relationship between a search area and a reference block. In FIG. 11, a shows a search area S for one of the reference blocks REF shown in FIG. 5B in the screen P. FIG. This search area S is a block that is vertical M × m and horizontal N × n. Each processor array PE11 to PEuv stores a search area S for the corresponding vector and then matches the reference block REF to obtain a MAE / MSE. At this time, the area to be compared is the same as the part indicated by hatched in Fig. 11C.

FIG. 11D shows, as an example, the order of data to be input when the search area S is input to the processor arrays PE11 to PEuv.

The processor arrays PE11 to PEuv cannot receive data of all the search areas S, and cannot obtain MAE / MSE for all vectors at once.

Because the search area S is large, since the processor arrays PE11 to PEuv must exist as many as the number of motion vectors, the search area is not easily wide. Therefore, the search must be divided into several times, and the whole area is divided by α to show an example. Matching is performed for the first (M × m) × (N × m /) and for the second (M × m) × (N × n / α). This comparison is performed on the α search region parts to obtain a motion vector having the smallest MAE / MSE.

However, when the number of motion vectors is small because the search area is small, the motion vectors may be obtained at one time using the processor arrays PE11 to PEuv. In the case where the search is continuously performed with respect to the adjacent reference block, as shown in FIG. 12, as many blocks as the new search area (M × m) × n escapes and exit, new blocks as much as (M × m) × n ( Since n1) is input to the processor arrays PE11 to PEuv, the search can be continued for adjacent reference blocks.

That is, in the present invention, as described above, when the processor arrays PE11 to PEuv cannot calculate the MAE / MSE for the entire search area at once, the processor arrays PE11 to PEuv can be divided by the search areas that can be calculated by the processor arrays PE11 to PEuv.

When a motion vector is predicted for one slice as described above, the data amount of the search area input to the processor arrays PE11 to PEuv is as follows.

Search area input data amount for estimating a motion vector with respect to one reference block REF = (M × m) × (N × n)

Search area input data amount for K reference blocks REF = K × (M × m) × (N × n)

When the motion vector is not found at once for the entire search area S due to the limitation of the processor arrays PE11 to PEuv, and the motion is predicted for only some blocks, the adjacent search block is adjacent to the previous search area as shown in FIG. By continuously inputting M × m) × n blocks into the processor arrays PE11 to PEuv, the motion vector can be estimated continuously only for blocks of some consecutively searched adjacent search areas.

In the case where motion is estimated with respect to successive adjacent reference blocks using the adjacent block (M × m) × n blocks starting from block 1 in FIG. 11D, starting with block 2 again, motion estimation is performed with respect to the adjacent reference block. And repeat for the α-th block.

During iteration α times, α motion vectors and MAE / MSEs can be obtained for the corresponding search region. The reason for obtaining α is because the search area is divided into α.

The obtained motion vectors for the smallest MAE / MSEs and the motion vectors for the smallest MAE / MSEs are the final motion vectors of the corresponding reference block.

The calculation amount required here is as follows.

Data amount into which the search area FD1 is input to the processor array with respect to the reference block REF1 = (M x m) x (N x n / a)

Input data amount for subsequent sequential blocks (REF2, ...) = (M × m) × n

Amount of data input to the processor array for K reference blocks in one slice for sub-block searching = (M × m) × ((K−1) × n × N × a / a)

The amount of data input to the processor array to repeat α to obtain the most motion vector.

In the two motion vector methods, the amount of input data shown is the input data amount of the search area read in order to find the motion for each reference block for the entire search area and then continue to search for the K reference blocks.

First, the motion is estimated only for the sub-search area block, and the block of the neighboring area of (M × m) × n is continuously read into the processor array to estimate the motion for K adjacent reference blocks, and then In order to estimate the final motion vector by repeating the sequence, the data of the input search area = B,

to be.

Since K is usually 10 times larger than N, and α is an integer of 1 or more, it is a method of predicting movement as fast as N / α as a whole.

That is, the present invention has an effect that the motion vector can be detected in a short time by sequentially inputting the previous frame data not overlapped with the reference block when the motion vector is detected using the processor array, and the complexity of the hardware Find the optimal point between processing speeds. That is, if the size of the search area is large enough to be processed by the processor array PE11-PEuv, α is 1, and the gain of the speed is N times, and the size of the search area is the processor array PE11-PEuv. If not, the hardware is simplified by α times and the processing speed gain is obtained by N / a.

Claims (1)

A method of detecting a motion vector of a reference region for a search region of a previous frame by using an array processor, comprising: a first step of inputting a pixel value of a predetermined search region for a previous frame to the array processor; Inputting a pixel value for the reference area into the array processor; Detecting a motion vector of the reference region with respect to the search region; And a fourth step of setting a new search area in the previous frame, inputting only pixel values not overlapping with the search area stored in the array processor to the new array processor, and returning to the third step. Motion prediction method