ONE-PASS VARIABLE BIT RATE MOVING PICTURES ENCODING
Field of the nvention
The present invention relates to a method and apparatus for encoding moving pictures. In particular, the present invention relates to a method and apparatus for performing variable bit rate control for a digital video encoder.
Background of the Invention
Digital compression has been applied to moving pictures for the purposes of transmissions bandwidth or storage size reduction. One current art of such compression techniques can be derived from the ISO/TEC MPEG Standards, the ISO/TEC 11172-3 (MPEG-1), the ISO/IEC 13818-2 (MPEG-2) and the MPEG-2 TM5 (test model 5), developed by the Moving Picture Experts Group of the International Organisation for Standardization. The disclosures of those standards documents are hereby expressly incorporated into this specification by reference.
In a standard MPEG compliant video encoder, a sequence of moving pictures (e.g. video) is input to the encoder where it is compressed with a user defined target bitrate. The target bitrate is set according to the communication channel bandwidth in which the compressed video is to be transmitted, or the storage media capacity in which the compressed video sequence is to be stored.
Several different forms of coding can be employed depending upon the character of the input pictures, referred to as I-pictures, P-pictures, or B-pictures. The I-pictures are intra-coded pictures used mainly for random access or scene update. The P-pictures use forward motion predictive coding with reference to previously coded I- or P^ pictures (anchor pictures), and the B-pictures use both forward and backward motion predictive/interpolative coding with reference to previously coded I-or P-pictures. Furthermore, a group of picture (GOP) is formed in encoded order starting with an I-picture and ending with the picture before the next
I-picture in the sequence.
The pictures are partitioned into smaller and non-overlapping blocks of pixel data called Macroblocks (MBs) before encoding. Each MB from a P- or B-picture is subjected to a motion estimation process in which forward motion vectors, and backward motion vectors in the case of a B-picture MB, are determined using reference pictures from a frame buffer. With the determined motion vectors, motion compensation is performed where the intra-or inter-picture prediction mode of the MB is first determined according to the accuracy of the motion vectors found, followed by generating the necessary predicted MB.
The predicted MB is then subjected to discrete cosine transform (DCT) and DCT coefficients quantization based on quantization matrices (QM) and quantization stepsize (QS). The quantized DCT coefficients of the MB is then run-length encoded with variable length codes (VLC) and multiplexed with additional information such as selected motion vectors, MB coding modes, quantization stepsize, and/or picture and sequence information, to form the output bitstream.
Local decoding is performed by inverse quantizing the quantized DCT coefficients, followed by inverse DCT, and motion compensation. Local decoding is performed such that the reference pictures used in the motion compensation are identical to those to be by an external decoder.
The quantization stepsize (QS) used for quantizing the DCT coefficients of each MB has direct impact on the number of bits produced at the output of the VLC encoding process, and therefore the average ouφut bit rate. It has also a direct impact on the encoding quality, which is output picture quality at the corresponding decoder. In general, larger QS generates lower output bit rate and lower encoding quality. In order to control ouφut bit rate and picture quality so that the resulting bitstream can satisfy channel bandwidth or storage limitations as well as quality requirements, rate control and quantization control algorithms are used.
Some methods for rate control and quantization control can be found in the abovementioned MPEG-2 TM5 (Test Model 5). These methods comprise generally a bit allocation process, a rate control process, and an adaptive quantization process. In the bit allocation process, a target number of bits is assigned for a new picture to be coded according to a number of previously determined and preset parameters. The rate control step then calculates a reference quantization stepsize QSτef for each MB based on the target bits for the picture, the number of bits already used from the target bits in encoding MBs from that picture, and a virtual buffer model as given in MPEG-2 TM5. In the adaptive quantization process, the calculated QSref is then scaled according to local activities of the MB, and an average MB activity determined from the previously coded picture. This scaling is done according to a level of masking effects of coding noise by human perception for MB with high or low activities within a picture. An example of an adaptive quantization technique is disclosed in United States Patent number 5,650,860, entitled "Adaptive Quantization". A video buffer verifier (VBV) may also be employed in such a way that underflow and overflow are prevented as required by the MPEG standard to ensure the target bit rate is maintained. Techniques for underflow detection and protection are also disclosed in US Patent number 5,650,860.
It is apparent that the fixed target bitrate in the process outlined above has little or no relationship to the actual or varying complexity of the video scenes contained in the input picture sequence. The target bitrate is actually defined by the communication channel bandwidth, or by the target storage capacity for the picture sequence, but the perceptual quality of the resulting pictures when decoded may vary from good to annoying from scene to scene according to scene complexity.
For applications where picture sequences are compressed for storage and retrieval, for example DVD (Digital Video Disc or Digital Versatile Disc), variable bit rate (VBR) may be applied on individual segments of the picture sequences depending on its scene complexity to maximize bit rate allocation and encoded picture quality. Data bits may be reduced for less complex scene to save storage space and increase potential recording duration of the medium, or the resulting storage saving can be used for coding of more complex scenes.
Similarly, VBR can also be applied to other applications such as a multi-channel video broadcasting network. Such channel bandwidth may be dynamically allocated to individual video sequences to be multiplexed together so that higher percentage of the bandwidth is used adaptively by sequences with complex scenes.
Existing VBR control algorithms such as that disclosed in US Patent 5,650,860 require multiple encoding passes to properly distribute data bits. In the first coding pass, the bit utilization information is determined for each scene or each picture in the input picture sequence. This may be done by fixing the reference quantization stepsize and disabling the VBV control.
The determined bit utilization information is then be used to generate a bit budget for each scene or picture such that an overall target number of bits to code the sequence is fixed, and so that a maximum bit rate is not violated. To accomplish this, the bit budget for each picture is modified so that the VBV buffer does not underflow. In cases that initial bit utilization information obtained is unrealistic for generating the bit budget, steps from the first coding pass must be repeated with an adjusted reference quantization stepsize. The input sequence is coded in a final pass using the generated bit budget information to achieve the target bits or bit rate. This form of multiple-pass VBR encoder requires very large storage memory for storing the intermediate bit utilization information, and large computational capacity for the additional passes and the bit budget generation. Furthermore, such a VBR technique such as this cannot process the input sequence in real-time.
Summary of the Invention
An object of the present invention is to provide a one-pass variable bit rate control technique for coding of moving pictures.
In accordance with the invention, there is provided a method for variable bit rate control in
a moving pictures encoder, comprising: selecting a target picture encoding quality; selecting upper and lower bit rate limits; encoding at least one picture according to a target bit allocation based on an initial bit rate within the upper and lower bit rate limits; comparing the encoding quality of the at least one encoded picture with the target picture encoding quality; and adjusting the target bit allocation on the basis of said comparison, for encoding subsequent pictures.
In one form of the invention, the picture encoding quality is based on encoding quantization step-size.
In another form of the invention, the picture encoding quality is based on a mean square error. Alternatively, the picture encoding quality may be based on a signal-to-noise ratio. Preferably the comparing step includes measuring a difference in picture encoding quality between corresponding input and locally decoded pictures.
In a particular form of the invention, the pictures to be encoded are arranged in groups of pictures comprising an I-picture and at least one P-picture and/or B-picture, and wherein the or plurality of pictures in each group of target bit allocation is adjusted for each picture j_ pictures. Preferably the target bit allocation is adjusted to achieve a target bit rate, determined on the basis of said comparison, for each picture or plurality of pictures in the group of pictures.
The target picture encoding quality may comprise a target encoding quantization step-size, wherein different target quantization step-sizes are selected for I-, P-, and B-pictures.
In one form of the invention, the method may include measuring an average quantization stepsize for at least one previously encoded picture, predicting a bit rate for a previously encoded I-, P-, and B-picture, and determining said target bit rate based on said predicted bit rate and
a difference between the target encoding quantization step-size and the measured average quantization step-size.
The method of the invention may include measuring an average picture activity for the moving pictures, and modifying the measured difference in picture encoding quality on the basis of the average picture activity.
The present invention also provides control apparatus for a moving picture encoder, comprising: an input for receiving a target picture encoding quality; an input for receiving upper and lower bit rate limits; and and a controller for controlling the encoder so as to encode at least one picture according to a target bit allocation based on an initial bit rate within the upper and lower bit rate limits, compare the encoding quality of the at least one encoded picture with the target picture encoding quality, and adjust the target bit allocation on the basis of said comparison for encoding subsequent pictures.
Preferably the moving picture encoder includes a frequency transform coefficient quantizer for quantization of the encoded picture data, and wherein the controller comprises a bit rate controller coupled to control the quantization step size of the quantizer, a quantization step size comparator for comparing, as a measure of encoding quality, an actual quantization step size with a target quantization step size based on the target picture encoding quality, a bit allocation processor coupled to control the bit rate controller according to a number of bits remaining for encoding a group of pictures, and a target bit rate estimator coupled to receive the upper and lower bit rate limits and coupled to the bit allocation processor and the bit rate controller for controlling the quantization so that the required bit rate for the quantized picture data is within the upper and lower bit rate limits.
The present invention further provides a video picture encoder comprising: a picture input for receiving data for a plurality of moving pictures;
a target quality input for receiving a target quality measure for encoded pictures; an encoder ouφut for supplying encoded picture data; a bit rate limit input for receiving upper and lower bit rate limits for the encoded picture data; a frequency transform processor for frequency transform encoding picture data from the picture input; a coefficient quantizer for quantizing the encoded picture data according to a quantization step size; and a bit rate controller for dynamically controlling the quantization step size of the coefficient quantizer based on the target quality so that the encoder ouφut remains within the upper and lower bit rate limits.
The video picture encoder may further include a frame bit counter for a number of remaining bits available for encoding a group of pictures, and a bit allocation processor for controlling the bit rate controller according to the remaining available bits.
The video picture encoder may further include a quality comparator for comparing the target quality with a measured encoding quality, and wherein the bit rate controller controls the quantization step size based on a difference between the target and measured qualities.
In one form of the encoder, the measured encoding quality comprises the quantization step size.
The v ideo picture encoder may further include a local decoder for decoding the quantized encoded picture data, and a quality measurement processor for determining a quality difference between corresponding input and locally decoded pictures.
In another form of the encoder, the quality measurement processor determines a difference in signal-to-noise ratio. Alternatively, the quality measurement processor may determine a mean square error.
8
It is possible in most applications that only a maximum bit rate and optionally a minimum bit rate are specified. Such applications may include a randomly accessible recording medium or packetized communication network with variable instantaneous bit rate but also a maximum bandwidth specification. In addition, such applications may also require that the input must be compressed in real-time; for example, live broadcasting or live recording cannot make use of multi-pass encoding. Therefore, it is also an object of the present invention to provide an encoder which can be operated in real time in a variable bit rate mode within the maximum and minimum bitrate boundary of the target application.
While encoding an input moving picmre sequence, the present invention continuously measures the resulting encoding picture quality, compares it to a defined target quality, and adjusts the encoding bit rate accordingly. By varying the target bit rate of the encoder within a defined maximum bit rate and a minimum bit rate according to a defined target encoded picture quality and the scene complexity, the encoder ensures consistent picture quality when possible and also that the maximum and minimum bit rate of the target application are not violated.
Brief Description of the Drawings
The invention is described in greater detail hereinafter, by way of example only, with reference to the accompanying drawings, wherein:
Figure 1 is a block diagram of an MPEG compliant constant bit rate video encoder constructed according to the prior art;
Figure 2 is a block diagram of a variable bit rate moving pictures encoder according to one embodiment of the present invention; and
Figure 3 is a block diagram of a moving pictures encoder according to an alternate embodiment of the present invention.
Detailed Description of the Preferred Embodiments
A block diagram of a generalized MPEG compliant video encoder 150 is illustrated in Figure 1. A sequence of moving pictures (e.g. video) is input at 100 to the encoder 150 where it is compressed with a user defined target bitrate provided at 111. The target bitrate is set according to the communication channel bandwidth in which the compressed video is to be transmitted, or to the storage media capacity in which a target duration of the compressed video is to be stored.
Input pictures, which may be pre-processed for noise reduction or other purposes, are labelled as either I-pictures, P-pictures, or B-pictures, and stored in a frame buffer 101 according to re-ordering and processing needs. The I-pictures are intra-coded pictures used mainly for random access or scene update. The P-pictures use forward motion predictive coding with reference to previously coded I- or P- pictures (anchor pictures), and the B-pictures use both forward and backward motion predictive/interpolative coding with reference to previously coded I- or P-pictures. Furthermore, a group of pictures (GOP) is formed in encoded order starting with an I-picture and ending with the picture immediately before the next I-picture in the sequence. In the case of an MPEG-2 encoder, the input picture sequence may be either a field or frame structure sequence derived for an interlaced or progressive source.
All pictures are partitioned into smaller and non-overlapping blocks of pixel data called Macroblocks (MBs) before actual encoding. Each MB from a P- or B-picture is subjected to a motion estimation process 102 in which forward motion vector(s), and backward motion vector(s) for the case of a B-picture MB, are determined using reference picture(s) from the frame buffer 101. With the determined motion vector(s), motion compensation 103 is performed where the intra- or inter-picture prediction mode of the MB is first determined according to the accuracy of the motion vector(s) found, followed by generating the necessary predicted MB. For the case of an I-picture MB or intra-mode MB, the predicted MB is the input MB (i.e. no prediction).
The predicted MB is subjected to discrete cosine transform (DCT) 104 and DCT coefficient quantization 105 based on quantization matrices (QM) and quantization stepsize (QS). Note
10
that for MPEG-2, field or frame based DCT can be selected in 104. The quantized DCT coefficients of the MB is then run-length encoded with variable length codes (VLC) at 106 and multiplexed with all necessary side information, for example selected motion vectors, MB coding modes, quantization stepsize, picture and sequence information, to form the ouφut bitstream. The bitstream is stored in an encoder ouφut buffer 107 for output at 108.
Local decoding is performed by inverse quantizing the quantized DCT coefficients (109), followed by inverse DCT 110, necessary motion compensation 103, and finally storing the reconstructed MB into frame buffer 101. Local decoding is performed such that the reference pictures used by motion compensation 103 is identical to those to be used by an external decoder.
The quantization stepsize (QS) used in 118 for quantizing the DCT coefficients of each MB has a direct impact on the number of bits produced at the ouφut of the run-length VLC encoding process 106, and therefore the average ouφut bit rate. It has also a direct impact on the encoding quality, which is ouφut picture quality at the corresponding decoder. In general, larger QS generates lower ouφut bit rate and lower encoding quality. In order to control ouφut bit rate and picture quality so that the resulting bitstream can satisfy channel bandwidth or storage limitations as well as quality requirements, rate control and quantization control algorithms are used.
Previously used methods for rate control and quantization control can be found in the MPEG- 2 TM5 (Test Model 5), which can be summarized in three steps. Step one is the bit allocation process 112 which assigns a target number of bits (T,, TP, or TB respectively) for a new I-, P- or B- picture to be coded. The target number of bits can be computed as follows:
^maxl R Bit Rate
I + N Pp Xp - . N BR X BR ^Picture Rate
X; Kp X, KB
1 1
R Bit Rate
7p=max|
NB Kp XB ' ^Picture _Rate
N„ KB Xp
R Bit Rate
-r £=max|
Np K„ Xp 8χ Picture Rate
NB+ p B p Kp XB
where
Λ = >->/ Qj, Λ.p = sip Qp, Λβ = ύg, g,
S„ Sp, Sg are number of bits generated by previously encoded I-, P-, B-picture respectively as computed by block 114,
Qι> Qp> QB are Λe average QS used by previously encoded I-, P-, B-picture respectively as computed by block 113,
Bit_Raie is the Target Bitrate supplied at 111,
Picture_Rate is number of pictures coded per second,
KP and Kg are constants (e.g. 1.0 and 1.4 respectively),
NP and NB are the number of P- and B-pictures remaining in the current GOP.
R is the remaining number of bits assigned to the GOP, where R is updated as follows:
After encoding a picture, R = R - Sl P B where S, P B is the number of bits generated in the picture just encoded (picture type is I, P, or B)
Before encoding the first picture in a GOP (an I-picture):
R = G+R G=Bit Rate* N/ Picture Rate
12
N is the number of picmres in the GOP.
In the second step, rate control (115) is performed. Basically, this step involves calculation of a reference quantization stepsize QSref for each MB based on the target bits (T,, TP, or TB) for the picture, the accumulated number of bits used from the target bits, and a virtual buffer model as given in MPEG-2 TM5.
Adaptive quantisation 118 is used in the third step for quantization control. In this step, the calculated QSref is scaled according to the local activities of the MB as generated by an MB activity calculation process 116, and the average MB activity of the previously coded picture (or optionally current picture) as produced by a frame activity average process 117. This scaling is done according to a level of masking effects of coding noise by human perception for MBs with high or low activities within a picture. As previously mentioned, an example of an adaptive quantization technique can be found in United States Patent number 5,650,860, "Adaptive Quantization".
The frame bit-count module 114 may contain a video buffer verifier (VBV) which is a virtual model of an input buffer of an external decoder. The VBV is maintained in such a way that underflow and overflow are prevented as required by the MPEG standard to ensure target bit rate 111 is maintained. Techniques for underflow detection and protection may also be found in US Patent number 5,650,860.
As mentioned, this form of encoding does not take account of the actual or varying complexity of the video scenes contained in the input picture sequence 100. Consequently, the perceptual quality of the resulting pictures may vary from good to annoying from scene to scene according to scene complexity. Variable bit rate (VBR) coding may be therefore be applied on individual segments of the picture sequences depending on its scene complexity to maximize bit rate allocation and encoded picture quality, and the bandwidth saving used for coding of more complex scenes, for example. However, existing VBR control algorithms such as that disclosed in US Patent 5.650,860 require multiple encoding passes to properly
13
distribute data bits. Multiple-pass VBR encoder require huge storage memory for storing the intermediate bit utilization information, and large computation needs for the additional passes and the bit budgets generation, and furthermore such techniques cannot process the input sequence in real-time.
Before encoding an input picture sequence, an encoder according to an embodiment of the present invention is set with maximum and minimum bit rate of target application (the minimum bit rate for some applications may be zero), and a target quality for encoded pictures. An initial target bit rate will then be set or generated. The encoder compresses the input sequence according to this target bit rate using conventional bit rate control method.
After encoding a picture or a certain number of pictures, the encoding quality (quality of encoded pictures) is measured against the target. If the encoding quality is lower than the target quality, the target encoding bit rate will be increased. On the other hand, if the encoding quality is higher than the target quality, the target bit rate will be decreased. The level of new target bit rate adjustment may depend on the current target bit rate and the predicted bit rate based on encoding history so that it is not over adjusted unnecessarily. The new target bit rate is also checked against the defined maximum and minimum bit rate.
The encoder will then re-adjust its rate control system with the new target bit rate, and continue to encode the rest of the input sequence. Furthermore, the encoder can ensure VBV compliant with existing methods using the defined maximum bit rate and/or the target bit rate. The process of determining new target bit rate based on encoding quality of past encoded picture(s) is repeated as needed.
In Figure 2 a block diagram of a one-pass variable bit rate moving pictures encoder 250 according to one embodiment of the present invention is illustrated. Those of ordinary skill in the art will recognise that a number of the functional blocks illustrated are common with the conventional MPEG encoder described above. The input moving picmre sequence 200 is processed picture by picture according to its possible assigned picture coding type (I-, P-,
14
or B-picture). A frame buffer 201 is used to temporarily store the input pictures according to re-ordering and processing needs. Each input picture is divided into Macroblocks (MB) before encoding. Temporal redundancy of each MB may be removed by motion estimation 202 and motion compensation 203.
After necessary motion compensation, a MB is subjected to discrete cosine transform (DCT) 204 and DCT coefficient quantization 205 based on quantization matrices and quantization stepsize. The quantized MB of I- or P-picture is inverse quantized 209, inverse discrete cosine transformed 210, and subjected to corresponding motion compensation 203 so that the quantized MB is reconstructed and stored locally in frame buffer 201 for future motion estimation and compensation needs. The quantized MB is run-length encoded with VLC (variable length coding) at 206, together with all necessary side information to form the encoded bitstream of the input sequence. The encoded bitstream is stored in an ouφut buffer 207 of the encoder for ouφut at 208 at desired data rates.
Methods of motion estimation 202, motion compensation 203, DCT 204, DCT coefficient quantization 205, run-length VLC encoding 206, inverse quantization 209, and inverse DCT 210 may be those defined and/or allowed in the ISO/IEC MPEG-1 or MPEG-2 standards.
Before encoding, a target quality of encoding (219) and maximum/minimum bit rate (211) are set. The maximum and minimum bit rates {BRmax and BR„in), given as bits per second, define the boundary bit rates which the encoder shall operate at, and the target quality defines the encoded picture quality that the encoder shall target by constantly adjusting a target bit rate within the given bit rate boundaries. The minimum bit rate may be set to zero.
Initially, target bit rate estimation 223 may be performed based the set target quality and the max/min bit rates to generate an initial target bit rate {BRlargel). Such initial estimate may be based on experimentally determined fixed values which may not directly relate to the picture(s) to be coded. With the target bit rate, the bit allocation 212 is performed to determine a target number of bits for the picture to be coded (TI P B). A rate controller 215
15
determines a reference quantization step size (QSrrj) for each MB based on the estimated number of bits for the picture (TI P B) and the bit utilization by the run-length VLC encoder 206. The QSref of each MB in the picmre to be coded can be computed, for example, as follows:
DI PS x Kι
~ ref (2 x BR target I picture _rate)
where DI P B is virtual buffer fullness of corresponding I-, P-, or B-picture, updated (after coding each MB) by the difference between the bits used by the MB and the bits allocated to the MB based on the corresponding Tl P B, a set of initial values for D,, DP, DB may be assumed at sequence start, and Kj is a constant (eg. 31).
Adaptive quantization 218 may be applied in which the determined QSref ι's scaled according to the local activities of the MB as generated by MB activity calculation process 216 and the average MB activity of the previously coded (or optionally current picture) as produced by the frame activity average process 217. Example implementations of the MB activity calculation 216, frame activity average 217, and adaptive quantization process 218 are found in MPEG-2 TM5. The ouφut quantization stepsize (QS) is used to quantized DCT coefficients of the MB.
After encoding a picmre, the number of bits (S1 P B , corresponding value for I-, P-, or B- picture) generated by encoding the picmre is accumulated by Frame Bit-Count 214, and the quantization stepsize (QS) is averaged by Frame Q Step Average 213 process. The encoded picmre quality is also determined by Frame Quality Measure 220 process. One method of determining the encoded picture quality is by the average value of the reference quantization stepsize (QSawrage) used for coding the picmre since it indicates roughly the amount of quantization noise in the encoded picmre. In this method, the target quality set at 219 is defined as the target reference quantization stepsize (QSωrjr«). Three different QSωrgfl values
16
may be set at 219 for corresponding I-, P-, and B-picmres; alternatively, the values can be determined by:
QS rg . for I-pictures = QSωrgel QSωrge, for P-pictures = KP * QS^, QS el for B-picmres = KB * QSurga
where KF and KB are constants which can be experimentally determined (example KF = 1.0 mά KB = 1.4).
A further condition may be applied to the final QSM el for B-picmres such that it is not lower than the QSaverπge of the last coded I- or P-picture so that bits can be efficiently used to improve the quality of anchor I- or P-pictures first before improving the quality of the B- pictures.
In the process of Frame Quality Measure 220 process, the target quality -2SMr^« 1S compared to the picmre quality QSav„age. Since a higher value of QSaverage implies higher quantization noise and therefore lower encoded picmre quality, when the value of QSaverage is found to be higher than QSarget, the difference of the two values will be used to increase the target bit rate (BRMrgel) at target bitrate estimator 222. On the other hand, when QSavemge is lower than QSMrga, BRlarge, is decreased at target bitrate estimator 222. Optionally, the value of QS^^, may be used at the rate controller 215 as a lower limit for the final ouφut QSref value so that when this target is reached (target quality reached), bits are saved for future encoding immediately. Hence, the ouφut reference quantization stepsize at rate controller 215 may be set according to:
if (QSref < QSurg , then QSrff =QSωrget .
For target bit rate estimation, a rate-quantization model for example one developed by Wei
Ding and Bede Liu, "Rate Control of MPEG Video Coding and Recording by Rate- Quantization Modelling" , IEEE Trans, on Circuit and Systems for Video Technology, Vol.
17
6, No. 1 , Feb. 1996, may be adopted. To avoid large estimation error or complex local fitting of the rate-quantization model, an alternative embodiment of the present invention may consist of a method of target bitrate estimation comprising the steps of predicting a bitrate (BRpredicled) at a Bitrate Predictor 221 based on bits used for encoding the last I-, P-, and B- picture, and estimating a new target bitrate (BRMrgel) at the target bitrate estimator 222 based on the said predicted bit rate and the difference between QSa„r„gt, and QSlarι/f.r The predicted bitrate (BRpredjaed) and the new target bitrate (BRωrgel) may be computed as follows:
BR predicted _(W (1 +« +„"βA) )
(OS - OS )
BRtargeA BRpredicιeX ^ * "T ' ^
'-■ average
if (BRurga > BR , then BRurgel = BR^ ϊf (BRurget< β ?TO , then BRuτg„ = BR^
where
S,, Sp, SB are number of bits generated by previously encoded I-, P-, B-picture, nP is the total number of P-pictures in the current GOP, nB is the total number of B-picmres in the current GOP, and K2 may be a constant, or a factor of BRpredictat, BR^, or BRurga.
The new target bit rate (BR^ can be computed after encoding each picmre or after encoding a certain number of pictures. The bit allocation process 212 and the rate controller 215 are updated with the new BR rge, value once it is determined. An embodiment of the bit allocation process 212 comprises the steps of:
a) Before encoding the first picmre in a group of picmres (GOP), determining R which
18
is the remaining number of bits assigned to the GOP:
R+= — xBR, „ then set BR , . , =BR picture _rate targe' °ldJarse' ,arge'
where:
N is the number of picmres in the GOP R = 0 before start of sequence, and
+ = is the accumulate function
b) Otherwise, if the picmre to be coded is not the first picmre of a GOP and a new target bit rate (BRωrg is determined, then updating R as given by:
R + ^e N e X (BR'^' ' **«--»> ' then S6t BR^ g t-B rget
where:
NP and NB are the number of P- and B-picmres remaining in the current GOP
c) Obtaining the average quantization stepsize (Q„ QP, or QB) from previously encoded picmres (I-, P-, or B-picmre respectively) as determined by Frame Q Step average process 213,
d) Obtaining from frame bit-count 214 the value of the number of bits (S„ SP, or SB) used to encode the past I-, P-, or B-picmres, and removing number of bits used by the last coded picmre (S, P B) from R, hence:
R - = S, if last coded picture is an I-picture
R -= Sp if last coded picmre is an P-picture
R -= SB if last coded picmre is an B-picmre
19
e) Finally, determining the target number of bits (T„ TP, or TB) to be used for coding the current picmre. T,, TP, or TB for respective I-, P-, or B-picmre are given, for example, by:
R
T
Np Xp ^ NB XB X, Kp Xj KB
T = R
NP KP X,
N - κB xp
T = R
ND Ka XP
Na κp xB
where
X] — S/ Qι> Xp ~ Sp QP, XB — SB, QB, Kp and Kg are constants. 0 A minimum value (Tmiπ) may be set as the lower boundary for T„ TP, TB. For example:
RR
T ■ = l^A. , whereK^ is a constant (e .8) mm K3* Picture _ Rate 3
Figure 3 is a block diagram illustrating another embodiment of a one-pass variable bit rate moving picmres encoder 350 according to the present invention. In this embodiment, the encoding quality 324 is determined by comparing the original input picmre taken from input
15 300 and the locally decoded picmre 325. Conventional methods such as the mean square
20
error (MSE) or the signal-to-noise ratio (SNR) measures can be used as comparison. This measure is further compared at Frame Quality Measure 320 with the definable Target Quality 319. When the encoding quality is lower than the target quality, the target bitrate estimator 322 will set the target bitrate higher; on the other hand, if the encoding quality is higher than 5 the target quality, the target bitrate is set lower.
An input picmre or a given number of input picmres are first encoded with an initial set of encoding parameters. Each of these picmres may be given a picmre coding type of either I-, P-, or B-picmre. The encoding technique such as those defined by the MPEG-1 or MPEG-2 10 standards may be utilized. Such techniques include motion estimation 302, motion compensation 303, discrete cosine transform 304, DCT coefficient quantization 305, and run- length VLC encoding 306 for temporal, spatial, and statistical redundancy reduction. A frame buffer 301 is used for storing the input picmres for re-ordering and processing needs.
15 Before each picmre is coded, its target amount of bits is set by bit allocation process 312 based on the initial set of coding parameters as well as any updated information from Frame Q Step averaging process 313 and Frame bit-count 314. Rate Controller 315 determines reference quantization stepsize for each MB while coding the picmre. Conventional adaptive quantization 318 which depends on MB activity calculation 316 and frame activity average
20 317 can be applied to each reference quantization stepsize to generate final quantization stepsize for the DCT Coeff quantizer 305.
The coded picmres are locally decoded by inverse quantization 309, inverse discrete cosine transform 310, inverse of the motion compensation 303 to form the locally decoded picmres
25 325. It is stored in the frame buffer 301 if necessary for fumre prediction use (ie. anchor pictures). The locally decoded picmres 325 are used with reference to the original input pictures 300 for encoding quality determination 324. The determined encoding quality is compared to the target quality given by 319 at frame quality measure process 320. It is also possible to offset the determined encoding quality by the picmre activity (ACTav„age) before
30 the comparison so as to take into account the masking effects of the human visual system to
21
the average level of activity within the picmre. For example if the encoding quality is determined by the peak signal-to-noise ratio, then the resulting encoding quality difference (SNRdiff) at the output of frame quality measure 320 can be given by:
SNRdiff = SNR,ar/!e! - (SNRp mr, + K5 x ACTmerttge)
where
SNRpiclure is the ouφut of encoding quality measure 324, SNR^, is the target quality definable at 319, K5 is a constant (eg. 0.018), and
ACTaveroge is the average MB luminance block variance.
One method of adjusting BRmτget based on SNRdjff at the target bitrate estimator 322 is given by:
BRiargei = BRp„diaed + 6 X BRcurre„, X SNR^
if (BRlargel> BR , then BR,argel = BR^ if (BR^ BRmiπ), then BR^, = BR^
where
BRpredicud is the result of bitrate predictor 321 ,
Recurrent ^ the previous value of -5/?Mrj-e, stored at current bitrate 323 memory,
K6 is a suitable constant, BR-max and BRmi„ are the maximum and minimum bitrate set at 311.
The bit allocation module 312 and the rate controller 315 take the updated BR„rgel from the target bitrate estimator 322 and perform necessary bit allocation and bit rate control such that the resulting compressed moving picmre bitstreams will have encoded bit rate close to BRM ι.,. Existing techniques of VBV (video buffer verifier) underflow detection and protection may
22
be applied using BR^ as the reference bit rate to ensure the maximum bitrate of the ouφut bitstream is not violated.
Because the target bit rate BR rg!, for encoding is adapted based on feedback on encoding quality, the encoding quality will always converge to a defined target quality. In such a way the encoding accommodates variable bit rates in real-time within single computational pass, and the encoding process can be better mned to changes in moving picmres scene content complexity. The overall percepmal quality of the encoded moving picmres sequence will be relatively consistent with reference to the defined target.
Embodiments according to the present invention can accommodate simplified models for estimation of target bit rate based on feedback and prediction. Overall impact to implementation and computation complexity as well as memory storage requirements of the encoder is minimized.
The foregoing detailed description of embodiments of the invention has been presented by way of example only, and is not intended to be considered limiting to the invention as defined in the claims appended hereto.