KR20000031283A - Image coding device - Google Patents

Abstract

PURPOSE: An image coding device is provided to improve an efficiency of a zero tree coding by rearranging DCT(Discrete Cosine Transform) coefficients and performing zero tree coding for them. CONSTITUTION: An image coding device comprises a conversion portion, a rearrangement portion(201), and a zero tree coding portion(202). The conversion portion converts space areas of each block to frequency areas after dividing an input frame into a plurality of block. The rearrangement portion classifies conversion coefficients form the conversion portion according to an important degree of the image information. The zero tree coding portion codes information related to locations and sizes of the rearranged coefficients according to an order of the important degree.

Description

Video encoding device

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an apparatus for encoding video signals, and more particularly, to an encoding apparatus suitable for low rate video transmission for video telephone systems over a public network (PSTN).

As an international standard of conventional video encoding methods, JPEG (Joint Photographic Coding Experts Group), which is a coding / decoding standard for still images, Moving Picture Experts Group (MPEG), which is a coding / decoding standard for video, and H.261, a low-rate video standard. , H.263, etc. are proposed.

In particular, as the need for video communication through a public switched telephone network (PSTN) increases, very low bit video coding studies are actively conducted in terms of coding algorithms and standardization activities. The H.263 Recommendation of the International Telecommunication Union-Telecommunication Standardization Sector (ITU-T), formerly CCITT, is a motion compensated hybrid suitable for very low bit rate transmission for video telephony systems. ) DPCM / DCT (Differential Pulse Code Modulation / Discrete Cosine Transform) coding method is used.

The video encoding method secures a desired bit rate by DCT converting and quantizing an input digital video signal, restoring the quantized video signal, detecting an error with the original video signal, and estimating the motion to control the quantization step.

FIG. 1 is a block diagram illustrating a conventional discrete cosine transform (DCT) based video encoding apparatus. Image encoding is classified into I (Intra) frame encoding and P (Predictive) frame encoding.

At this time, in the case of an I frame, the input video bitstream is output to the DCT unit 101 as it is, and in the case of a P frame, the output of the subtractor 110, that is, the difference between the motion predicted data and the currently input bitstream is DCT. It is output to the unit 101.

The DCT unit 101 removes the correlation of data through two-dimensional axis transformation. To this end, the input frame is divided into block units, and then the respective blocks are divided. That is, the image of each block is converted from the spatial domain to the frequency domain. The DCT data tends to be concentrated in one direction (low pass). Only the collected data is quantized in the quantization unit. In this case, a quantization parameter, that is, a weighting matrix and a quantizer scale code is used for the quantization. Here, the weight matrix represents the weight of each DCT coefficient, and the quantization scale code determines the quantization step.

Each coefficient after quantization is output to an entropy coding unit 103. The entropy encoder 103 applies variable length coding (VLC) to reduce the total number of bits by reducing the number of bits by frequently displaying a small number of bits and a rare value of a large number of bits. Send it through.

In addition, the quantized data is inversely quantized by the inverse quantization unit 105 and IDCT by the IDCT unit 106 and output to the adder 107. The adder 107 adds the motion predicted data and the IDCT data by the motion predictor 109 and stores the IDCT data in the frame memory 108.

On the other hand, since pictures that are continuous on the time axis mainly have a motion of a person or an object in the center portion of the screen, the motion predictor 109 removes the redundancy of the time axis using this property. That is, the amount of data to be transmitted can be greatly reduced by taking and filling a similar portion from the previous picture even if it is not changed or moved.

That is, when the output of the moving prediction unit 109 and the output of the IDCT are added to the adder 107 and stored in the frame memory 108, the frame memory may be estimated when the motion of the picture currently input by the motion predictor 109 is estimated. The data stored in 108 becomes the previous picture.

Finding the most similar block among pictures is called motion estimation, and a motion vector (MV) representing a displacement of how much motion is moved is called motion estimation. The motion vector is transmitted with the VLC transform coefficient information. At this time, the motion vector is also VLC to obtain maximum coding efficiency.

That is, in order to estimate the motion in the motion predictor 109, the motion vector MV must first be obtained. In this case, up to four MVs can be output per macroblock. If this signal is sent, the MV can be transmitted.

The motion compensation process of the motion predictor 109 uses forward and backward prediction blocks, and there are two motion compensation frames. Among them, the P frame is motion compensated only by omnidirectional prediction, and is used to predict the next P frame by itself. The P frame is also used for forward and backward prediction of B frames (frames predicted in both directions). However, B frames are not used for prediction by themselves.

On the other hand, the I frame is a reference screen when compressing and encoding an arbitrary picture, and removes only redundancy in the spatial direction by applying the original signal to DCT transformation and quantization processes for each block.

That is, basically, the first frame is I-frame coded, and if there is a loss of the transmitted packet according to the standard, the receiver sends an I frame when the receiver makes an I frame request from time to time. In this case, since the I frame is used for the motion prediction of the P and B frames, the I frame encoding result greatly influences the coding efficiency of subsequent frames (P and B frames). Moreover, since the background part of the image is hardly encoded in the subsequent P frames until the scene change occurs after the I frame encoding, the result of encoding the I frame continues to improve the subjective quality. Affects. In other words, if the encoding result of the I frame is good, the encoding result of the subsequent P frame may be good.

In the case of P frame encoding, an input frame is a process of encoding two parts, that is, a motion compensated prediction part using high visual correlation between adjacent frames and a DFD (Displaced Frame Difference), which is a prediction error after motion compensation. Divided into

The DFD is an output signal of the subtractor 110, that is, a difference signal between the current frame and the previous frame shifted by the motion vector, and DFD encoding generally occupies most of the P frame generation bits.

In most standards, the DFD encoding method is encoded in the same manner as the I frame image encoding. However, this does not properly exploit the difference between the natural image and the DFD image characteristics. The reason is that DFD images have less spatial correlation, that is, they have much more mid and high frequency components than natural images that mainly contain smoothing areas. Therefore, the energy compression efficiency is lower than that of a natural image having a high energy compression during DCT conversion. As a result, the efficiency of the run-length coding method by conventional zigzag scanning is also reduced.

In addition, since the number of DCT coefficients encoded at a very low bit rate transmission is very small, and each coefficient is also represented by a very large quantization level, block and ringing effects are remarkable in the reconstructed image.

And another issue of video encoding is bit rate adjustment. Since bits are generated by fixed quantization parameters and coded by coefficients, an iterative method is required to meet certain target bit rates, and even then, accurate bit rate adjustment is very difficult.

In addition, I frame coding has a structure of adjusting the bit rate by simply adjusting only the quantization interval (quantization parameter x2). As a result, DCT conversion errors due to large quantization intervals are increased during very low bit rate transmission, and the result continues to affect encoding, including motion estimation and compensation, of P frames, thereby reducing the overall encoding performance. do.

Moreover, since the amount of bits required for I frame encoding when transmitted at a very low bit rate (12-48 kbps) is usually 40% -70% of all generated bits, efficient encoding of I frames is essential for improving overall encoding performance.

Meanwhile, the introduction of Shapiro's concept of Embedded Zerotree Wavelet (EZW) has resulted in an improved zero-tree distortion rate improvement in conventional still image compression over conventional DCT-based image encoders (JPEG). (embedded zerotree) Image coding research is being actively conducted.

The embedded zero-tree coding is a method commonly used for wavelet transformed coefficient coding. In other words, by encoding the position and amplitude information in order of importance using a zero-tree structure of wavelet coefficients, a bit stream, for example, an embedded bit stream, sorted according to importance is generated. Get

This method has not only excellent compression performance, but also a simple algorithm, scalability characteristics that enable various resolutions and image quality, and accurate bit rate adjustment. That is, even if the transmission of the bit stream stops at any point, it is possible to obtain a high quality image at a given bit rate and to have very easy rate control.

The main feature of the embedded zero-tree coding is the continuous estimation quantization, which uses the self-similarity of the wavelet transform to predict the position of the inter-band importance coefficients and sequentially approximates the magnitude of the wavelet coefficients. Successive Approximation Quantization (SAQ).

The approximate method is as follows.

The input image is decomposed into subbands of various resolutions using wavelet transform.

At this time, the low frequency components of the original image are gathered in the coarse band and the detail high frequency components are gathered in the other band. And all coefficients in a given band except for the highest frequency band are related to coefficients in similar directions of the next subband.

Thus, a coherent set of coefficients in the coarse band at the same location in the same direction in a similar direction is called the children.

At this time, a parent node in the lowest frequency band has three children in different directions.

That is, the EZW creates a data structure called zero tree created in the parent-child relationship described above.

The zero tree structure uses the property that if the wavelet coefficients in the sparse band are smaller than any given threshold, then its children are also likely to be small. This zero-tree structure is very similar to the concepts of zigzag scanning and end of block (EOB) commonly used to encode DCT coefficients.

For example, the EZW scans coefficients by band.

That is, the parents are scanned after all neighboring parents of the same band are scanned.

Each coefficient is then compared against the current threshold. At this time, if the absolute value of the coefficient is larger than the threshold, it is encoded with one of negative or positive significance symbols.

In addition, the zero tree root symbol is used to encode parents whose children of the zero tree structure have sub-threshold values.

An isolated zero symbol encodes a coefficient whose at least one children is above a threshold.

In this case, the EZW further encodes coefficients determined to be important information by using continuous estimation quantization (SAQ). The continuous estimation technique for quantization of the wavelet coefficients produces an embedded bitstream arranged in order of significant bit.

In the EZW, since the wavelets have frequency and spatial information, spatial grouping and quantization of data are possible.

In addition, the improvement of the efficiency of the continuous estimation algorithm for the band in the same direction is a key to efficiently predict the zero and nonzero values.

Among many wavelet-based image encoders, Said and Pearman's Set Partitioning in Hierarchical Trees (SPIHT) encoder using the improved zero-tree method is an embedded image encoder with excellent compression performance.

The SPIHT encoding method efficiently removes the redundancy between coefficients by removing a tree node that is already encoded and updating a critical node with respect to a threshold value that decreases by half.

The main reason for this method's performance improvement over EZW is due to the improved zero-tree structure using the property that important coefficients are mainly distributed in the lowest band.

However, these methods are very efficient for still image compression coding using mainly high resolution (512 × 512) images, but filter characteristics and insufficient band resolution in low bit rate video compression using mainly low resolution (176 × 144) such as QCIF. That is, the encoding performance is remarkably degraded due to the wavelet spatial frequency characteristic degradation (zero tree coding efficiency is lowered).

In addition, when EZW is applied to DFD encoding, the decomposition of the pyramid structure based on the existing wavelets decreases the energy compression efficiency to the upper level and does not properly reflect the intermediate and high frequency components, resulting in lower overall coding efficiency than the DCT based method. Shows.

Moreover, compatibility with existing video codecs is also a major problem given that DCT transform is used to encode errors (i.e., DFDs) with I frames and motion compensated images in most video compression standards.

The present invention is to solve the above problems, an object of the present invention is to rearrange the DCT coefficients in a two-level pyramid structure and then zero-tree coding, thereby improving the efficiency of zero-tree coding and compatibility with existing DCT-based encoders The present invention provides an image encoding apparatus that maintains.

1 is a block diagram of a conventional DCT-based video encoding apparatus

2 is a block diagram illustrating a video encoding apparatus according to the present invention.

3 is a diagram for explaining a relationship between a DCT and a wavelet.

FIG. 4 is a diagram illustrating a relationship in which DCT coefficients are rearranged in a two-level pyramid structure in FIG. 2. FIG.

FIG. 5 shows a parent-child relationship in the rearrangement of DCT coefficients to a two-level pyramid structure in FIG.

6 shows an example of performance evaluation between the present invention and the prior art;

Explanation of symbols for main parts of the drawings

101: DCT unit 102: quantization unit

103: entropy encoder 104: channel

105: dequantization unit 106: IDCT unit

107: adder 108: frame memory

109: motion prediction unit 110: subtractor

111: bit rate control unit 201: relocation unit

202: zero tree encoding unit 203: reverse rearrangement unit

According to an aspect of the present invention, a video encoding apparatus includes a transform unit that divides an input frame into a plurality of blocks, and then converts a spatial domain of each block into a frequency domain, and a degree of conversion of the transform coefficients of the transformer to information necessary for image reproduction. And a zero-tree encoder for relocating and classifying the order of importance according to the order of importance, and outputting a bit stream arranged according to the importance by encoding the position and size information of the rearranged coefficients in order of importance.

In the case of an I frame, the converting unit converts input data into a frequency domain by DCT.

In the case of a P frame, the transform unit performs motion compensation prediction using a previous I or P frame, and then converts a difference between the motion compensated data and the currently input data into a frequency domain by DCT.

The converting unit encodes each block of the P frame in an intra or inter mode.

The rearrangement unit classifies the DCT coefficients of the transform unit using wavelet analysis and rearranges the DCT coefficients into a two-level pyramid structure.

The two-level pyramid structure of the rearrangement unit is a structure in which the four bands decomposed by the uniform band decomposition method are independently wavelet transformed.

The rearrangement unit decomposes the DC coefficient of the DCT coefficient and the AC coefficient value according to the frequency, and classifies and rearranges the lowest frequency band only in a rearranged form from the parent-child relationship defined in order to use the spatial correlation of DC. It is done.

The rearrangement unit biases the DC value to quantize the rearranged DCT coefficients.

The relocator may bias the DC value by the DC average by subtracting the average value of the DC coefficients from the DC coefficient with respect to the I frame.

The relocator may subtract the average value of the DCT coefficients of the intra blocks per frame from the DC coefficient with respect to the P frame to bias the DC coefficients with the DC average.

The relocation unit is characterized in that the DC average value is set to 0 for the inter mode block of the P frame so as not to bias the DC value.

The zero-tree coding unit encodes the position and the sign of the significant coefficients by comparing each of the rearranged coefficients with a series of thresholds, and performs the initial threshold with a maximum factor of 2 smaller than the maximum coefficient value and decreases it by half. It is done.

The zero tree encoder is characterized in that it consists of two stages: the classification path for encoding the location and code of the important information, and the fine division path required for continuously estimating the critical coefficient.

The output bit stream of the zero tree encoder is characterized by consisting of a significance test result, a sign, and finely classified bits.

The zero tree encoder is characterized by gradually transmitting the most important information by first transmitting an approximation of the most important coefficients and classifying all the important coefficient values in detail one bit at a time.

The zero tree encoder may further include an entropy encoder configured to perform lossless entropy encoding on the significance test result, the code, and the detailed classification symbol generated in the zero tree encoding process by an adaptive arithmetic encoding method.

The output of the zero tree encoder is rearranged to have a plurality of uniform decomposition bands for motion compensation and prediction.

Other objects, features and advantages of the present invention will become apparent from the following detailed description of embodiments taken in conjunction with the accompanying drawings.

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

The present invention aims to increase the efficiency of zero-tree coding by rearranging DCT coefficients into a two-level zero-tree structure by using wavelet analysis that 8x8 DCT is a wavelet transform having 64 uniform decomposition bands.

Simulation results show that the video encoding apparatus according to the present invention has an efficient zero-tree structure and has superior performance in terms of image quality than the H.263 I-frame encoding method, which is a low bit rate video compression standard. It shows higher rate distortion improvement than EZW method using wavelet of.

Still image compression also shows bit rate distortion improvements comparable to those of JPEG and other representative wavelet embedded image encoders.

FIG. 2 is a block diagram illustrating a video encoding apparatus according to the present invention. Since the remaining blocks except for the quantizer 200 are the same as in FIG. 1, the same reference numerals are used for the same block and detailed description thereof will be omitted.

That is, the present invention adopts an embedded zero-tree quantizer 200 that quantizes each coefficient gradually by a predetermined bit, deviating from the conventional approach of encoding by coefficient.

In this case, each frame of the video input to the image encoder as shown in FIG. 2 is intra for the first frame I, and the remaining frames P are encoded with Inter.

The encoding is performed in macroblock units, and the brightness and color components are in a 4: 2: 0 format.

That is, in the case of an I frame, after dividing the brightness (Y) and the color (Cb, Cr) components of the first input image into 8 × 8 blocks, each block is input to the DCT unit 101. The DCT unit 101 removes spatial redundancy by converting an image of each block from the spatial domain to the frequency domain by DCT.

The coefficients of the blocks DCT in the DCT unit 101 are input to the rearrangement unit 201 of the quantizer 200 and rearranged into a two-level pyramid structure. That is, the DCT coefficients transformed by the DCT unit 101 into the frequency domain are sorted in order of high importance and rearranged into a two-level pyramid structure.

The DCT coefficients relocated to the two-level pyramid structure in the repositioner 201 are quantized by being coded into an embedded zerotree structure by the zerotree coding unit 202. In this case, the SPIHT coding method may be used.

Each quantized coefficient is entropy coded by the entropy encoder 103 and then transmitted through the channel 104.

In addition, the motion estimation and compensation method used in H.263 + is used for all consecutive frames of the video sequence, for example, P frames. Accordingly, the image encoder of the present invention may selectively use block motion estimation, ANNEX D (Unrestricted Motion Vector mode), or ANNEX F (Advanced Prediction mode).

After the blocks are motion predicted, residuals errors are also DCTed by the DCT unit 101 and then rearranged by the repositioner 201 to the remaining frames of the two-level pyramid structure for quantization.

At this time, when an overlapping or exposed area or a scene change occurs due to a very fast movement or movement of an object, prediction fails for such a macro block. In such a case, it is preferable to encode a block of the original image rather than to encode DFD after motion estimation in terms of coding efficiency and subjective quality.

The present invention selects and encodes such blocks in intra mode.

In this case, the determination of the intra / inter mode is shown in Equation 1 below, and is similar to the mode decision of the H.263 encoder.

If A <(SAD (x, y)-T), the intra mode is selected and no motion estimation is done. However, T is a given threshold and may vary according to the motion state of each block, and the intra / inter mode is transmitted as additional information such as H.263. Here, MB _mean is an average value of the macro block, A is a difference between each pixel and the average value for the macro block, that is, a deviation of each pixel, and SAD is a difference value at the same position of the previous frame.

Here, the frequency spectrum characteristic after DCT conversion of a block selected as an intra mode has a lot of differences from other inter mode blocks. In particular, the magnitude of the DC coefficient of the block selected in the intra mode is large.

Therefore, in the present invention, by subtracting the average value of the DCT coefficients of the intra blocks per frame from the DC coefficients, the DC coefficients are biased to the DC average to reduce the number of bits that are wasted for unnecessary scanning due to the large DC coefficients.

Of course, the interblock is relocated without bias.

The DCT coefficients of the two-level pyramid structure constructed after such processing are quantized by the embedded zerotree coding in the zerotree coding unit 202 as performed in the I frame coding.

In this case, the purpose of relocating the DCT coefficients to the two-level pyramid structure in the repositioning unit 201 is to make DCT well coupled with the embedded zero-tree coding, thereby increasing the interdependence between the spatial and frequency of the coefficients, and thus the embedded zero tree. It is to improve the coding efficiency.

Figure 3 is a simple diagram that can intuitively know the relationship between the DCT and the wavelet. For example, considering the 2 × 2 block DCT, each block after the conversion may have a degree of difference, as shown in the left figure of Figure 3 a (the most Low frequency) to d (highest frequency).

This is achieved by decomposing an image using a suitable 2x2 decomposition filter for the same image and subsampling by 2 in each direction, and the band-decomposed image is equivalent to the transformed block. That is, as shown in the right figure of FIG. 3, the highest band (low-pass / low-pass) is as if the lowest frequency components of each DCT conversion block are collected. The other conversion coefficients are collected in a similar way for the remaining bands.

Extending this concept, an 8x8 block DCT can be seen as a wavelet transform with 64 resolution bands.

4 shows a process of rearranging DCT coefficients into a two-level pyramid structure, where 401 indicates a DCT coefficient and 402 indicates a rearrangement thereof into a two-level pyramid structure.

That is, the repositioner 201 regards each DCT 8 × 8 block as a three-level wavelet pyramid structure defined in the EZW. In order to explain the parent-child relationship indicating the spatial correlation between the coefficients, the coefficients are numbered as shown in FIG. 4.

At this time, if i is a coefficient from 1 to 63, the parent of the coefficient i is an integer value of i / 4, whereas if j is a coefficient from 1 to 15, the children of the coefficient j are {4j, 4j + 1,4j + 2,4j +3}.

And, DC coefficient 0 is just the root of the tree with three children 1,2,3 children.

Here, important information in most images is included in the DC coefficient and the first few AC coefficients, so the DC coefficients and children 1,2,3 of each DCT block are mapped to the uppermost band of the memory of the image size and each coefficient Children 4,5,6,7 for each parent of 1,2,3; 8,9,10,11; 12, 13, 14, and 15; respectively, correspond to the next band, and children of parents in the next band correspond to the next band, respectively.

Figure 5 shows the parent-child relationship of the structure according to the present invention (501-502-502 '-503-503').

The difference between the structure of the present invention and the two-level pyramid structure by the conventional wavelet is the frequency characteristic by the decomposition method.

In other words, the conventional wavelet two-level pyramid structure is a hierarchical band decomposition method, and the wavelet transform is applied to only one low frequency band, but the two-level pyramid structure according to the present invention has four bands decomposed by the uniform band decomposition method. The wavelet transformed structure is considered to be a 2-level pyramid structure.

In addition, the relocation structure of the present invention is simple, and the difference from the uniform band decomposition method is that only the lowest frequency band is rearranged again in a parent-child relationship defined to use the spatial correlation of DC.

Therefore, while the conventional wavelet two-level pyramid structure reflects low frequency components well but does not reflect middle frequency components well, the structure of the present invention is a reflection characteristic of low frequency components compared to the conventional wavelet two-level structure. Is somewhat lower but reflects the intermediate frequency components well.

This feature is advantageous over hierarchical decomposition in low bit rate transmissions. The reason for this is that coding efficiency that can not be encoded from low bit rate to high frequency component without losing the decaying spectrum property that determines the efficiency of zero tree coding can be improved by reflecting the intermediate frequency component well. Because.

On the other hand, the quantization of the DCT coefficients rearranged in the two-level pyramid structure is performed by encoding the zero tree structure in the improved method of the EZW, that is, the zero tree encoder 202.

That is, the zero tree encoder 202 first encodes coefficients having the most visually significant influence by encoding the highest band. In the case of an I frame, by subtracting the average value of the DC coefficients of each block, the DC coefficients are biased to the average of the DC coefficients, thereby improving coding efficiency.

First, the average value of the DC components is subtracted with respect to the DC component (each node) of the uppermost band. Since the DC values of neighboring neighboring blocks are highly spatially correlated, by doing so, each DC value is biased to an average value to some extent, thereby reducing efficiency of unnecessary bits during zero-tree coding and increasing efficiency.

However, in the case of the P frame, since the spatial correlation between neighboring neighboring blocks DC is not large, the DC average value is set to 0 so that the above process is not performed.

Each coefficient encodes the position and sign of a significant coefficient compared to a series of thresholds. The initial threshold level is a maximum factor of 2 less than the maximum coefficient, and is reduced by half. Perform as you go.

In this case, coefficients larger than the threshold are classified as important coefficients, and coefficients smaller than the threshold are classified as not important.

In other words, three lists of LIS (list of significant sets), LIP (list of insignificant pixels) and LSP (list of significant pixels) are used.

The LIP is the root of the highest layer, and the LIS is initialized with a set of children of each parent of the LIP.

The encoding consists of two stages: a sorting pass for encoding the location and code of the important information and a refinement pass necessary for continuously estimating the critical coefficient.

In the classification path, first, LIP pixels are compared with a current threshold, and a sign is output when the significant factor is significant, and then moved to the LSP.

Next, we can examine the coefficients of the LIS and represent them with only 1 bit if they are all insignificant, and if there is a significant pixel, the subset with the roots of the descendants of the current root. After dividing by (subset), repeat the previous process.

After one sorting pass is completed for the LIS and the LIP, a refinement pass is performed for the LSP, and each coefficient value is classified by one bit.

At the end of this entire path, the threshold is divided by 2 and the next classification path is passed.

The output bit stream is composed of the significance test result, the sign, and the finely divided bits. Since the encoder and the decoder share the same algorithm, all the determination results during the encoding process are output. It does not need to be sent.

This method results in a progressive transmission that always selects the most important information by first sending an approximation of the most important coefficients and then finely classifying all the important coefficient values one bit at a time. Likewise, if the unitary and geometric rules (euclidean norm) are preserved, this gradual transmission is the best way to reduce mean square error (MSE).

In the meantime, the entropy encoding unit 103 performs lossless entropy encoding on the significance test result, the code, and the detailed classification symbol generated in the embedded zero tree encoding process of the zero tree encoder 202 by an adaptive arithmetic encoding method. do.

Since this method implicitly transmits the model to the decoder, it does not require any prior information on the image and can utilize the statistical dependence between important coefficients very efficiently.

In the present invention, an adaptive arithmetic coding algorithm such as Witten is used.

In addition, since the present invention is an embedded encoder, the bit rate adjustment of the bit rate controller 111 may be any bit rate adjustment algorithm based on the existing bit rate distortion at the output terminal, and the bit rate may be easily adjusted.

Meanwhile, the output of the zero tree encoder 202 must be dequantized for motion prediction. To this end, if the output of the zero tree encoder 202 is rearranged in the state before the input to the rearrangement unit 201, that is, the rearrangement unit 203 has a plurality of uniform decomposition bands, the result of dequantization is performed. Get

Then, the output of the reverse rearrangement unit 203 is IDCT in the IDCT unit 106 and output to the adder 107. Subsequent operations are the same as in FIG.

6 is a graph comparing the performance of the conventional encoder and the encoder according to the present invention, wherein the encoder of the present invention has superior bit rate distortion performance of I frame encoding than H.263 + image encoder having H.263 + and ANNEX I. It can be seen that the improvement of. Therefore, excellent coding efficiency can be obtained for areas that are exposed or overlapped in the scene due to scene change and rapid movement of an object.

As described above, the algorithm of the present invention encodes the quantization coefficients in the order of MSB to LSB by rearranging the DCT, and in-band (spatial correlation of DC coefficients) and inter-band spatial dependence (dequering spectral characteristics from low frequency to high frequency of one block). The coding efficiency of the embedded zero tree can be increased.

This method shares some characteristics with bitplane image coding, but by using a zero-tree structure with a single coded symbol in which bits are not transmitted line by line, the decoder is able to The difference is that it infers that almost every bit is zero.

Meanwhile, the image decoder which decodes the encoded and transmitted image uses the same algorithm as the image encoder. Image decoding according to the present invention is performed by performing the same process as image encoding.

According to the image encoder according to the present invention, DCT transform of picture data of 8 × 8 block unit and transform the DCT coefficient into wavelet structure to obtain the result of sorting in order of information of high importance, and then by transmitting embedded zero-tree encoding This has the following advantages:

First, it is suitable for low-rate video transmission for video telephony systems through the public network (PSTN) due to improved compatibility with existing encoders and excellent bit-distortion performance.

Second, the zero-tree coding used in the present invention has the advantage that the scalable characteristics and the bit rate control are very easy. Thus, when there is usually no bit budget available or when the buffer crosses a given threshold, the bitstream can be truncated at any time. This feature is particularly useful for browsing in image databases.

Third, due to the excellent bit rate distortion performance of I frame coding, excellent coding efficiency can be obtained for areas that are exposed or overlapped in a scene due to scene change and rapid movement of an object. In particular, in the case of natural video input directly to a camera rather than a test video of MPEG-4 or H.263, an intra macroblock is frequently generated, and thus, greater coding efficiency can be obtained. In other words, the efficient coding of the I frame is very important for the performance of the low bit rate video encoder, so that the overall coding efficiency is improved when the encoder of the present invention is actually applied to the picture encoder.

Fourth, the block distortion phenomenon is reduced. This is the result of DC coefficients being preferentially coded and intermediate and old frequency component reflections.

Fifth, it is suitable for MPEG-4, H.263 ++, and H.26L requirements, which are currently being standardized, and are applicable in practice.

Claims (19)

In the video encoding apparatus that decomposes and encodes one frame into a plurality of blocks, Converting means for dividing an input frame into a plurality of blocks and converting a spatial domain of each block into a frequency domain; Repositioning means for classifying and rearranging the transform coefficients of said converting means in order of importance according to the degree of containing information necessary for image reproduction; And zero-tree encoding means for encoding the position and size information of the rearranged coefficients in order of importance and outputting bit streams arranged in accordance with the importance. The method of claim 1, wherein the conversion means is I frame An image encoding device, characterized by converting input data into a frequency domain by discrete cosine transform (DCT). The method of claim 1, wherein the converting means is a P frame And performing a motion compensation prediction using previous I and P frames, and then converting the difference between the motion compensated data and the currently input data into a frequency domain by using a discrete cosine transform (DCT). The method of claim 3, wherein the converting means is adapted to encode a P frame. And an intra mode / inter mode of each block according to the following equation. MB _mean is an average value of the macro block, and A is a difference between each pixel and the average value of the macro block. The method of claim 4, wherein the conversion means And A <(SAD (x, y)-T) to determine the intra mode in which motion estimation is not performed. Where SAD is the difference at the same location of the previous frame and T is the given threshold. The method according to claim 5, wherein the threshold T of the converting means is The image encoding apparatus may be variable according to the motion state of the screen. The method of claim 1, wherein the repositioning means, And classifying the DCT coefficients of the transform means using wavelet analysis and rearranging the DCT coefficients into a two-level pyramid structure. 8. The structure of claim 7, wherein the two-level pyramid structure of the repositioning means is And a wavelet transform of four bands separated by a uniform band decomposition method. The method of claim 7, wherein the repositioning means, Image coding characterized by reclassifying the DC coefficients of DCT coefficients and AC coefficient values according to frequency, and classifying and rearranging only the lowest frequency bands in the form of rearrangement only from the parent-child relationship defined in order to use the spatial correlation of DC. Device. The method of claim 1, wherein the repositioning means, And biasing the DC value for quantization of the rearranged DCT coefficients. The method of claim 11, wherein the repositioning means And an average value of the DC coefficients is subtracted from the DC coefficients for the I frame to bias the DC value by the DC average. The method of claim 11, wherein the repositioning means In the case of the P frame, the image encoding apparatus characterized in that the DC coefficients are biased by the DC average by subtracting the average value of the DCT coefficients of the intra blocks per frame from the DC coefficients. The method of claim 11, wherein the repositioning means And a DC average value of 0 for an inter mode block of a P frame to bias the DC value. The method of claim 1, wherein the zero tree encoding means, Each of the rearranged coefficients encodes the position and the sign of the significant coefficients in comparison with a series of thresholds, and the initial threshold value is a maximum factor of 2 smaller than the maximum coefficient value, and is performed by decreasing it by half. . 15. The method of claim 14, wherein the zero tree encoding means And a classification path for encoding the location and code of the important information and a fine division path necessary for continuously estimating the critical coefficient. 15. The method of claim 14, wherein the output bit stream of the zero tree encoding means A video encoding apparatus comprising a test result of importance, a sign, and bits classified in detail. 15. The method of claim 14, wherein the zero tree encoding means And transmitting the most important information gradually by first transmitting an approximation of the most important coefficients and classifying all the important coefficient values one bit at a time. The method of claim 1, wherein the zero tree encoding means, And an entropy encoding unit for lossless entropy encoding by the adaptive arithmetic encoding method. The method of claim 1, wherein the output of the zero tree encoding means And repositioned to have a plurality of uniform decomposition bands for motion compensation and prediction.