KR100287214B1 - Method and system for encoding motion pictures - Google Patents

Abstract

PURPOSE: A method and a system for encoding motion pictures are provided to reduce the amount of generation of bits, improve picture quality, and simplify a processing configuration. CONSTITUTION: A system for encoding motion pictures includes the first and second buffers, a motion estimation and compensation unit(52), an MC error generation and division unit(53), the first and second quad-tree processors(54,55), a multiplexer(56), a quantizer(57), a variable-length coder(58), a bit rate controller(60), an inverse quantizer(61), the first and second inverse quad-tree processor(62,63), an adder and a filter. The system divides MC error generated after motion estimation into accumulation error and compensation error, and encodes the errors according to their characteristics.

Description

Video encoding method and apparatus

1 illustrates the structure of a general video encoder.

FIG. 2 shows the number of lowermost regions generated when the quadtree is applied to a memc error, an acc error, and a motion compensation (MC) error.

3 is a block diagram showing a video encoding apparatus according to the present invention.

4 is a detailed block diagram of the motion estimation and compensator and the MC error generation and separator shown in FIG.

FIG. 5 is a flowchart for explaining the operation of the first and second quadtree processors shown in FIG.

FIG. 6 is a detailed block diagram of the adder and the filter shown in FIG.

7 is a diagram schematically showing a processing structure for cumulative errors and compensation errors.

8 shows the experimental results of filtering cumulative errors remaining after the compensation error is processed.

<Explanation of symbols for main parts of drawing>

51, 59: buffer 52: motion estimation and compensator

53: MC error occurrence and separator 54, 55: Quadtree processor

56 multiplexer 57 quantizer

58: variable length encoder 60: bit rate controller

61: inverse quantizer 62, 63: inverse quad tree processor

64: adder and filter

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a video encoding method and apparatus, and more particularly, to a method and apparatus for separating an MC error generated after motion compensation into a cumulative error and a compensation error according to a characteristic, and encoding the same according to each characteristic.

In general, video encoders that are widely used have the effect of data compression by removing temporal redundancy through motion processing and spatial redundancy through discrete cosine transform (hereinafter, referred to as DCT).

1 is a basic structure of a video encoder that is designed to remove temporal redundancy through motion processing and spatial redundancy through DCT, and is used in many standardized encoders such as H.261, MPEG-1, and MPEG-2.

Referring to FIG. 1, the motion estimator 11 generates a motion vector for configuring a current image by referring to a previous image. Most encoders perform a full search based on the minimum absolute error (MAE) within a fixed range in units of fixed blocks to generate a motion vector. The motion vector generated by the motion estimator 11 is sent to the frame store and the motion compensator 17 to form a motion compensation image through motion compensation by referring to the reconstructed image of the previous image already stored in the memory. Therefore, the motion compensation image may be very similar to the current image through the motion estimation of the motion estimator 11 and the motion compensation of the frame store and the motion compensator 17.

The subtractor 18 calculates a difference between the current image and the motion compensation image to generate a motion compensation (hereinafter, MC) error. That is, combining the motion compensation image with the MC error produces a complete current image. The MC error is sent to the DCT 12 and converted from the spatial domain to the frequency domain.

The quantizer Q 13 performs quantization of the MC error converted by the DCT 12 according to the quantization interval generated by the bit rate controller 18. The MC error quantized by the quantizer 13 is subjected to variable length encoding by the variable length encoder 19 to be transmitted to the buffer 21 which is the last stage of the encoder through the multiplexer 20 and transmitted to the decoder.

The inverse quantizer (Q ^-1 ; 14) and the inverse discrete cosine transformer (DCT ^-1 ; 15) perform inverse quantization and inverse discrete cosine transforming the MC error through inverse transformation processes of the quantizer 13 and the DCT 12, respectively. The adder 16 then combines the motion compensated image to generate a reconstructed image of the current image, and then stores it in the frame store and the motion compensator 17. The reconstructed image generated at this time is the same image as the reconstructed image generated by the decoder.

Therefore, the data sent from the encoder to the decoder is an MC error in which a motion vector and some data are lost. The decoder constructs a motion compensation image using the input motion vector, and inversely transforms an MC error to generate a reconstructed image such as a reconstructed image stored in the frame store and the motion compensator 17.

The MC error output from the subtractor 18 is caused by two causes. First, if two or more different motion components exist in the block by performing motion estimation based on the fixed block of the motion estimator 11, the frame store and the motion compensator 17 cannot perform accurate motion compensation. An error with the original image is generated. The error at this time will be referred to as a compensation error. Compensation errors occur frequently at the boundary edges of objects moving in the image and rarely occur elsewhere. Strobach calls this effect the 'line drawing effect'.

The second cause of MC error is cumulative error. Since the MC error is encoded by the following various processes, if all MC errors are encoded, no error occurs in the reconstructed image. However, an error occurs while quantizing according to the quantization interval set by the bit rate controller 18 according to the target bit rate. Therefore, an error exists in the reconstructed image generated by the decoder, the frame store, and the motion compensator 17. These errors accumulate as the image processing continues and do not have large values, but are very wide and random.

After all, since the MC error is generated by adding the compensation error and the cumulative error, it has both the characteristics of the compensation error and the cumulative error. That is, it has a high error value at the boundary of a moving object in the image such as a compensation error, and low error values are irregularly distributed over all regions such as a cumulative error.

The above-mentioned MC error is converted from the spatial domain to the frequency domain through the DCT, and the power concentration can be expected. However, the concentration of the power through the DCT can be achieved only when there are many co-relations of the values of the pixels in the spatial domain. It is possible. Since the MC error has the property of irregular distribution of cumulative errors as it is, when the DCT is performed on the MC error having an irregular distribution, power is not concentrated and is more often distributed.

It is already experimentally proved by Strobach that the concentration of power cannot be expected by performing DCT on the MC error. Strobach pointed out the DCT's inefficiency for MC errors and proposed a quadtree encoding method as an alternative.

The encoding using quadtree encodes the average value of the values in the block if the pixels in the block have a similar value for a block of a constant size, and divides the data into four equal parts if the pixels in the block have very different values. We will examine the similarity of the values in the block for three small blocks. In this way, the block is divided into four quarters until the condition is met, and the mean value in the block and the quadtree structure are encoded.

Although MC errors are composed of cumulative error and compensation error in real video coding, Strobach assumes only MC compensation characteristics as a simple compensation error, and considers the characteristics of MC errors due to irregular distribution of cumulative errors. I'm not doing it. That is, since the dispersion is large due to the irregular distribution of the cumulative errors in other areas as well as the error concentrated on the boundary edge of the moving object, the regions are very finely divided in all regions when the encoding is performed using the quadtree. Therefore, the information about the quadtree structure to be encoded, as well as the average value for each region, must be encoded, resulting in a phenomenon in which the information amount is called out.

Also, due to the irregular distribution of cumulative errors, the average value often has a value close to '0'. In this case, even if the average value is transmitted to compensate for the error, the effect is not remarkable.

FIG. 2 shows the number of leaf nodes generated when divided into quadtrees by applying the same dispersion threshold 400 for compensation error, cumulative error, and MC error (compensation error + cumulative error). In this case, the cumulative error can be seen that the number of the lowest region is much more than the compensation error, and accordingly the MC error is also increased.

As described above, the DCT method widely used for encoding the MC error is inefficient due to the irregular distribution of the MC error and the high frequency component at the edge. In addition, it can be seen that the quadtree method proposed as an alternative is also inefficient due to the irregular distribution of MC errors.

In addition, due to data loss occurring in the encoding process, an error is caused in the quantization process according to the quantization interval generated in the bit rate controller 18. In this case, when the target transmission rate is high, the error is small because the quantization interval is small, whereas when the target transmission rate is low, the error is very severe because the quantization interval is large, so that the image quality of the reconstructed image is very poor.

Especially, in the case of ultra low bit rate video encoding of 64 Kbps or less, due to limited bit rate, the blocking effect is that the size of the processed block appears on the screen because a lot of data is lost because the quantization interval becomes very large in the general DCT or quadtree processing structure. In addition, the deterioration of the image quality is very prominent due to the mosquito effect in which errors are concentrated in the form of a dot on the boundary of a moving object and a phenomenon in which the boundary of the object does not match.

Accordingly, an object of the present invention is to solve the above problems by separating the MC error generated after motion compensation into cumulative errors and compensation errors according to characteristics, and encoding them according to the characteristics to be effective in terms of bit generation and reconstruction quality. The present invention provides a video encoding method for performing encoding.

Another object of the present invention is to provide an apparatus most suitable for realizing the video encoding method.

In order to achieve the above object, a video encoding method according to the present invention

A motion compensation process of generating a motion vector by performing motion evaluation on the current original image with reference to a previous original image, and generating a motion compensated image by referring to the motion vector and the previous reconstructed image;

An error separation step of separating a motion compensation error generated after the motion compensation process into a cumulative error and a compensation error;

A quadtree processing step of quadtree processing the accumulated error and the compensation error separated in the error separation process so as to match the characteristics of each error;

A quantization step of performing quantization on the mean value in each quadtree of the cumulative error and the compensation error processed in the quadtree processing, and determining a quantization coefficient according to a predetermined quantization interval;

A variable length encoding process of performing variable length encoding on the quantized data in the quantization process;

A storage process of storing data output in the variable length encoding process;

A bit rate control process of determining a quantization interval to be provided to the quantization process according to the amount of data stored in the storage process;

An inverse quantization process of converting quantized data into data before being quantized in the quantization process;

An inverse quadtree process of generating a reproduction compensation error and a reproduction accumulation error by performing an inverse quadtree process on a block having an average value in a specific subblock among the output data of the inverse quantizer; And

A reconstructed image that generates a final reconstructed image through low pass filtering after a first reconstructed image is generated from a reproduction compensation error generated during the inverse quadtree processing, a reproduction accumulation error, and a motion compensation image generated in the motion compensation process. Characterized in that it comprises a generation process.

In order to achieve the above object, a video encoding apparatus according to the present invention

A first buffer for storing a previous original image and a current original image;

A motion vector is generated by performing motion evaluation on the current original image by referring to the previous original image stored in the first buffer, and a motion for generating a motion compensation image by referring to the generated motion vector and the previous reconstructed image. Estimator and compensator;

A motion compensation error generation and separator for generating a motion compensation error with the motion compensation image generated by the motion estimation and compensator and the current original image, and separating the compensation error and the cumulative error;

First and second quadtree processors for performing quadtree processing for encoding on the compensation error and the cumulative error output from the motion compensation error generator and the separator;

A switching unit for switching and outputting the motion vector generated by the motion estimation and compensator and output data of the first and second quadtree processors;

A quantizer for performing quantization on an average value in the quadtree among the output data of the switching unit and determining a quantization coefficient according to a predetermined quantization interval;

A variable length encoder for performing variable length coding on the data quantized by the quantizer;

A second buffer for storing data output from the variable length encoder;

A bit rate controller for determining a quantization interval to be provided to the quantizer according to the amount of data stored in the second buffer;

An inverse quantizer for converting quantized data into data before being quantized;

First and second inverse quad tree processors for generating a compensation compensation error and a cumulative reproduction error by performing inverse quadtree processing on a block having an average value in a specific subblock among the output data of the inverse quantizer; And

After generating the first reconstructed image by adding the reproduction compensation error and the reproduction accumulation error output from the first and second inverse quad tree processors to the motion compensation image output from the motion estimator and the compensator, they are irregularly distributed through filtering. And an adder and a filter for processing an error to generate a final reconstructed image.

Hereinafter, the present invention will be described in detail with reference to the accompanying drawings.

3 is a block diagram showing a video encoding apparatus according to the present invention, wherein the first and second buffers 51 and 59, the motion estimation and compensator 52, the MC error generator and separator 53, and the first and second buffers are shown. Quadtree processors 54, 55, switchers for example multiplexer 56, quantizer 57, variable length encoder 58, bit rate controller 59, dequantizer 61, first and second inverse Quadtree processors 62, 63, adders and filters 64 are included.

Referring to the operation of the video encoding apparatus shown in FIG. 3, the first buffer 51 stores current and previous original images, and the motion estimation and compensator 52 stores the previous circles stored in the first buffer 51. A motion vector is generated by performing motion evaluation on the current image with reference to the image, and a motion compensation (MC) image is generated by referring to the generated motion vector and the reconstructed image of all the stored images.

The MC error generation and separator 53 generates an MC error using the MC image generated by the motion estimation and compensator 52 and the current original image, and separates the MC error to generate a compensation error and a cumulative error.

The first and second quadtree processors 54 and 55 perform quadtree processing for encoding on the compensation error and the cumulative error output from the MC error generator and the separator 53, respectively.

The multiplexer 56 receives a motion vector generated by the motion estimator and compensator 52 and output data of the first and second quadtree processors 54 and 55, sequentially selects the motion vectors, and transmits the same to a decoder (not shown).

The quantizer 57 performs quantization on the average value in the quadtree among the data supplied from the multiplexer 56 and determines the quantization coefficient according to the quantization interval supplied from the bit rate controller 60.

The variable length encoder 58 performs variable length coding on the data quantized by the quantizer 57, and the second buffer 59 stores the output data of the variable length encoder 58.

The bit rate controller 60 determines the quantization interval to be provided to the quantizer 57 according to the amount of data stored in the second buffer 59, and the inverse quantizer 61 performs operations opposite to the quantizer 57. And converts the quantized data into a form similar to the data before being quantized.

The first and second inverse quad tree processors 62 and 63 extract only a block having an average value from a specific subblock among the output data of the inverse quantizer 61 to generate a reproduction compensation error and a reproduction accumulation error.

In the adder and the filter 64, the reproduction compensation error and the accumulated accumulation error are added to the MC image output from the motion estimation and the compensator 52 to form a first reconstructed image, and then an irregular distribution error is processed by filtering to obtain a final reconstructed image. Create a reconstruction image.

First, since the characteristics of the MC error are a mixture of the characteristics of the cumulative error and the compensation error, the MC error e _t at the time _t can be expressed by the following equation (1).

In the first equation, ME (k) is a motion vector generated by referring to k image, MC (v, k) is an image reconstructed by using motion v and referring to k image, Are images reconstructed at time t, and x _t represents the original image at time t.

By separating the MC image represented by the first equation into a compensation error (memc) and cumulative (acc) error, as shown in the second equation to enable the processing suitable for the characteristics of each error.

In the second equation e _t can be obtained from the first equation, e _memc is Therefore, cumulative and compensation errors can be obtained through e _acc = e _t -e _memc . Compensation errors are distributed around the outer edge of the moving object and are very small elsewhere. Compensation error is already in the state of gathering the error value, if you divide by using the quadtree, it can be effectively encoded because it is divided into only small areas where the error is gathered. In the rest of the regions, since the average value is close to '0', the encoding is not performed.

On the other hand, the cumulative error is distributed throughout the image and is very irregular and mostly has a small value, and the average of the cumulative error in a large block is mostly '0'. Therefore, the low pass filtering process can be processed effectively without any data transmission. However, the cumulative error is an error that results from incomplete processing of the compensation error, so the cumulative error is processed in the same way as the compensation error using the quadtree. At this time, the cumulative error is rarely divided into smaller areas than the compensation error, and even after processing with the average value, the low pass filtering is performed after the average value processing because it has a small value and an irregular distribution.

4 is a detailed block diagram of the motion estimator and compensator 52 and the MC error generation and separator 53 shown in FIG. 3, and the motion estimator and compensator 52 includes the motion estimator 101, Two motion compensators (102, 103), MC error generation and separator 53 is composed of a buffer 107, first to third subtractors (108, 109, 110).

Referring to the operation of the motion estimation and compensator 52 shown in FIG. 4, the motion estimator 101 generates a motion vector, and the first motion compensator 102 refers to the original image of the previous image and the motion estimator 101. ) Configures the first MC image using the motion vector output from the motion vector, and the second motion compensator 103 refers to the reconstructed image of the previous image and configures the second MC image using the motion vector output from the motion estimator 101. do.

Meanwhile, referring to the MC error generation and the operation of the separator 53, the buffer 107 stores the second MC image output from the second motion compensator 103 and composes a reconstructed image. Compensation error is generated by obtaining a difference between the first MC image output from the motion compensator 102 and the original image of the current image, and the MC image output from the second motion compensator 103 and the current are output from the second subtractor 109. The MC error is generated by obtaining the difference between the original image of the image, and the third subtractor 110 calculates and accumulates the difference between the MC error output from the second subtractor 109 and the compensation error output from the first subtractor 108. Generate an error.

FIG. 5 is a flowchart for explaining the operation of the first and second quadtree processors 54 and 55 shown in FIG.

When an error image of a predetermined block size is input in operation 151, in operation 152, an average value M of the entire block and average values M1, M2, M3, and M4 of four subblocks divided into quadtrees are calculated.

In step 153, it is determined whether the average value M is smaller than the preset threshold TH. If the average value M is smaller than the threshold TH, it is determined whether the average values M1, M2, M3, and M4 of the sub-blocks are smaller than the threshold TH (th). Step 154). If the average values M1, M2, M3, and M4 of the subblocks are smaller than the threshold value TH in step 154, the quadtree-type split is stopped and the quadtree structure stored so far is transmitted to the multiplexer (56 of FIG. 4). (Step 155).

On the other hand, if the average value M is greater than the threshold value TH in step 153, the average value M is subtracted from all values in the block, and then a new block value is generated (step 156), and the average value of the new subblocks is obtained. (Step 157). In step 158, an average value of the new subblocks generated in step 157 is used. It is judged whether it is smaller than this threshold value TH, and the average value of subblocks If all of these are smaller than the threshold value TH, the quadtree type split is stopped and the quadtree structure stored so far is transmitted to the multiplexer (56 in FIG. 4) (step 155).

On the other hand, the average value of the sub-blocks in step 158 If any one is larger than the threshold TH, it is determined whether the size of the current block is larger than the minimum block (step 159). If the size of the current block is greater than the minimum block in step 159, the block is divided into quidtrees, the quadtree structure is stored, and the flow returns to step 152 to continue the quadtree processing for each sub-block (step 160). .

FIG. 6 is a detailed block diagram of the adder and filter 64 shown in FIG. 3, which includes first and second adders 201 and 202, filter tap and coefficient determiner 203, and low pass filter 204. FIG. It is composed.

Referring to the operation of the adder and the filter 64 shown in FIG. 6, the first adder 201 uses the MC image and the first inverse quad tree processor (FIG. 4 of FIG. The reproduction compensation error output from 62 is added, and the second adder 202 adds the output of the first adder 201 and the accumulated accumulation error output from the second inverse quadtree processor (63 of FIG. 4).

The filter tap and coefficient determiner 203 determines the number of taps according to the subblock size of the reproduction accumulated error output from the second inverse quadtree processor (63 of FIG. 4), and the reconstructed image output from the first adder 201. The filter coefficients are determined by the variance at the same position and the same size, and the low pass filter 204 filters the output of the second adder 202 by the filter coefficients determined by the filter tap and the coefficient determiner 203.

7 is a diagram schematically showing a processing structure for cumulative errors and compensation errors.

In FIG. 7, it is assumed that a) is an original signal and b) is a cumulative error. Cumulative error b) is irregular and exists over the entire signal area. When the cumulative error b) and the original signal a) are combined, it becomes as c), and low pass filtering is performed on c) to generate a reconstruction signal as in d). Thus, a low pass filtering can generate a reconstruction signal very similar to the original signal because the average value of the cumulative errors is '0' and is irregularly distributed. e) indicates a compensation error. When the original error of a) and the compensation error of e) are combined, the signal becomes a signal such as f). When f) is divided into quadtrees (binary trees in the primary signal), only regions where compensation errors exist, such as g), are finely divided. Since the errored area is replaced with the mean value, the compensation error restored using the quadtree and the mean value is equal to h). Therefore, subtracting h) from f) yields a reconstructed image like i). The quadtree method for the compensation error can be effectively handled because most of the errors in the compensation error are collected in a specific area.

8 is a result of experimenting with the degree to which the cumulative error is processed through filtering in the processing of the cumulative error and the compensation error. The video data used is an image called "foreman" and is sampled at 10Hz and has a size of 352 × 288. In the experiment, the result is filtered using a simple two-dimensional 5-tap and fixed coefficient values instead of the adaptive filtering according to the image. Therefore, more effective results can be expected when the adaptive filtering is used. As can be seen from the experimental results, it can be seen that the picture quality improvement of 3 ~ 4dB is achieved without additional coding process for cumulative error.

As described above, in the video encoding method and apparatus according to the present invention, the motion compensation error signal is divided into a compensation error and a cumulative error, and the error signal concentrated in the compensation error is processed by a quadtree method to encode an average value and the remaining signals. The small size signal is filtered and the cumulative error is mostly filtered according to its characteristics, and only a part of the average value is encoded, thereby reducing bit generation as compared to the conventional discrete cosine transform or quadtree method. There is this.

Also, the effective coding of compensation error can greatly reduce the disparity of the boundary of the object, and greatly reduce the phenomenon such as mosquito effect and blocking effect through filtering. Compared with this, the image quality is improved.

In addition, the processing structure used in the present invention can be processed in the spatial domain without changing the area of the data, thereby eliminating the processing delay caused by the area transformation like the discrete cosine transform, and the general quadtree structure in the processing of cumulative error and compensation error. Can be used to eliminate the need for additional hardware due to the separation of errors.

Claims (9)

A motion compensation process of generating a motion vector by performing motion evaluation on the current original image with reference to a previous original image, and generating a motion compensated image by referring to the motion vector and the previous reconstructed image; An error separation step of separating a motion compensation error generated after the motion compensation process into a cumulative error and a compensation error; A quadtree processing step of quadtree processing the accumulated error and the compensation error separated in the error separation process so as to match the characteristics of each error; A quantization step of performing quantization on the mean value in each quadtree of the cumulative error and the compensation error processed in the quadtree processing, and determining a quantization coefficient according to a predetermined quantization interval; A variable length encoding process of performing variable length encoding on the quantized data in the quantization process; A storage process of storing data output in the variable length encoding process; A bit rate control process of determining a quantization interval to be provided to the quantization process according to the amount of data stored in the storage process; An inverse quantization process of converting quantized data into data before being quantized in the quantization process; An inverse quadtree process of generating a reproduction compensation error and a reproduction accumulation error by performing an inverse quadtree process on a block having an average value in a specific subblock among the output data of the inverse quantizer; And reconstructing the final reconstructed image through low pass filtering after creating a first reconstructed image from the regeneration compensation error generated during the inverse quadtree processing, the reproducing accumulation error, and the motion compensation image generated in the motion compensation process. Video encoding method comprising the image generation process. The video encoding method of claim 1, wherein, in the quadtree processing, most of the cumulative errors are lowpass filtered, and the quadtree processing is performed only on a portion of which the average value is not '0'. The video encoding method of claim 1, wherein the error among the compensation errors in the quadtree processing is quadtree-processed to encode an average value, and the remaining small error is not encoded. A first buffer for storing a previous original image and a current original image; A motion vector is generated by performing motion evaluation on the current original image by referring to the previous original image stored in the first buffer, and a motion for generating a motion compensation image by referring to the generated motion vector and the previous reconstructed image. Estimator and compensator; A motion compensation error generation and separator for generating a motion compensation error with the motion compensation image generated by the motion estimation and compensator and the current original image, and separating the compensation error and the cumulative error; First and second quadtree processors for performing quadtree processing for encoding on the compensation error and the cumulative error output from the motion compensation error generator and the separator; A switching unit for switching and outputting the motion vector generated by the motion estimation and compensator and output data of the first and second quadtree processors; A quantizer for performing quantization on an average value in the quadtree among the output data of the switching unit and determining a quantization coefficient according to a predetermined quantization interval; A variable length encoder for performing variable length coding on the data quantized by the quantizer; A second buffer for storing data output from the variable length encoder; A bit rate controller for determining a quantization interval to be provided to the quantizer according to the amount of data stored in the second buffer; An inverse quantizer for converting quantized data into data before being quantized; First and second inverse quad tree processors for generating a compensation compensation error and a cumulative reproduction error by performing inverse quadtree processing on a block having an average value in a specific subblock among the output data of the inverse quantizer; And a reproduction compensation error outputted from the first and second inverse quad tree processors and a cumulative reproduction error respectively added to the motion compensation image output from the motion estimator and the compensator to form a first reconstructed image, and then irregularly distributed through filtering. And an adder and a filter for processing the error to generate a final reconstructed image. The apparatus of claim 4, wherein the motion estimation and compensator comprises: a motion estimator for generating a motion vector by performing motion estimation on the current original image with reference to a previous original image; A first motion compensator for referring to a previous original image and generating a first motion compensated image by using the motion vector; And a second motion compensator configured to refer to a previous reconstructed image and to generate a second motion compensated image by using the motion vector. The apparatus of claim 5, wherein the motion compensation error generation and separation unit comprises: a buffer configured to store a second motion compensation image output from the second motion compensator to generate a reconstructed image; A first subtractor for generating a compensation error by obtaining a difference between the first motion compensation image output from the first motion compensator and the current original image; A second subtractor for generating a motion compensation error by obtaining a difference between a second motion compensation image output from the second motion compensator and a current original image; And a third subtractor for generating a cumulative error by obtaining a difference between a motion compensation error output from the second subtractor and a compensation error output from the first subtractor. 5. The apparatus of claim 4, wherein the adder and the filter comprise: a first adder for adding a motion compensation image output from the motion estimation and compensator and a reproduction compensation error output from the first inverse quad tree processor; A second adder for adding an output of the first adder and a reproduction accumulated error output from the second inverse quadtree processor; A filter tap and coefficient determiner for determining a filter tap according to a block size of a reproduction accumulated error output from the second inverse quad tree processor, and determining a filter coefficient by variance of a reconstructed image output from the first adder; And a low pass filter for filtering the output of the second adder by the filter coefficient determined by the filter tap and the coefficient determiner. 8. The video encoding apparatus of claim 7, wherein the number of filter taps in the filter tap and coefficient determiner is proportional to a size of a target block among reproduction accumulated errors output from the second inverse quadtree processor. The filter coefficients of claim 8, wherein the filter coefficients of the filter tap and coefficient determination unit are equal to the target block among the reproduction accumulation errors output from the second inverse quadtree processor, and the dispersion of the motion compensation image in which the compensation errors are combined at the same position. The larger the coefficient value toward the center of the low-pass filter is larger than the other coefficient value, and if the variance is small, all the coefficient values are similar.