JP2003219430A - Method and closed-loop transcoder for reduced spatial resolution transcoding of compressed bitstream of sequence of frames of video signal - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a method and closed-loop transcoder for reducing the spatial resolution of a compressed bitstream of a sequence of frames of a video signal by first decoding the frames and storing the decoded frames in a first frame buffer. <P>SOLUTION: While performing decoding, motion compensating is performed with full resolution motion vectors of the stored decoded frames. The decoded frames are then down-sampled to reduced resolution, and stored in a second frame buffer. The reduced resolution frames are partially encoded to produce a reduced resolution compressed bitstream of the video. While performing the partial encoding, motion compensation is performed with reduced resolution motion vectors of the stored reduced resolution frames. <P>COPYRIGHT: (C)2003,JPO

Description

Detailed Description of the Invention

[0001]

TECHNICAL FIELD OF THE INVENTION [Cross-reference of related patent applications]
This patent application was assigned to the same assignee as this patent application,
“Video Transcoder with” filed on May 11, 2001
This is a continuation-in-part application of US patent application Ser. No. 09 / 853,394 entitled "Spatial Resolution Reduction". The present invention relates generally to the field of transcoding bitstreams, and more particularly to reducing spatial resolution including drift compensation. Accordingly, the present invention relates to methods and video transcoders that include spatial resolution reduction and drift compensation. BACKGROUND OF THE INVENTION Video compression enables storage, transmission and processing of image information with low storage, network and processor resources. The most widely used video compression standards are MPEG-1 for video storage and retrieval, MPEG-2 for digital television and H.263 for video conferencing. For these, see ISO / IEC 111
72-2: 1993 `` Information Technology-Coding of Mo
ving Pictures and Associated Audio for Digital Sto
rage Media up to about 1.5 Mbit / s-Part 2: Vide
o ", D. Le Gall's" MPEG: A Video Compression Standa "
rd for Multimedia Applications, Communications o
f the ACM, Vol. 34, No. 4, pp. 46-58, 1991, ISO / IE
C 13818-2: 1996, `` Information Technology-Generi
c Coding of Moving Pictures and Associated Audio In
formation-Part 2: Video '', 1994, ITU-T SG XV, DR
AFT H. 263, `` Video Coding for Low Bitrate Communi
cation '' 1996, ITU-T SG XVI, DRAFT13 H. 263 + Q15-A
-60 rev. 0, `` Video Coding for Low Bitrate Communic
ation ”1997.

The above-mentioned standards are relatively low-level specifications that mainly deal with spatial compression of images or frames and spatial and temporal compression of frame sequences. As a common feature, these standards provide frame-based compression. According to these standards, high compression rates can be achieved for a wide range of applications.

MPEG-4 for multimedia applications
New video coding standards such as ISO / IEC 14496-2: 1
999, `` Information technology-coding of audio / visu
al objects, Part 2: Visual ”), it is possible to encode or decode objects of arbitrary shape as separate video object planes (VOPs). In this case, the object may be a visual object, an audio object, a natural object, an artificial object, a primitive object, a mixed object, or an object composed of a combination thereof. Also, a considerable amount of error-tolerant measures are incorporated to enable robust (ie, robust) transmission over error-prone channels such as wireless channels.

The emerging MPEG-4 standard is for multimedia applications such as interactive video, namely
It is intended to allow multimedia applications where natural and artificial materials are integrated and access is universal (non-unilateral). In connection with video transmission, compression standards are needed to reduce or reduce the amount of bandwidth on the network. In this case, the network may be wireless or the Internet. In any case, the capacity of the network is limited, so contention for low resources should be minimized.

Much effort has been expended on systems and methods that enable devices to robustly transmit content, ie, consistently and adapt the quality of content to available network resources. In this regard, when encoding or encoding content, it is necessary to first decode (decode) the bitstream so that the bitstream can be transmitted through the network at a low bit rate or a low resolution. There are cases.

This can be accomplished by a transcoder 100, as shown in FIG. In the simplest configuration, the transform encoder 100
Has a decoder (decoder) 110 and an encoder (encoder) 120 connected in cascade. The compressed input bitstream 101 is the input bitrate

[Equation 1] Fully decoded, then re-encoded at bit rate

[Equation 2] Output bitstream 103 is generated. The output bit rate is usually smaller than the input bit rate. However, in a practical example, the re-encoding of the decoded bitstream is so complicated that the transcoder does not perform full decoding and full re-encoding.

As an early research on the transcoding of the MPEG-2 standard, a paper "Architectures for MPEG compresse" published by Sun et al.
d bitstream scaling '', IEEE Transactions on Circui
ts and Systems for Video Technology (April 1996)
There is a month. This paper describes four rate reduction methods due to complexity and architectural changes.

FIG. 2 illustrates a first example method 200 referred to as an open loop architecture. In this architecture, the input bitstream 201 is only partially decoded. Specifically, macroblocks in the input bitstream are variable length decoded (VLD).
210 and fine quantizer

[Equation 3] Subject to inverse quantization 220 at which the discrete cosine transform (DCT) coefficients are generated. For a given desired output bit rate 202, the DCT block is the quantizer 2
30 coarse level quantization

[Equation 4] Is requantized by. These requantized blocks are
Variable length coding (VLC) is then performed, as shown at 240, resulting in the formation of a new output bitstream 203 at a low bitrate. This scheme is considerably simpler than the scheme shown in FIG. The reason is that the motion vector is reused and the inverse DCT operation is unnecessary. here,

[Equation 5] as well as

[Equation 6] The exact choice depends on the bitrate characteristics of the bitstream. Other possible factors, such as the spatial characteristics of the bitstream, are not taken into account.

FIG. 3 illustrates a second example method 300. This method is called a closed loop architecture. In this method, the input video bitstream is partially decoded again, that is, the macroblocks of the input bitstream are variable length decoded (VLD) 310 as well as the quantizer.

[Equation 7] Subject to dequantization 320, which produces discrete cosine transform (DCT) coefficients 321. In contrast to the first example method described above, the correction DCT coefficient 332 is added to the input DCT coefficient 321 (see 330), thereby compensating for the mismatch caused by the requantization. . This correction improves the quality of the reference frame that is ultimately used for decoding or decoding, that is, the quality. After correction, the newly formed block will have the new bit rate.

[Equation 8] Is re-quantized (340) and is variable-length coded (350) as described above. Also in this case,

[Equation 9] as well as

[Equation 10] Note that is based on the bit rate.

To obtain the correction component 332, the requantized DCT coefficients are dequantized (36
0) and subtracted (370) from the original partially decoded DCT coefficient. The difference resulting from this subtraction is the inverse DC
It is converted to the spatial domain via T (IDCT) 365 and stored in the frame memory 380. Here, the motion vector 381 associated with each input block is
90) is used to reread the corresponding difference block. Therefore, the corresponding block is DCT3.
Transformed via 32, which produces a correction component. Assuncao is a derivative example of the method shown in FIG.
Outside paper `` A frequency domain video transcoder for
dynamic bit-rate reduction of MPEG-2 bitstream
s '', IEEE Transaction on Circuits and System for V
ideo Technology, pp. 953-957, 1998.

That is, Assuncao et al. Also propose another method for the same task. This alternative method uses a motion compensation (MC) loop operating in the frequency domain for the purpose of drift compensation. An approximate matrix is derived for performing high speed calculation of the MC block in the frequency domain. Transcoding
On the other hand, Lagrange's optimization is adopted to calculate the best quantizer scale. According to this method, I
The need for DCT / DCT components is eliminated.

According to the conventional compression standard, the number of bits allocated for encoding texture information is controlled by a quantization parameter (QP). The method described above uses a Q based on the information contained in the original bitstream.
There is similarity in that the texture bit rate is reduced by changing P, the quantization parameter.
For efficient implementation, the information is usually extracted directly from the compressed domain to determine the macroblock motion or D
A measure of the residual energy of the CT block can be included. The above method is applicable only for bit rate reduction.

In addition to bit rate reduction, other formats of bit streams can be converted. For example, the object-based conversion is 2000 outside of Vetro.
US patent application 09 / 504,323 dated February 14, 2013
Specification (Title of Invention: Object-Based Bitstream Trans
coder). In addition, the conversion related to spatial resolution is described in Shanableh and Ghanbari's paper "Heterogeneous.
video transcoding to lower spatio-temporal resolu
, and different encoding formats ", IEEE Trans
It is described in action on Multimedia (June 2000).

In these methods, a bitstream is generated with a low spatial resolution which is unsatisfactory in terms of quality, and complexity is increased if quality is attempted. Also, no proper consideration is given to the means of forming the reconstructed macroblock. This not only has a great impact on both quality and complexity, but is especially problematic when considering reduction factors different from two. Furthermore, these methods are not accompanied by specific architectural details. Much of the interest has been devoted to various means of scaling motion vectors by the factor "2".

FIG. 4 illustrates transform coding (tr) an input bitstream into an output bitstream 402 at a low spatial resolution.
Details of a method 400 for anscoding). This method corresponds to an extension of the method shown in FIG. 1, but shows details of the decoder 110 and the encoder 120 and has a downsampling block 410 between the decoding and encoding processes. Decoder 110
Performs partial decoding of the bitstream. The downsampling block 410 reduces the spatial resolution of groups that partially contain macroblocks. In motion compensation 420 at the decoder, motion vectors of full resolution

[Equation 11] 421 is used, while the motion compensation 430 in the encoder uses a low resolution motion vector.

[Equation 12] 431 is used. Low resolution motion vectors are downsampled spatial domain frames

[Equation 13] Estimated from 403 or mapped from full resolution motion vectors. Transform encoder (transcoder) 4
Details of 00 will be described below.

FIG. 5 shows details of an open loop method 500 for transcoding an input bitstream 501 into an output bitstream 502 at low spatial resolution. In this method, the bit stream 10
The 1 is also partially decoded. That is, the macroblock of the input bitstream is variable length decoded (VLD) (51
0) and dequantized (520), which produces discrete cosine transform (DCT) coefficients. Note that these processing steps are well known.

The DCT macroblock is then 16 ×
Each 8 × 8 (2 ³ × 2 in 16 (2 ⁴ × 2 ⁴ ) macroblocks
³ ) Downsampling with a factor of "2" by masking the high frequency coefficients of the luminance block (53
0), thereby generating four 4 × 4 DCT blocks. In this regard, Ng US Pat. No. 5,262,854 entitled “Low-resoluti
on HDTV receivers ". In other words, by downsampling, for example, a block group of four blocks is converted into a group of four small blocks, that is, a block group.

By performing downsampling in the transcoder, the transcoder must take an additional step to reshape the subordinate 16 × 16 macroblocks. That is, an inverse transform into the spatial domain followed by a retransform into the DCT domain. After downsampling, the block is requantized (540) with the same quantization level and then variable length coded (550). It should be noted that there is no description of a method relating to the implementation of the bit rate control for the reduced resolution block.

In order to perform the motion vector mapping 560 from the total motion vector 559 to the reduced motion vector 561, several suitable methods have been proposed in the past for frame-based motion vectors. A simple averaging or median is used to map the four frame-based motion vectors, one for each macroblock in a group, to one motion vector for the newly formed 16x16 macroblock. -A filter can be applied. This is called 4: 1 mapping.

However, MPEG-4 and H.26
Some compression standards, such as 3, support advanced prediction modes that allow one motion vector per 8x8 block. In this case, each motion vector is 1 at the original resolution.
The 6x16 macroblock is mapped to an 8x8 block with a reduced resolution macroblock. this is,
It is called 1: 1 mapping.

FIG. 6 maps (600) motion vectors from four 16 × 16 macroblock groups 601 to either one 16 × 16 macroblock 602 or four 8 × 8 macroblock groups 603. An example is shown. Always use 1: 1 mapping is 4
Inefficient because many bits are used to encode one motion vector. Also, in general, extensions to field-based motion vectors for interlaced images are not nonsensical. As is well known, for downsampled DCT coefficients and mapped motion vectors, the data can be variable length encoded to form a reduced resolution bitstream.

It would be desirable to provide a transform coding method for a bitstream that overcomes the problems of the conventional methods for reducing spatial resolution. Furthermore, it is desirable to compensate for drift and to provide a balance between complexity and quality in transcoders.

[0023]

SUMMARY OF THE INVENTION A method and system compresses a sequence of frames of a video signal by first decoding the frame and storing the decoded frame in a first frame buffer. Reduce the spatial resolution of the bitstream. While performing the decoding, motion compensation is performed using the maximum resolution motion vector of the stored decoded frame. The decoded frame is then downsampled to the reduced resolution and stored in the second frame buffer. The reduced resolution frame is partially encoded to produce a reduced resolution compressed bitstream of the video. Motion compensation is performed using the reduced resolution motion vectors of the stored reduced resolution frames while performing the partial encoding.

[0024]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Introduction The present invention transcodes a compressed bitstream of a digital video signal to a spatial resolution reduced with minimal drift, ie, transform coding. System and method. First, some application examples relating to the distribution of content in which the transcoder or transcoder according to the present invention can be used will be described. Next, a basic method for generating a bitstream with a low spatial resolution will be analytically described. Based on this analysis, some embodiments for the basic method and the corresponding architecture associated with each embodiment are described.

The first embodiment (see FIG. 9) employs an open loop architecture, while the other three embodiments (see FIGS. 10 and 11A-11B) use downsampling and requantization. And a closed-loop architecture that constitutes means for compensating for drift caused by motion vector truncation. It should be noted that one of these closed-loop architectures performs the compensation at reduced resolution, while the other two closed-loop architectures perform the compensation at the original resolution in the DCT domain to ensure good quality.

As will be explained in more detail below, the open loop architecture of FIG. 9 is less complex. If there is no restoration loop, there is no DCT / IDCT block, and there is no frame memory, so
The quality is commensurate with the low image resolution and bit rate. This architecture is suitable for Internet use and software implementation. The first closed-loop architecture shown in Figure 10 is also of medium complexity. This first closed loop architecture comprises a reconstruction loop, an IDCT / DCT block and a frame memory. This architecture can compensate for drift and improve quality in the reduced resolution domain. The complexity of the second closed loop architecture shown in FIG. 11A is also moderate. This second architecture comprises a decompression loop, an IDCT / DCT block and a frame memory. According to this architecture
Although it is possible to compensate for drift and improve quality in the original resolution domain, up-sampling of reduced resolution frames is required. In the third closed loop architecture, the correction signal obtained in the reduced resolution domain is used.

In order to gain a deeper understanding of the architecture according to the invention, a "mixed mode" with reduced resolution is provided.
Mode)) macroblocks are also described for some additional techniques.

A group of blocks to be down-sampling, eg, four blocks, were coded in both intra-mode and inter-mode. When the block is included, it is called "empress block (mixed block)". In the MPEG standard, I-frames contain only macroblocks coded according to intra mode, but P-frames may contain blocks coded in intramode and intermode. These modes need to be taken into account especially during downsampling. This is because there is a possibility that the quality of the output, that is, the image quality may be deteriorated.

Further, drift compensation and DCT base
A data upsampling method will also be described.
These methods are suitable for the second and third closed loop architectures as they allow the post-upsampling operation or operation to be performed properly without additional conversion steps.

Reduced spatial resolution transform coding (transcodin
Application to g) The main intended use of the present invention is for digital television (DTV) broadcast and Internet content, eg wireless telephones, pagers and PDAs (personal d).
igital assistance) to distribute or deliver to devices or devices that have low-resolution displays such as. Currently, MPEG-2 is adopted as a compression format for DTV broadcasting and DVD recording, and MPEG-1 content can be used via the Internet.

Since MPEG-4 has been adopted as a compression format for video transmission over a mobile network, the present invention converts MPEG-1 / 2 contents into low resolution MPEG-4 contents. Let's take a look at how to encode.

FIG. 7 shows a first example of a multimedia content distribution system 700 utilizing the present invention. The system 700 comprises an adaptive server 701 connected to a client 702 via an external network 703. One feature of this system is that the client's display is small in size and connected via low bit rate channels. Therefore, it is necessary to reduce the resolution of the content delivered to the client 702.

Input source 7 for multimedia content
04 is stored in the database 710. Content is feature extraction and indexing processing (indexing processing) 7
Receive 20. The database server 740 allows the client 702 to scan the content of the database 710 and make requests for specific content. The search engine 730 can be used to search for multimedia content. Once the desired content is retrieved, the database server 740
Sends the multimedia content to a transcoder 750 according to the invention.

Transform encoder 750 reads network and client characteristics. If the spatial resolution of the content is higher than the display characteristics of the client, the method according to the invention is used to reduce the resolution of the content to match the display characteristics of the client. The present invention can also be applied when the bit rate on the network channel is lower than the bit rate of the content.

FIG. 8 shows a second embodiment of the content distribution system 800. The system 800 comprises a local "home" network 801, an external network 703, a broadcast network 803 and the adaptive server 701 described in connection with FIG. In this example, high quality input source content 804
It can be transferred to the client 805 connected to the home network 801 via a broadcasting network 803, for example, a cable, terrestrial or satellite broadcasting network. The content is received by the set top box or gateway 820 and stored in local memory or hard disk drive (HDD) 830. The received content can be delivered to the client 805 at home. In addition, the content can be transform encoded (850) for use by clients that do not have the ability to decode and display the full resolution content. This is, for example, in high definition television (HDT
This corresponds to the case where the V) bit stream is received by a standard definition television set. Therefore, the content should be transcoded or transcoded to satisfy the capabilities of the client in the home.

Furthermore, when access to the content stored in the local memory (HDD) 830 is requested from the low resolution external client 806 via the external network 802, the conversion encoder 850 is used. It is possible to distribute low-resolution multimedia contents to clients.

Analysis of the Basic Method The signal generated by the method shown in FIG. 4 will be further analyzed and explained in order to design a transcoder having variable complexity and quality. Note that, in relation to the notation in the formula, variables in lowercase represent the signal in the spatial domain, while
Variables in upper case shall represent the equivalent signal in the DCT domain. Also, subscripts attached to variables shall represent time, while superscripts equal to 1 represent signals with drift and superscripts equal to 2 represent signals without drift. Note that drift can occur in lossy processes such as requantization, motion vector truncation (rounding), or downsampling.
The drift compensation method will be described later.

I-frame Motion-compensated prediction is not possible for I-frames.
That is,

[Equation 14] And therefore the signal is downsampled (4
10). That is,

[Equation 15] Is. Next, the encoder or encoder 120 encodes according to the following equation.

[Equation 16]

Signal

[Equation 17] Receives the DCT 440 and then the quantization parameter

[Equation 18] Is quantized by (450). Quantized signal

[Formula 19] Is subjected to variable length coding (460) and is written in the transcoded bitstream 402. As part of a motion compensation loop in an encoder

[Equation 20] Is dequantized (470) and undergoes IDCT 480.
In this way, the reduced resolution reference signal

[Equation 21] 481 is stored in the frame buffer 490 as a reference signal for future frame prediction.

P-frame In case of P-frame,

[Equation 22] From, a reconstructed or reconstructed full resolution image is generated. As with the I-frame case, this signal is then down-converted according to equation (2). Next, the reduced resolution residual (reduced reso
solution residual) is generated.

[Equation 23] The above equation (5) is equivalently expressed as follows.

[Equation 24]

The signal given by equation (6) represents the reference signal to which the architecture according to the invention approximates. It is needless to say that the complexity of the generation of the reference signal is high, and therefore it is desirable to realize the quality approximation while considerably reducing the complexity.

Open loop architecture As an approximation,

[Equation 25] Then, the reduced resolution residual signal of equation (6) is expressed by the following equation.

[Equation 26]

The above equation is the transform encoder 900 shown in FIG.
Suggest an open-loop architecture for.

In the transform encoder 900, the signal of the input bit stream 901 is variable length decoded (91
0), thereby the inverse quantized DCT coefficient 911 and the full resolution motion vector

[Equation 27] 902 is generated. All resolution motion vectors are MV
Mapping (MV mapping) 920 reduces motion vector with reduced resolution

[Equation 28] 903, that is, mapped. Quantized DCT
Coefficient 911 is quantized

[Equation 29] The signal is dequantized at 930

[Equation 30] 931 is generated. This signal is then provided to the block of processors 1300, as described in more detail below. The output of processor 1300 is downsampled (950), thereby

[Equation 31] 951 is generated. After downsampling, this signal is quantized

[Equation 32] (960) is performed. Finally, the reduced resolution requantized DCT coefficients and motion vectors are coded by variable length coding (97
0) and written in the transform-coded output bit 902.

Although details regarding a preferred embodiment of the block group processor 1300 are provided below, for purposes of brief description of this processor, the purpose of this processor is a downsampling process 950 in which the sub-blocks are encoded in different coding modes, such as interlace. It should be mentioned that the selected macroblocks are pre-processed so that macroblocks with block and intra block modes will not occur. It should be noted that the mixed coding mode (mixed coding mode) in the macroblock is not supported by any known video coding standard.

If only the approximation given by the drift compensation equation (7b) in the reduced resolution is given, equation (6)
The reduced resolution residual signal of is expressed by the following equation.

[Expression 33]

The above equations imply the closed loop architecture 1000 shown in FIG. 10 with drift compensation at reduced resolution.

In this architecture, the input signal 1001 is variable length decoded (1010), thereby quantized DCT coefficients 1011 and full resolution motion vectors.

[Equation 34] 1012 is generated. Full resolution motion vector 101
2 is mapped by MV mapping 1020. This reduces the motion vector with reduced resolution.

[Equation 35] 1021 is generated. Quantized DCT coefficient is quantized

[Equation 36] Dequantized (1030) by

[Equation 37] 1031 is generated. This signal is then provided to the block group processor 1300 for downsampling (1
050). After downsampling 1050, the reduced resolution drift compensation signal 1051 is added to the low resolution residual signal 1052 in the DCT domain (1060).

The signal 1061 is a spatial quantizer.

[Equation 38] Quantized at 1070. Finally, the reduced resolution requantized DCT coefficient 1071 and motion vector 1021.
Is variable-length coded (1080) to generate a transform-coded output bitstream 1002.

The reference frame from which the reduced resolution drift compensation signal is generated is requantized residual.

[Formula 39] Dequantized (1090) 1071 and downsampled residual

[Formula 40] It is obtained by subtracting (1092) from 1052. This difference signal is given to the IDCT 1094 and added to the low resolution prediction component 1096 of the previous macroblock stored in the frame memory 1091 (1
095). This new signal is the difference

[Formula 41] 1097, which is used as a reference for low resolution motion compensation for the current block.

A low resolution motion compensation 1098 is performed on the stored reference signal and a prediction is made to the DCT 1099. This DCT domain signal is the reduced resolution drift compensation signal 1051. This motion is a set of low-resolution motion vectors.

[Equation 42] 1021 is used on a macroblock basis.

First Drift Compensation Method at Original Resolution Approximation

[Equation 43] On the other hand, the reduced resolution residual signal of equation (6) is expressed as follows.

[Equation 44]

The above equation suggests the closed loop architecture 1100 shown in FIG. 11 that compensates for drift in the original resolution bitstream.

In this architecture, input signal 1
001 is a variable-length decoded (1110), quantized DCT coefficient 1111 and motion vector of full resolution.

[Equation 45] 1112 is generated. Quantized DCT coefficient 111
1 is a quantizer

[Equation 46] Is dequantized (1130) by

[Equation 47] 1131 is generated. This signal is then provided to the block group processor 1300. Block group processing (13
00), the original resolution drift compensation signal 1151 is converted to DCT.
It is added to the residual signal 1141 of the domain (116
0). The signal 1162 is then downsampled (11
50) and quantizer

[Equation 48] Is quantized (1170). Finally, the reduced resolution restored DCT coefficient 1171 and motion vector 1121.
Is variable length coded (1180) and written to the transcoded bitstream 1102.

The reference frame in which the original resolution drift compensation signal 1151 is generated is the requantization residual component.

[Equation 49] It is obtained by dequantizing (1190) 1171 and upsampling (1191). In this example,
After the upsampling, the upsampled signal is subtracted (1192) from the original resolution residue 1161. This difference signal is provided to the IDCT 1194 to provide the original resolution prediction component 119 of the previous macroblock.
It is added to 6 (1195). This new signal is the difference

[Equation 50] 1197, used as a reference for motion compensation for the current macroblock at the original resolution.

Motion compensation 1198 at the original resolution is performed on the reference signal stored in the frame buffer 1181 to make a prediction for the DCT 1199. This DCT
The domain signal is the original resolution drift compensation signal 1151.
Is. This operation is the motion vector set of the original resolution

[Equation 51] 1121 is used on a macroblock basis.

Second Drift Compensation Method at Original Resolution FIG. 11B shows another variation of the closed loop architecture shown in FIG. 11A. In this example, the requantization residual

[Equation 52] The inverse quantized output 1190 of 1172 is subtracted (1192) from the reduced resolution signal before upsampling 1191.

The above two drift compensation architectures at original resolution do not use motion vector approximation to generate the drift compensation signal 1151. this is,
This can be realized by adopting up-sampling 1191. The two alternative architectures described above differ primarily in the selection of the signals used to generate the difference signal. In the first method, the difference signal represents errors due to requantization and resolution conversion, whereas in the difference signal in the second method only errors due to requantization are considered.

Since the upsampled signal is not considered in future decoding or decoding of the transform-coded bitstream, any error measured by successive downsampling and upsampling in the drift-compensated signal will be accounted for. It is reasonable to exclude also. However, upsampling is adopted for two reasons. That is, the motion vector 1121 of the full resolution is used to avoid the subsequent approximation, and the drift compensation signal is set to the original resolution and added to the input residual component 1161 before the down sampling 1150 (11
60) This is to be able to do it.

The purpose of the mixed block processor block group processor 1300 is to select macroblocks such that downsampling does not produce macroblocks in which the subblocks have different coding modes, eg, interblock mode and intrablock mode. It is a pretreatment.
Coding modes mixed within a macroblock are not supported by any known video coding standard.

FIG. 12 shows an example of a macroblock group 1201 that can generate a block group 1202 with reduced resolution after transform coding 1203. In this example, there are three inter mode blocks and one intra mode block. Note that the motion vector (MV) of the intra mode block is zero. The determination as to whether a specific block group is a mixed mode group (mixed mode group) depends solely on the macroblock mode. The block group processor 1300
A group of four macroblocks 1201 forming a single macroblock 1202 with reduced resolution is handled. In other words, for the luminance component, MB (0) 1210
Corresponds to the subblock b (0) 1220 in the reduced resolution macroblock 1202, and similarly MB (1)
1211 corresponds to b (1) 1221, and MB (k) 12
12 corresponds to b (2) 1222, and MB (k +
1) 1213 corresponds to b (3) 1223. Here, k represents the number of macroblocks for each column in the original resolution. Chrominance components are also processed in a similar manner to match luminance modes.

The MB mode group determines whether the block group processor 1300 should process a particular MB (multi-block). If the block group includes at least one intra mode block and at least one inter mode block, processing of the block group is performed. After the macroblock is selected, its DCT coefficient and motion vector data are modified or changed.

FIG. 13 shows a block group processor 1300.
Shows the components, or components. For the selected mixed block group 1301, the block group processor performs mode mapping 1310 and motion vector correction 13
20 and DCT coefficient modification 1330 to produce an unmixed mode block output 1302. Block group 1301
Is identified or identified, the mode of the macroblock is changed so that all macroblocks are the same. This is done according to a pre-specified strategy that matches the mode of each sub-block in the reduced resolution block.

According to the selected mode map, the MV data is then subjected to a modification or change process 1320. Possible modifications or changes to fit the corresponding mode maps are described in detail below with reference to FIGS. 14A-14C. In this case, the new MB (macroblock) mode and the MV (motion vector) data have corresponding D
The CT coefficients are also modified or changed (1330) to match the mapping.

In the first embodiment of the block group processor shown in FIG. 14A, M of the block group 1301 is used.
The B mode is changed to the inter mode by the mode map 1310. Therefore, the MV data of the intra block is reset to zero by the motion vector process, and the DCT coefficient corresponding to the intra block is also reset to zero by the DCT process 1330. In this way, the transformed block is duplicated with the data from the corresponding block in the reference frame.

In the second embodiment of the block group processor shown in FIG. 14B, the MB of the mixed mode block group is
The mode is changed to inter mode by mapping, mapping 1310. Therefore, unlike the first preferred embodiment, the MV for intra MB (macroblock)
The (motion vector) data will be estimated. This estimation is based on data in adjacent blocks that may include both texture and motion data. And
Based on this estimated motion vector, a new residue for the modified block is calculated. In the final step 1320, the inter DCT (discrete cosine transform) coefficient is reset to the intra DCT coefficient.

In the third embodiment shown in FIG. 14C, the MB mode of the block group is changed to the intra mode (1310). In this case, since there is no motion information associated with the reduced resolution macroblock, all associated motion vector data is reset to zero (132).
0). This needs to be done in the transcoder.
The reason is that the motion vector of the adjacent block is estimated or predicted from the motion of this block. To ensure proper reconstruction at the decoder, the MV data for the blocks must be reset to zero at the transcoder. In the final step 1330,
As described above, the intra DCT coefficient that should replace the inter DCT coefficient is generated.

In order to realize the second and third embodiments described above, a decoding loop (recording loop) for reproducing at full resolution can be used. The reproduced data can be used as reference data for converting the DCT coefficient between the intra mode and the inter mode or between the inter mode and the intra mode. However, the use of a coding loop for that is not necessarily required. Alternatively, the conversion can be performed in the drift compensation loop.

For a series of frames with a small amount of motion and a low level of detail, the low complexity strategy shown in FIG. 14A can be used. In other cases, the strategy with the appropriate complexity shown in FIG. 14B or 14C should be adopted.
It should be noted that the strategy shown in FIG. 14C guarantees the best quality.

Drift Compensation in Block Processing Blocks processor 1300 can also be used to control or minimize drift. Since the intra-coded block does not undergo drift, the inter-coded block (inter-coded block)
The effect of drift can be reduced by converting d block) into an intra-coded block.

In the first step 1350 of FIG. 14C, the amount of drift in the compressed bitstream is measured. In the case of a closed loop architecture, this drift can be measured according to the energy of the difference signal generated by 1092 and 1192 or the energy of the drift compensation signal stored at 1091 and 1191. A known method can be used to calculate the signal energy. The calculated energy is considered in various approximations including requantization, downsampling and motion vector truncation (rounding).

Another method of computing drift that is also applicable to open loop architectures estimates the error introduced by truncation or rounding of motion vectors. It is known that the half-pixel motion vector at the original resolution causes a large reproduction error when the resolution is reduced. In contrast, the full pixel motion vector does not have such an error. This is because the full pixel motion vector can be correctly mapped to the half pixel region. Therefore, one possible way to measure drift is to record the percentage or percentage of the half-pixel motion vector. However, since the impact of the motion vector approximation depends on the complexity of the content, it is possible that the measured drift is a function of the residual components associated with blocks with half-pixel motion vectors.

The methods using the energy of the difference signal and the motion vector data for the purpose of measuring the drift can be implemented in combination, or can be adopted over a partial region in the frame. Since the most convenient macroblock position can be identified or identified by the drift compensation method, it is advantageous to apply the method to a partial area in a frame. To use the above methods in combination, the drift is measured by the energy of the difference signal or the drift compensation signal for the macroblock with half-pixel motion vector at the original resolution.

In the second step, the measured drift value is
"Intra refresh rate" 13 used as input to block group processor 1300
Converted to 51. Controlling the percentage of intra-coded blocks has traditionally been considered in encoding video with error-resident transmission. For example, "Analysis of Vide
o Transmission over Lossy Channels '' Journal of Sel
ected Areas of Communications, by Stuhlmuller, et
See al, 2000. In this paper, the inverse channel from the receiver to the encoder conveys the amount of loss introduced by the transmission channel and intra-codes from the source side to prevent the error rate due to the lost data in the predictive coding scheme. The blocks are coded directly.

In contrast, in the present invention, a new intra block is generated in the compression domain for a video that has already been encoded and the conversion from inter mode to intra mode is accomplished by the block group processor 1300. It

When the drift exceeds the drift threshold amount, the block group processor 1300 shown in FIG. 14C is activated to convert an inter mode block into an intra mode block. In this case, the conversion is a pre-specified fixed intra refresh rate.
Done in. Alternatively, the conversion can be done at an intra refresh rate that is proportional to the amount of measured drift. Further, it is possible to set an appropriate compromise between the intra refresh rate and the quantizer used for encoding the intra block and the inter block in consideration of the rate distortion characteristic of the signal.

The present invention now generates a new intra block in the compression domain, and this drift compensation scheme can be performed with any transcoder or transform encoder with or without resolution reduction. it can.

Downsampling The transform encoder according to the present invention can employ any downsampling method. However, the preferred downsampling method is Sun et al.
U.S. Pat. No. 5,855,151 dated May 10 "Method
and apparatusfor down-converting a digital signa
The downsampling method described in "1" is advantageous. Note that the disclosure content of this US patent specification is incorporated herein by reference.

The concept of this downsampling method is shown in FIG.
5A. One group has four

[Equation 53] A DCT block 1501 is included. That is, the size or size of the group is

[Equation 54] Is. Applying frequency synthesis or filtering 1510 to a block group

[Equation 55] Generate DCT block 1511. DCT block 1 downsampled from this synthesized block
512 can be extracted.

Although the above operations have been described with respect to the DCT domain using 2D operations, operations can also be performed using separable 1D filters.
Furthermore, the operations can also be performed entirely in the spatial domain. Also, US patent application Sn. 09 / 135,969 “Three layer sca” dated March 6, 1998 outside Vetro.
An equivalent spatial domain filter can be derived using the method described in "Lable decoder and method of decoding". Note that the disclosure content of the specification of this US patent application is also incorporated herein by reference.

The main advantage of using the downsampling method in the transform encoder according to the invention is that the correct dimension of the sub-blocks within the macroblock is directly obtained. For example, a single 8x8 block can be formed from four 8x8 DCT blocks. On the other hand, conventional downsampling methods produce data that is downsampled in a dimension that is not equal to the required dimension in the output subblock of the macroblock. For example, four 4 × 4 DCT blocks are obtained from the 8 × 8 DCT block. Therefore, the conventional method would require an additional step to construct a single 8x8 DCT block.

The filter described above is a useful component for efficiently implementing the architecture shown in FIG. 11 which requires upsampling. In general, the filters derived here are upsampled D with or without resolution reduction or drift compensation.
It can be applied to any system that requires calculation on CT data.

Upsampling Any conventional upsampling means may be used in the present invention. However, the previously cited US patent application “Three layer scalable decode”
The "r and method of decoding" states that the optimal upsampling method depends on the downsampling method. Therefore, the downsampling filter

[Equation 56] Upsampling filter corresponding to

[Equation 57] It is advantageous to use The relationship between the above two filters is given by the following equation.

[Equation 58]

There are two problems associated with the filter derived from the above equation. The first problem is that the DCT filters are not invertible, so they can only be applied to spatial domain filters. However, this is not a big problem, since the corresponding spatial domain filter can be derived and transformed into the DCT domain.

However, the second problem is different in that the upsampling filter thus obtained corresponds to the process shown in FIG. 15B. In this process, for example,

[Equation 59] Block 1502 is single

[Equation 60] Upsampled (1520) to block 1530. If all the upsampling is done in the spatial domain then no problems occur. However, if upsampling is done in the DCT domain, one

[Equation 61] One has to deal with the DCT block, ie one DCT component. This is because the upsampled DCT block is in standard MB format, ie 4

[Equation 62] It is not suitable for operations that require a DCT block (however, N = 4). That is, the upsampled block has the same format or dimension as the larger number of original blocks.

The above upsampling method in the DCT domain is not suitable for use in the transform encoder described in connection with the present invention. Referring to FIG. 11A, the upsampled DCT data is subtracted from the DCT data output from the mixed block processor 1300. The two DCT data of these two blocks must have the same format. Therefore, FIG.
A filter capable of performing the upsampling shown in C is required. Where single

[Equation 63] Block 1502 has four

[Equation 64] Upsampled (1540) to block 1550. Since such a filter has not been considered in the past and does not exist in the prior art, the equation for the 1D case will be described below.

Note that in relation to the notation in the equations described below, lowercase variables represent spatial domain signals, while uppercase variables represent equivalent signals in the DCT domain.

As shown in FIG. 16, C1601 is a DCT
The domain represents the DCT block to be upsampled, and c1602 represents the equivalent block in the spatial domain. These two blocks are N-ptDCT and I
They are associated with each other by the definition of DCT 1603. For example, Rao and Yip's Discrete Cosine Transform: Alg
orithms, Advantages and Applications '' Academic, Bo
See ston, 1990. For convenience, it is represented by a mathematical formula below.

The definition of DCT is given by the following equation.

[Equation 65] The definition of IDCT is given by the following equation.

[Equation 66] In the above equations (13) and (14),

[Equation 67] Is.

From the top, block E1610

[Equation 68] 1611 represents an upsampled DCT block based on filtering C in 1611, where e is given by equation (12)

[Equation 69] 16b represents an upsampled spatial domain block based on filtering c at 1621. e and E are 2
Note that it is related by N-ptDCT / IDCT1630. Input of filtered input /
The output relation is given by the following equation.

[Equation 70]

Referring to FIG. 16, the desired DCT block is represented by A1611 and B1612. This purpose is a filter that can be used to calculate A and B, respectively, directly from C.

[Equation 71] 1641 and

[Equation 72] 1642 is to be derived.

In the first step, equation (14) is substituted into equation (16b).

The equation thus obtained is the DCT input C
Is an expression of the spatial domain output e as a function of, and is expressed as

[Equation 73]

When A and B are represented by C using the equation (17),
The spatial domain relationship between a, b and e is as follows.

[Equation 74] In the above formula, i represents a spatial domain index (index).
The DCT domain expression for a is given by:

[Equation 75]

The following equation is obtained from the equations (17) to (19).

[Equation 76] The above equation is equivalently expressed as follows.

[Equation 77] In the above formula,

[Equation 78] Is. Similarly, the following equation holds.

[Equation 79] The above equation is equivalently expressed by the following equation.

[Equation 80] In the above formula,

[Equation 81] Is.

The filter described above can then be used to upsample a single block of a given dimension or size into multiple blocks, each having the same dimension as the original block. In general, the filter derived here is applicable to any system where computation is required on upsampled DCT data.

To realize the filter given by equations (22) and (25), k × q of filter taps
Consider the matrix. Here, k is an index (index) of the output pixel and q is an index (index) of the input pixel. For 1D data, the output pixel is calculated as a matrix multiplication. Two steps are taken for 2D data. First, the data is upsampled in a first direction, for example the horizontal direction. Next, the up-sampled data in the horizontal direction is up-sampled in the second direction, for example, the vertical direction. Reversing the order of directions for upsampling does not affect the result.

In the case of upsampling in the horizontal direction, each column in the block is individually operated and treated as an N-dimensional input vector. Each input vector is
Filtering is performed according to equations (21) and (24). Two standard DCTs as the output of this process
You get a block.

In vertical upsampling, each row in the block is individually operated on and treated as an N-dimensional input vector. As in the case of upsampling in the horizontal direction, each input vector is
The filtering process is performed according to 1) and (24). The output of this process is 4 as shown in FIG. 15C.
Two standard DCT blocks are obtained.

Open Loop Transcoder Drift Error Analysis An analysis of the drift error caused by a reduced resolution transcoder is described below. The analysis is based on the open loop transcoder shown in FIG.
In this transcoder, the reduced resolution residual is given by:

[Equation 82] Drift error compared to equation (6)

[Equation 83] Is represented by the following formula.

[Equation 84] However, the following equation holds.

[Equation 85] In the above equation, the drift error has two components. First component

[Equation 86] Represents the error in the reference frame used for motion compensation. This error is caused by requantization removing non-zero DCT coefficients and arithmetic error due to integer truncation. This is a common drift error in many transcoders. "A freq by Assuncao et al.
uency domain video transcoder for dynamic bit-rate
reduction of MPEG-2 bitstreams "(IEEE Transactio
ns on Circuits and Systems for Video Technology, p
p. 953-957, 1998). In this case, the frame initially used by the transcoder as a reference is different from its corresponding frame in the decoder, thereby causing a mismatch between the prediction and residual components. Second component

[Equation 87] Due to the non-commutative nature of motion compensation and downsampling, which is inherent in reduced resolution transcoding.

[Equation 88] There are mainly two factors that contribute to the effect of: motion vector (MV) mapping and downsampling. In mapping the MV from the original resolution to the reduced resolution, the MV is truncated due to the limited accuracy of encoding the MV. In downsampling to a lower spatial resolution of the compressed area, block constraints are often adhered to to avoid filter overlap between blocks. Due to these constraints, the complexity of the system is reduced, but the quality of the downsampling process is degraded and usually introduces some error. Regardless of the magnitude of these errors in the case of a frame, the combination of these two transforms generally produces a larger mismatch between the prediction and residual components, which results in successive frames being predicted. Increase every time. To show this mismatch between the prediction and residual components due to non-commutativeness of motion compensation and downsampling, we consider an example using a 1-D signal and consider all Drift error (ie

[Equation 89] ) Is ignored. All at the original resolution,

[Equation 90] Denote the reconstructed block, and

[Formula 91] Denote the reference block,

[Equation 92] Denote the error (residual) block. further

[Equation 93] Denote the maximum resolution motion compensation filter,

[Equation 94] Shall represent a motion compensation filter with reduced resolution. The reconstructed block at the original resolution is then given by:

[Formula 95] When applying the down-conversion process to both sides of equation (30), the following equation is obtained.

[Equation 96] The quality of the signal produced by the above equation is

[Numerical Expression 97] Is not affected by the drift error contained in. However, this is not the signal produced by the reduced resolution transcoder. The actual reconstructed signal is given by:

[Equation 98] Because

[Numerical expression 99] Therefore, there is a mismatch between the reduced resolution prediction component and the residual component. In order to achieve the quality produced by equation (31), either or both of the prediction and residual components need to be modified to match each other. In the reference transcoder of FIG. 4, this mismatch is removed using a second coding loop that determines a new reduced resolution residual. The second loop is used to readjust the prediction and residual components.

Reduced Resolution Transcoding with Drift Compensation

[Equation 100] The reduced resolution residual signal of equation (6) can be expressed as:

[Equation 101] The above equation suggests the closed loop transcoder shown in Figure 17, which compensates for drift in the reduced resolution signal. Video Transcoder with Drift Compensation by Partial Coding FIG. 17 is a block diagram of a closed loop transcoder 1700 for reducing spatial resolution with drift compensation in a reduced resolution signal according to the present invention. Transcoder 1700 includes decoder 1703 and partial encoder 1704. Transcoder 1700
Then, the input signal 1701, that is, the sequence of frames of the compressed video signal bitstream is supplied to the decoder 1703. The decoder 1703 has VLD1
710, inverse quantization 1720, IDCT 1730,
Motion compensation 1740. The decoded frame is
When the maximum resolution motion vector of each previously decoded frame is added (1780) to the motion vector of the next frame to be decoded, a first value is added for motion compensation (1740) during decoding (1703). Frame buffer 1 of 1
It is stored in 760. Each frame of the decoded bitstream is downsampled by downconversion block 1750. The reduced resolution frame is subtracted from the current reduced resolution frame for motion compensation of the previous reduced resolution frame for motion compensation during subcoding (1704) ( 1782) when partially coded (1704)
It is stored in the second frame buffer 1760 for motion compensation (1770) therein. Motion compensation at the decoder 1703 for a full resolution frame is performed using the full resolution motion vector.

[Equation 102] While motion compensation in the partial encoder 1704 for reduced resolution frames (1770)
Is a low resolution motion vector

[Equation 103] To use. The low resolution motion vectors are either estimated from downsampled spatial domain frames or mapped from the full resolution motion vectors (1765). The reduced resolution residual is obtained by subtracting the motion compensated prediction of the previous reduced resolution frame from the current low resolution frame (1782). The reduced resolution residual is then subjected to DCT1783, quantization 1784 and VLD1786 operations to produce a reduced resolution and drift compensated output transcode bitstream 1702. The transcoder 1700 according to the present invention is

[Equation 104] Reduce drift errors caused by.

[Equation 105] Is usually

[Equation 106] Being significantly larger, transcoder 1700 minimizes the complexity associated with completely reconstructing a reference frame, which is typically used by prior art decoders to form motion compensated predictions. . Therefore, the dequantization 470, IDCT 480 and addition operations in the prior art decoder 400 shown in FIG. 4 are omitted. The drift compensation according to the present invention can be considered as full decoding and partial coding, and the requantization error is shown in FIG.
It is not compensated as in the case of. Finally, it should be noted that since full resolution decoding is performed in transcoder 1700, there is no mixed block problem as in the prior art. Although the present invention has been described by way of example of preferred embodiments, it is to be understood that various other adaptations and modifications can be made within the spirit and scope of the invention. Therefore, the purpose of the appended claims is to cover all such variations and modifications as fall within the true spirit and scope of the invention.

[Brief description of drawings]

FIG. 1 is a block diagram of a conventional cascaded transform encoder or transcoder.

FIG. 2 is a block diagram of a conventional open loop transform encoder for bit rate reduction.

FIG. 3 is a block diagram of a conventional closed loop transform encoder for bit rate reduction.

FIG. 4 is a block diagram of a conventional cascaded transform encoder for spatial resolution reduction.

FIG. 5 is a block diagram of a conventional open loop transform encoder for spatial resolution reduction.

FIG. 6 is a block diagram of a conventional motion vector mapping or mapping.

FIG. 7 is a block diagram showing a first embodiment of the present invention in which the first bitstream transform coding according to the present invention is applied to spatial resolution reduction.

FIG. 8 is a block diagram showing a second embodiment of the present invention to which a coding conversion of a bitstream into a reduced spatial resolution according to the present invention is applied.

FIG. 9 is a block diagram of an open-loop coding converter for reducing spatial resolution according to the present invention.

FIG. 10 is a block diagram of a first closed-loop transform encoder for spatial resolution reduction that performs drift compensation with reduced resolution according to the present invention.

FIG. 11A is a block diagram of a second closed-loop transform encoder for spatial resolution reduction that performs drift compensation at the original resolution according to the present invention.

FIG. 11B is a block diagram of a third closed loop transform encoder for spatial resolution reduction that performs drift compensation at the original resolution according to the present invention.

FIG. 12 is a diagram showing an example of a macroblock group including a macroblock mode, DCT coefficient data, and corresponding motion vector data.

FIG. 13 is a block diagram of a block group processor according to the present invention.

FIG. 14A is a block diagram illustrating a first block group processing method according to the present invention.

FIG. 14B is a block diagram illustrating a second block group processing method according to the present invention.

FIG. 14C is a block diagram illustrating a third block group processing method according to the present invention.

FIG. 15A is a diagram illustrating the conventional idea of downsampling in the DCT or spatial domain.

FIG. 15B is a block diagram illustrating conventional upsampling in the DCT or spatial domain.

FIG. 15C is a block diagram illustrating upsampling in the DCT domain according to the present invention.

FIG. 16 is a diagram illustrating upsampling in the DCT domain according to the present invention.

FIG. 17 is a block diagram of a closed loop transcoder for reducing spatial resolution with drift compensation according to the present invention.

   ─────────────────────────────────────────────────── ─── Continued front page    (72) Inventor Anthony Vetro             Stay, New York, United States             Ton Island, Regis Drive             113 (72) Inventor Hai Phan Seung             Ku, New Jersey, United States             Ranberry, Kinglet Drive             Us 61 (72) Inventor Penn In             Pue, New Jersey, United States             Linston, Hallsey Street             222 bees (72) Inventor Bede Liu             Pue, New Jersey, United States             Linston, Hartley Avenue             248 F-term (reference) 5C059 KK41 LB04 MA00 MA01 MA23                       MA31 MC11 ME01 NN21 PP05                       PP06 SS02 SS08 UA02 UA33                 5J064 AA01 AA02 BA09 BA16 BB01                       BB03 BC01 BC16

Claims (8)

[Claims] 1. A method for transcoding a compressed bitstream of a sequence of frames of a video signal to a reduced spatial resolution, the decoding of the frames, the decoding of the decoded frames Storing the decoded frame in a second frame buffer, storing the decoded frame in a second frame buffer, down-sampling the decoded frame to a reduced resolution, Partially encoding a frame to produce a reduced resolution compressed bitstream of the video. 2. The decoding includes variable length decoding the bitstream to produce an output including a full resolution motion vector and quantized DCT coefficients for each block in each frame. Dequantizing the quantized DCT coefficient for each block in each frame, applying an inverse DCT to the dequantized block of the frame, the stored Motion compensation with a full resolution motion vector of the decoded frame. 3. The partially encoding comprises motion compensating with a reduced resolution motion vector of the stored reduced resolution frame; and a DCT comprising the reduced resolution motion compensation. The method of claim 1, further comprising: applying a quantized difference; quantizing a DCT block of the frame; and variable length coding the quantized block of the frame. 4. The method of claim 2, wherein the motion compensation during decoding further comprises adding a motion compensated prediction value at a maximum resolution of a previously decoded frame to a current frame. . 5. The motion compensation during the sub-coding further comprises subtracting a reduced resolution motion compensated prediction value of a previously reduced resolution from a current reduced resolution frame. The method of claim 3 including. 6. The method of claim 3, further comprising estimating the reduced resolution motion vector from the reduced resolution frame. 7. The method of claim 2, further comprising mapping the full resolution motion vector to the reduced resolution motion vector from the variable length decoded frame. 8. A closed-loop transcoder for transcoding a compressed bitstream of a sequence of frames of a video signal to a reduced spatial resolution, partially decoded from said compressed bitstream. A decoder for motion compensation using a maximum resolution motion vector stored in a first frame buffer to generate a frame; and a down downsampler for downsampling the decoded frame into a reduced resolution frame. A conversion block and a partial encoder for motion compensation using a reduced resolution motion vector stored in a second frame buffer to generate a reduced spatial resolution compressed bitstream of the video; Closed-loop transcoder including.