一個利用動作補償來達成影片更新率提高的方法

國

立

交

通

大

學

資訊科學系

碩

士

論

文

一個利用動作補償來達成影片更新率提高

的新方法

A New Method for Frame Rate Up-Conversion Based

on Motion Compensation

研 究 生：楊雅茹

指導教授：陳玲慧 教授

一個利用動作補償來達成影片更新率提高的新方法

A New Method for Frame Rate Up-Conversion Based on Motion

Compensation

研 究 生：楊雅茹 Student：Ya-Ru Yang

指導教授：陳玲慧 Advisor：Ling-Hwei Chen

國 立 交 通 大 學

資 訊 科 學 研 究 所

碩 士 論 文

Submitted to Institute of Computer and Information Science College of Electrical Engineering and Computer Science

National Chiao Tung University in partial Fulfillment of the Requirements

for the Degree of Master

in

Computer and Information Science

June 2004

A New Method for Frame Rate Up-Conversion Based on Motion Compensation

Ya-Ru Yang and Ling-Hwei Chen

Department of Computer and Information Science, National Chiao Tung University

1001 Ta Hsueh Rd., Hsinchu, Taiwan 30050, R.O.C.

Abstract

Frame rate up-conversion (FRC) is a conversion between any two display

formats with different frame rates. For improving visual quality, the frame rate

required for high definition television (HDTV) is much higher than the current frame

rate. That is, the frame rate should be up-converted in HDTV.

In this thesis, we present a frame interpolation algorithm for FRC. In the

proposed scheme, forward motion estimation (ME) is performed to generate the

motion vectors (MVs) that determine how the frame will be interpolated. Based on the

naturally perceptual scene, we set a threshold for ME, and only refer to one frame

information when the interpolation frame is constructed in the process of motion

compensated (MC). To solve the problem of overlapped and hole regions introduced

in MC, a weighted average compensation technique is presented to process the

overlapped regions, and a block based algorithm is provided to process the hole

experimental results of the proposed method are compared with the results of the

hierarchical motion compensated frame interpolation (HMCFI) method. The PSNR of

the interpolated frames by the proposed method and by HMCFI are near enough,

一個利用動作補償來達成

影片更新率提高的新方法

研究生：楊雅茹 指導教授：陳玲慧 博士

國立交通大學資訊科學研究所

摘要

播放，更新率的提高(frame rate up-conversion)是必要的。在這篇論文裡，我

們將利用影片壓縮的技巧內插生成補圖，來達到更新的提高。為了容易研究與使

誌謝

TABLE OF CONTENTS

ABSTRACT……….I

ABSTRACT(IN CHINESE)………III

ACKNOWLEDGEMENT(IN CHINESE)……….………IV

TABLE OF CONTENTS……….………….V

LIST OF FIGURES……….…………..…………..VI

CHAPTER 1 INTRODUCTION………….……….1

CHAPTER 2 THE PROPOSED METHOD………9

2.1 Motion estimation and motion compensated interpolation………...7

2.2 Overlapped and hole regions processing………...8

2.2.1 Overlapped region processing……….9

2.2.2 Hole region processing………..10

2.2.2.1 The process of block compensated with partial hole region………...11

2.2.2.2 The processing of block with complete hole region………...12

2.2.3 Motion estimation with overlapping………..13

CHAPTER 3 EXPERIMENTAL RESULTS……….………15

3.1 Comparison with hierarchical MCFI………15

3.2 Effectiveness of overlapping block………..19

CHAPTER 4 CONCLUSIONS………23

LIST OF FIGURES

Fig. 1-1 Hole and overlapped regions in the interpolated frame……….2 Fig. 1-2 (a) Estimated MV, Vb, computed during ME. (b) Interpolation using blocks

from (k-1)th and kth frames………..……….3

Fig. 2-1 Block diagram of the proposed method...………..6 Fig. 2-2 Frame interpolation by the proposed method with Macroblock I inserted…....7 Fig. 2-3 (a) The previous and current frames. (b) The interpolated frame by the proposed

motion estimation and motion compensated interpolation with frames of (a)…………8

Fig. 2-4 Three types of blocks in the interpolation frame………...9 Fig. 2-5 The interpolated frame about Fig. 2-3(b) after processing of the overlapped

regions by the proposed method………..10

Fig. 2-6 (a) A block with partial hole region (black part). (b) Scan the block row by row,

and stop scanning when the first scanned pixel in a row belongs to the hole region. (c) Change the scanning direction and scan the block column by column. (d) The compensated block………11

Fig. 2-7 The order of the block scanning direction with bold arrow indicating the first

scanning row(column)………..12

Fig. 2-8 The interpolated frame about Fig. 2-5 after processing of the compensated block

with the partial hole regions by the proposed method………..12

Fig. 2-9 The neighboring pixels marked by white color around block………13 Fig. 2-10 The interpolated frame about Fig. 2-8 after processing of the complete hole

block regions by the proposed method……….13

Fig. 2-12 The interpolated frame about Fig. 2-3(a) after processing of the overlapped

regions with half block overlapped………....14

Fig. 3-1 “Calendar” sequences: (a) previous frame, (b) current frame, (c) frame

interpolated by the HMCFI, (d) frame interpolated by the proposed method, (e) an enlarged region of (c) by a factor of two, (f) an enlarged region of (d) by a factor of two………...16

Fig. 3-2 “Foreman” sequences: (a) previous frame, (b) current frame, (c) frame

interpolated by the HMCFI, (d) frame interpolated by the proposed method, (e) an enlarged region of (c) by a factor of two, (f) an enlarged region of (d) by a factor of two……….17

Fig. 3-3 “Table” sequences: (a) previous frame, (b) current frame, (c) frame interpolated

by the HMCFI, (d) frame interpolated by the proposed method without overlapping...18

Fig. 3-4 The PSNR computation of the sequences: (a) the sequence of “foreman”, (b) the

sequence of “table”………....20

Fig. 3-5 “Table” sequences: (a) previous frame, (b) current frame, (c) frame interpolated

by the proposed method without overlapping, (d) frame interpolated by the proposed method with overlapping, (e) and (f) enlarged regions of (c) and (d) by a factor two, respectively………22

CHAPTER 1

INTRODUCTION

Frame rate up-conversion (FRC), a conversion between any two display formats

with different frame rates, is one of main issues that have arisen in recent years with

the emergence of new television and multimedia system [1]. For improving visual

quality, the frame rate required for HDTV is much higher than the current frame rate

(e.g., 24, 25 or 30 frames/sec) for currently available motion pictures. The frame rate

should be up-converted in HDTV and multimedia PC environments. Moreover, FRC

technique can be used for video transmissions.

In recent years, simple FRC algorithms such as frame repetition and linear

interpolation have been already used. A noted solution [2] for conversion is widely

accepted interoperability practice called “3-2 pulldown”. This technique converts the

frame rate from 24 frames/sec up to 60 frames/sec by a 3:2 frame repeating sequence.

However, some “jerkiness” would be produced when conversion from low and

medium resolution to high resolution sequences as in HDTV. Besides, linear

interpolation causes blurring in the moving areas of the video sequence depending on

the amount of motion.

solution for high quality FRC [3]. Motion estimation (ME) and motion compensation

(MC) are employed in MCI [4-5]. Motion estimation is used to find motion vectors

(MVs) that represent the motion information of objects in an image sequence, while

motion compensation is used to generate the interpolated frame according to MVs

from ME. Hence, the correctness of MVs affects the quality of FRC using MCI. There

are some algorithms proposed for the true motion estimation [6-7]. And several FRC

approaches [8-12] using motion information have been proposed.

FRC using block-based motion compensation may produce overlapped and hole

regions in the interpolated frame, as shown in Fig. 1-1. To avoid the above-mentioned

shortcoming, a kind of new motion compensated frame interpolation (MCFI)

algorithms [10-11,13] is proposed. It [11] compensates the interpolation frame block

Hole region Overlapped region

Previous frame

Interpolated frame

Current frame

by block. To find the blocks in the (k-1)th (previous) and kth (current) frames to be

used for interpolation, start with the estimated motion vector (Vb), as shown in Fig.

1-2(a). Using the assumption of linear motion , we estimate that each block will have

moved one half of its motion vector distance in the interpolated frame. Before

interpolating a block in the interpolated frame, one block from each of the (k-1)th

(previous) and kth (current) frames is searched as shown in Fig. 1-2(b). Although the

method can keep the overlapped and hole regions off, a new problem occurs. It is that

the motion of the compensated block is not always the same as the estimated block

motion. To handle this problem, a refinement of the MVs using bi-directional motion

estimation has been proposed [10]. However, some blurring regions would occur

Vb

Vb / 2

Previous frame (k-1) Interpolated frame Current frame (k)

Fig. 1-2 (a) Estimatied MV, Vb, computed during ME, (b) Interpolation using blocks from (k-1)th and kth frames.

(a)

(b)

since there are not similar enough blocks in motion estimation. Other methods to

estimate the MV field are hierarchical MCFI (HMCFI) [1,14-15]. An approach [1]

based on the Gaussian/Laplacian pyramid structures has been proposed to improve

visual quality of interpolated frames. It focuses on refinement of the MCFI based on a

hierarchical progressive video scheme. However, this method requires large and

complicated operators since frames are decomposed into lowpass and highpass

components and the operators are quite different at top, intermediate and bottom

levels. Beside the effect on the top level affects the quality of the interpolated frame

so much. In [1,10-11,13], the block used to compensate to the interpolated frame is

usually formed by a linear combination from two frames. It would introduce

incomplete or overlapped shadow in the interpolated frame, since the blocks picked to

combine are not exact the same.

In this thesis, a new FRC based on MCI algorithm is proposed. To avoid

unnatural shadow, we construct the interpolated frame by only referring to one frame.

To get a better visual perception, there is a threshold used to abandon the MVs that

provide a large prediction error. To solve the problems of overlapped and hole regions

that are introduced in motion compensation, a weighted average compensation

technique is employed to process the overlapped regions, and a block-based algorithm

the proposed method will be described. Experimental results will be presented in

CHAPTER 2

THE PROPOSED METHOD

Fig. 2-1 shows the block diagram of the proposed method. It consists of two

phases: forward motion estimation and motion compensated interpolation, and

overlapped and hole regions processing. First, to determine how to interpolate the new

frame, the motion vectors are generated by forward motion estimation. Second, the

interpolation frame with motion vectors introduced from the first phase will be

constructed. Based on the natural perception, the information for interpolation only

refers to one frame. Third, the overlapped regions are processed by using the

weighted average compensation. Finally, we segment the interpolation frame into

blocks and compensate the hole regions block by block. In what follows, we will

describe the details of the proposed method.

Previous frame

Current frame

Motion

compensation

Overlapped

region

processing

Hole region

processing

Interpolated

frame

Motion

estimation

2.1 Motion estimation and motion compensated interpolation

A block matching algorithm is employed for forward motion estimation to obtain

the motion vectors. Let x be the spatial domain index, n be the time domain index,

and a frame image is denoted as f x( ,n). Then the motion vector, v, of a macroblock

(MB) is obtained by

{

}

{

}

{

}

∑

∑

B is a macroblock in the search area, S is the search range of ME and t is a threshold

value for abandoning MVs. Note that if the motion of object is not linear, the

difference between the estimated macroblock and the macroblock in the related search

ranges will be large. That is, no appropriate motion vector can be found to estimate

the macroblock. To solve the above shortcoming, in this thesis, there is a threshold

used for abandoning this kind of motion vectors.

Fig. 2-2 shows how a frame is interpolated. For each macroblock I centered at x in

V

Previous frame Interpolated frame Current frame Fig. 2-2 Frame interpolation by the proposed method with Macroblock I inserted

X X + V/2

I I

frame n, we first find its motion vector v, then paste macroblock I in the interpolation

frame 2 1 −

n with center located at 2 v

x+ . That is the interpolated frame is

generated by ) , ( ) 2 1 , ( n f n f x 2 v x+ − = , (3)

where v is the estimated motion vector. To avoid unnatural shadow and blurring, our

compensation only refers to one frame. Fig. 2-3 is an example of result after this step.

2.2 Overlapped and hole regions processing

Overlapped and hole regions are introduced in the interpolated frame since some

adjacent motion vectors are not the same. Moreover, some hole regions are produced

(a)

(b)

The overlapped regions

The hole regions

Fig. 2-3 (a) The previous and current frames. (b) The interpolated frame by the proposed motion estimation and motion compensated interpolation with frames of (a).

due to null motion vectors. Fig. 2-4 shows that the overlapped and hole regions are

introduced in the interpolated frame. In this thesis, we use a counter array to detect the

overlapped and hole regions [1]. Hole regions are pixels with the count value equal to

zero while overlapped regions are pixels with the count value larger than one. A

motion vector array is also needed for hole regions processing.

2.2.1 Overlapped region processing

For reducing blocking artifact, we use a weighted average way [13] to

compensate the interpolated frame for overlapped region. It is calculated as

∑

∑

where v is the motion vector of macroblock that covers the pixel x. That is, all motion

vectors of macroblocks covering the same location x are utilized. For processing hole

Fig. 2-4 Three types of blocks in the interpolation frame. Compensated completely block

Compensated block with partial hole region Complete hole block

Compensated Overlapped Hole

regions which will be described latter, a motion vector array is provided. For each

pixel x covered by several motion compensated blocks, the array will only store the

motion vector of the motion compensated block with minimum prediction error. As a

result, the motion vector array is filled with motion vectors except for hole regions.

Fig. 2-5 is an example of result after this step.

2.2.2 Hole region processing

To compensate the hole regions of the interpolated frame efficiently, the

interpolated frame will be divided into blocks of the same size and the hole regions

will be compensated block by block. Different kinds of blocks need different methods

to compensate the hole region. As shown in Fig. 2-4, there are three kinds of blocks:

compensated completely, compensated with partial hole region, complete hole region.

In what follows, we will introduce two kinds of methods to perform hole region

processing.

The hole regions

Fig. 2-5 The interpolated frame about Fig. 2-3(b) after processing of the overlapped regions by the proposed method.

2.2.2.1 The process of block compensated with partial hole region

Fig. 2-6 shows the flow diagram of the compensated method. Basically, the

proposed algorithm detects the hole region in a block by scanning it row by row

(column by column), and compensates the hole region by the motion vector of the

adjacent pixel. The order of scanning direction and the start row (column) are shown

as in Fig. 2-7 respectively. As shown in Fig. 2-6(b), if the first pixel on a row (or

column) according to scan direction belongs to the hole region, stop the block scan

and rescan from a new direction according to the proposed order. In our experiment,

the effective scanning runs two directions at most. Since the second direction

scanning can compensate the hole region totally referring to the information

compensated in the first direction scanning if it needs (see Fig. 2-6(c)). Fig. 2-8 is an

example of result after this step.

(a) (b) (c) (d)

Fig. 2-6 (a) A block with partial hole region (black part). (b) Scan the block row by row, and stop scanning when the first scanned pixel in a row belongs to the hole region. (c) Change the scanning direction and scan the block column by column. (d) The compensated block.

2.2.2.2 The processing of block with complete hole region

If the values of counter array on the four corners of a block are zeros, then the

block is a complete hole region. In this thesis, based on the motion information from

the block’s neighbors, we propose a method to decide the motion vector, v, of the

hole block (see Fig. 2-9). That is,

{

}

where N(u) is the times of u occurring in S and S is the neighboring pixels around the

1(5)

2

3

4

Fig. 2-7 The order of the block scanning direction with bold arrow indicating the first scanning row (column).

The hole regions

Fig. 2-8 The interpolated frame about Fig. 2-5 after processing of the compensated block with the partial hole regions by the proposed method.

hole block and shown in Fig. 2-9. Fig. 2-10 is an example of result after this step.

2.2.3 Motion estimation with overlapping

Since there are hole regions produced in motion compensation, overlapping

processing block would reduce the hole regions. A frame is divided into several

blocks with half block overlapping and each block is used to do motion estimation

and compensation, as shown in Fig. 2-11. Based on this overlapping schema, the

number of hole regions is reduced, and there are fewer blocking artifact. Fig. 2-12 is

an example of result after this step.

Fig. 2-10 The interpolated frame about Fig. 2-8 after processing of the complete hole block regions by the proposed method.

Fig. 2-11 Motion estimation using overlapping blocks.

The hole regions

Fig. 2-12 The interpolated frame about Fig. 2-3(a) after processing of the overlapped regions with half block overlapped.

CHAPTER 3

EXPERIMENTAL RESULTS

In this section, we present some experimental results. The experiments are

focused on doubling the frame rate. The size of the marcoblock is set as 16×16 and

the search range is ± 16 pixels for both horizontal and vertical directions. The

threshold for abandoning motion vectors in motion estimation is set as 400. We

select three different video sequences for testing. They are “calendar”, “foreman” and

“table”. Test sequences are color sequences with YUV 4:1:1 format. The image size of

test sequences is 176×144 quarter common intermediate format (QCIF). In order to

show the brief and efficiency of the proposed method, we compare the results with an

algorithm of hierarchical MCFI (HMCFI) [1].

3.1 Comparison with hierarchical MCFI

Fig. 3-1 is an example of scene with multiple motions. It contains camera

zooming and panning. There are also several object motions such as horizontal

translation (train), vertical translation (calendar) and rotation (ball). The speeds of

motions are not fast. As shown in Fig. 3-1, especially Fig. 3-1(e) and Fig. 3-1(f), the

the interpolated frame by HMCFI is usually compensated by two frames with

averaging while that by the proposed method is only compensated by one frame.

(a) (b)

(c) (d)

(e) (f)

Fig. 3-1 “Calendar＂ sequences : (a) previous frame, (b) current frame, (c) frame interpolated by the HMCFI, (d) frame interpolated by the proposed method, (e) an enlarged region of (c) by a factor of two, (f) an enlarged region of (d) by a factor of two.

Fig. 3-2 is an example of scene with motion of panning. The camera moves from

left to right and the scene shifts to right with a medium speed. Although the motion is

simple, some determinations of motion vectors are difficult. It is easy to find that

(a) (b)

(c) (d)

(e) (f)

Fig. 3-2 “Foreman＂ sequences : (a) previous frame, (b) current frame, (c) frame interpolated by the HMCFI, (d) frame interpolated by the proposed method, (e) an enlarged region of (c) by a factor of two, (f) an enlarged region of (d) by a factor of two.

there are two faulted macroblocks marked by circles compensated with faulted motion

vectors on the right-up corner in the interpolated frame by HMCFI. In this case, the

proposed method provides a more precise motion estimation and a better

compensation.

Fig. 3-3 is an example of scene with motion of zooming. The camera zooms

from the near to the distant and the scene becomes larger gradually. Among the frames,

the left hand moves fast. From Fig. 3-3(c), we can see that the shape of the left hand

in the HMCFI interpolated frame is blurred seriously, even the table and the hand are

indistinct. And the shape of the hand in the proposed interpolated frame approaches

(a) (b)

(c) (d)

Fig. 3-3 “Table＂ sequences : (a) previous frame, (b) current frame, (c) frame interpolated by the HMCFI, (d) frame interpolated by the proposed method without overlapping.

the original one. The poster on the left up corner in the interpolated frame by HMCFI

is more blurred than that in the proposed interpolated frame. This is due to that

HMCFI adapt to two average compensation blocks from two frames.

Fig. 3-4 shows the PSNR for the interpolated frames by the proposed method and

HMCFI method. In order to compute the PSNR of an interpolated sequence, we drop

the frames of even number on purpose and reproduce each dropping even frame by its

previous and next odd number frames. And then we compare the interpolated frames

with original ones. Fig. 3-4(a) is the result of the “foreman” sequence. The PSNRs of

the proposed method are a little lower than those of HMCFI. It is due to that the frame

constructed by the proposed method only refers to one frame. However, the

interpolated frame by the proposed method is clearer from the visual point. Fig. 3-4(b)

is the result of the “table”. The PSNR of the 67th interpolated frame by the proposed

method is much lower since a scene change appears here. From Fig. 3-4, we found

that PSNRs of both methods in most parts are similar. However, the proposed method

provides higher quality from visual point and is simpler than HMCFI.

3.2 Effectiveness of overlapping block

In order to reduce the hole regions produced from MC and the blocking artifact,

20 25 30 35 40 45 50 1 10 19 28 37 46 55 64 73 82 91 100 109 118 127 136 145 interpolated frames PSN R (d b )

the proposed method HMCFI

(a) 20 25 30 35 40 45 50 1 14 27 40 53 66 79 92 105 118 131 144 157 170 183 196 the interpolated frames

PSN

R

(d

b

)

the proposed method HMCFI

(b)

Fig. 3-4 The PSNR computation of the sequences: (a) the sequence of “foreman”, (b) the sequence of “table”.

schema affects on the perceptual quality of the interpolated frame, we simulated two

types of frame rate up-conversion:

(a) the proposed method without overlapping block

(b) the proposed method with overlapping block

Fig. 3-5 shows that the perceptual quality is improved if the overlapping schema is

used. In Fig. 3-5(e) and 3-5(f), we can see that, the interpolated frame produced with

overlapping is smoother and has fewer block artifacts. It is obvious that the block

(a) (b)

(c) (d)

(e) (f)

Fig. 3-5 “Table＂sequence: (a) previous frame, (b) current frame, (c) frame interpolated by the proposed method without overlapping, (d) frame interpolated by the proposed method with overlapping, (e) and (f) enlarged regions of (c) and (d) by a factor two, respectively.

CHAPTER 4

CONCLUSIONS

In this thesis, we have presented a new motion compensated interpolated

algorithm for frame rate up-conversion. The proposed method consists of two

component: forward motion estimation and motion compensation, and overlapped and

hole regions processing. To get a naturally perceptual and non-blur scene, two new

concepts are provided. The motion vectors are abandoned when the prediction errors

are too large, and the process of compensation only refers to one frame. Moreover,

overlapping blocks make the hole regions fewer and the compensation precisely.

REFERENCES

[1] B.-W. Jeon, G.-I. Lee, S.-H. Lee, and R.-H. Park, “Coarse-to-Fine Frame

Interpolation for Frame Rate Up-Conversion Using Pyramid Structure,” IEEE

Trans. Consumer Electronics, vol. 49, no. 3, pp. 499-508, Aug. 2003.

[2] K. A. Bugwadia, E. D. Petajan, and N. N. Puri, “Progressive-Scan Rate

Up-Conversion of 24/30 Source Materials for HDTV,” IEEE Trans. Consumer

Electronics, vol. 42, no. 3, pp. 312-321, Aug. 1996.

[3] S. C. Han and J. W. Woods, “Frame-rate Up-conversion Using Transmitted Motion

and Segmentation Fields for Very Low Bit-rate Video Coding,” in Proc. Int. Conf.

Image Processing, vol. 1, pp. 747-750, Oct. 1997.

[4] ITU-T Recommendation H.263, Video Coding for Low Bit-Rate Communications.

Nov. 1995.

[5] ISO/IEC JTC1/SC29/WG11 MPEG93/N457, MPEG-2 Test Model Version5. Mar.

1993.

[6] J. N. Youn, M. T. Sun, and C. W. Lin, “Motion estimation for high performance

transcoding,” IEEE Trans. Consumer Electronics, vol. 44, no. 3, pp. 649-658,

1998.

[7] G. D. Haan, P. W. A. C. Biezen, H. Huijgen, and O. A. Ojo, “True-motion

Video Technol., vol. 3, no. 5, pp. 368-379, Oct. 1993.

[8] F. Dufaux and F. Moscheni, “Motion estimation techniques for digital TV: A review

and a new contribution,” Proc. IEEE, vol. 83, pp. 858-876, June. 1995.

[9] S.-H. Lee, Y.-C. Shin, S.-J. Yang, H.-H. Moon, and R.-H. Park, “Adaptive motion

compensated interpolation for frame rate up-conversion,” IEEE Trans. Consumer

Electronics, vol. 48. no. 3, pp. 444-450, Aug. 2002.

[10] B.–T. Choi, S.–H. Lee, and S.–J. Ko, “New frame rate up-conversion using

bi-directional motion estimation,” IEEE Trans. Consumer Electronics, vol. 46, no.

3, pp. 603-609, Aug. 2000.

[11] K. Hilman, H.–W. Park, and Y.-M. Kim, “Using motion compensated frame-rate

conversion for the correction of 3:2 pulldown artifacts in video sequences,” IEEE

Trans. Circuits Syst. Video Technol., vol. 10, no. 6, pp. 869-877, Sept. 2000.

[12] R. Castagno, P. Haavisto, and G. Ramponi, “A method for motion adaptive frame

rate up-conversion,” IEEE Trans. Circuits Syst. Video Technol., vol. 6, no. 5, pp.

436-446, Oct. 1996.

[13] S.-H. Lee, O. Kwon, and R.-H Park, “Weighted-Adaptive Motion-Compensated

Frame Rate Up-Conversion,” IEEE Trans. Consumer Electronics, vol. 49, no. 3,

pp. 485-492, Aug. 2003.

covered and uncoverd background,” Signal Processing: Image Compression, vol.

1, pp. 192-212, 1989.

[15] M. Bierling and R. Thoma, “Motion Compensating field interpolation using a

hierarchically structured displacement estimator,” Signal Processing, vol. 11, no. 4,