Decoding-Workload-Aware Video Encoding

Transcription

1 Decoding-Workload-Aware Video Encoding Yicheng Huang, Guangming Hong, Vu An Tran and Ye Wang Deartment of Comuter Science ational University of Singaore Law Link, Singaore Reulic of Singaore {huangyic, honggm, ABSTRACT This aer resents a novel decoding-workload-aware video encoding scheme. It takes raw video data and decoding workload constraint of a moile client as inut and generates a video itstream which matches such a constraint while striving to achieve the est video quality. For a given constraint, the est overall video quality of the encoded itstream is selected with a tradeoff etween satial and temoral distortions. The main contriutions of this aer include: 1) the roosal of an efficient scheme which selects the most suitale target frame rate efore the actual encoding; 2) The design of a workload control (analogous to the rate control) scheme which ensures an accurate control of the decoding workload when the itstream is generated using the roosed encoding scheme. Exerimental results demonstrate the feasiility and erformance of the roosed scheme. Categories and Suject Descritors I.6.3 [Simulation and Modeling]: Alications H.5.1 [Multimedia Information Systems]: Video General Terms Algorithms, Design, Exerimentation eywords Video Encoding, Workload Control, Frame Rate Selection 1. ITRODUCTIO The recent develoments in communication and microrocessor technologies have made moile devices ecoming attractive entertainment latforms for multimedia esecially video alications. However, suorting video alications on moile devices is challenging due to limited rocessing ower on the low to middle-end moile devices with rocessor frequency of MHz. Even with the fastest moile rocessor in the market today with rocessor frequency of around 600 MHz, it is still difficult to Permission to make digital or hard coies of all or art of this work for ersonal or classroom use is granted without fee rovided that coies are not made or distriuted for rofit or commercial advantage and that coies ear this notice and the full citation on the first age. To coy otherwise, or reulish, to ost on servers or to redistriute to lists, requires rior secific ermission and/or a fee. OSSDAV 08, May 28 30, 2008, Braunschweig, Germany. Coyright 2008 ACM /05/2008 $ $5.00. satisfy the decoding workload requirement of high quality video with a high frame rate. To address this rolem, we roose a decoding-workload-aware video encoding scheme to generate a video itstream which matches the workload constraint of the target moile devices. The analysis in our revious work on decoding workload model [1] shows the relationshi etween arameters in the video itstream and decoding workload. Given the arameters such as the numer of Huffman codes, Macrolock (MB) tye and motion vectors recision, the decoding workload can e estimated accurately using the roosed decoding workload model. This insires us to design a decoding-workload-aware video encoding scheme, which controls the decoding workload from the encoder side y adjusting these arameters during the encoding rocess. The concet of workload control is analogous to the rate control which controls the target itrate to satisfy the andwidth/storage constraint. The roosed workload control scheme is imlemented with an MPEG-2 video codec for the roof of concet. For video encoding, the frame rate lays a significant role on the decoding workload. Intuitively, a video itstream with the same image quality ut a lower frame rate has a lower decoding workload. For a conventional video encoder, the target frame rate is tyically fixed at 25 or 30 fs. However, moile devices with low rocessing ower might not e ale to decode a video itstream with such a high frame rate in real time. To reduce the decoding workload, we can either reduce the frame rate or reduce the quality of the individual frame. The rolem is, given a decoding workload constraint, there can e more than one candidate with different cominations of frame rate and individual frame quality. The challenge is how to choose a candidate with the est overall quality. Conventional ojective video quality measures such as MSE and PSR cannot comare quality etween video clis with different frame rates. A ossile solution roosed in [2, 3] is to relace the droed frame y its revious frame in dislay order and calculate average PSR or MSE of the new video sequence as the video quality. The rationale ehind such aroach is that a video layer tyically maintains the current frame on the screen efore dislaying the next frame. With this method, oth satial distortion (image quality) and temoral distortion (frame rate) are considered. This aroach, however, is rolematic for our roosed encoding scheme, ecause it requires encoding, decoding and calculating PSR/MSE for every candidate of all ossile frame rates. This could ecome rohiitively exensive comutationally. To solve this rolem, we roose a new frame rate selection scheme which rovides a fast estimation of the resulted distortion for every candidate efore the actual encoding. This is followed y a workload control scheme which accurately 45

2 determines the decoding workload of the generated video itstream when eing decoded on the targeted clients. The architecture of the roosed decoding-workload-aware encoder is shown in Figure 1. Raw Video Data Client Decoding Workload Constraint Frame Rate Selection Scheme Target Frame Rate Encoding with Workload Control Figure 1: The encoder architecture Encoded Video Bitstream The encoding rocedure includes two hases. In the first hase, the raw video data and client decoding workload constraint together with all ossile frame rates are fed into the Frame Rate Selection Scheme which selects the most suitale frame rate for the actual encoding. In the second hase, the encoder uses the selected frame rate and client decoding workload constraint to comress the raw video into the target video itstream using the Workload Control Scheme. The rest of the aer is organized as follows: Section 2 introduces the workload control scheme. Section 3 resents the frame rate selection scheme. Section 4 includes evaluations of the roosed schemes and Section 5 concludes this aer. 2. WORLOAD COTROL SCHEME According to the decoding workload model roosed in [1]: an MPEG video is made u of a sequence of Macrolocks (MBs), and modeling the video decoding workload is decomosed to modeling three major tasks of decoding one MB: Variale Length Decoding (VLD), Inversed Discrete Cosine Transform (IDCT) and Motion Comensation (MC). The decoding workload of VLD is modeled as a linear function of the numer of Huffman codes. The decoding workload of IDCT is modeled as a looku tale indexed y the last osition of Huffman codes. The workload of MC is modeled as a looku tale indexed y motion vectors recisions. And for different tyes of MB, the arameters of the models can e different. Thus, we can control the decoding workload y adjusting the numer of Huffman codes, MB tye and motion vector recision. The workload control can work at three levels: frame level, MB level or task level. For an encoder, workload control in frame and MB level is similar to the conventional it rate control. However, in task level, rate control and workload control are significantly different. Rate control scheme only considers the quantization level of DCT coefficients, which is roortional to the it rate. Workload control needs to consider multile factors and their tradeoff. For examle, if we allocate more workload to the VLD or IDCT task, we can have more Huffman codes. This increases the video quality. However, allocating more workload to the VLD or IDCT task will result in allocating less workload to the MC task. This may limit the motion vectors recision and thus decreasing the video quality. This rolem ecomes even harder if we also consider the MB tye. In this section, as the first ste of our research on the decoding-workload-aware encoder, we do not consider the task level workload control: we simly fix the workload of the MC task y fixing the MB tye and motion vectors from the conventional motion estimation rocedure. We only control the decoding workload of IDCT and VLD tasks y adjusting the numer of Huffman codes. We design two strategies for the frame level and MB level workload control, resectively. The strategies allocate the workload so that the encoded itstream can have a etter video quality within the constraint of decoding workload. The two strategies can e summarized as follows:! In frame level, the workload is allocated ased on statistical ratio of different frame tyes and the ratio is adjusted according to the recent history.! In MB level, the workload is allocated ased on the image comlexity which can e estimated y the variance or MSE. The exeriment results in Section 4 show that these two strategies can imrove the video quality. Algorithm 1 descries how the workload control scheme works. 1) Allocate the workload for the current GOP according to the constraint, GOP size and frame rate on an average asis. 2) For all the frames in the GOP: 3) Allocate the workload to the current frame according to the frame tye and history record. 4) Run motion estimation for all the MBs of the current frame, decide their MB tyes, record their MSEs (or VAR for I-MB) and motion vectors. 5) Estimate the workload of MC for all the MBs ased on the results from 4) using the workload model. 6) For all the MBs in the current frame: 7) Allocate the workload of current MB y its MSE/VAR and MB tye. 8) From the motion vectors get in the line 4, estimate the workload of VLD+IDCT for all ossile numer of quantization scales using workload model. Select out the numer of quantization scale having the workload closest to W VLD_IDCT =W m -W MC, where W m is the workload allocated for the MB, and W MC is the workload of MC get in line 5. 9) Encode the MB. 10) Udate the status. The details are as follows: Algorithm 1: Workload Control Scheme In line 3, the decoding workload for current frame (W i, W and W for I-frame, P-frame and B-frame) is allocated as: Wi! W /(1 " " ) i P i W! W /( " ) W! W ( " ) where i, and are the decoding workload for the revious I-, P- and B-frames; and are the arameters reresenting the (1) (2) (3) 46

3 ratio etween I-, P- and B-frame. In our imlementation, =1.0 and =2.0, which are otained emirically. W is the remaining workload of the GOP, which is udated after encoding a frame., are the numer of P-and B-frames in a GOP. At the eginning, i, and are initialized as ( & i! W 1" '! W! W ( & ' " " ( & " ' % # $ % # $ % # $ In line 4, we use conventional motion estimation, with which the MB tye is decided y comaring the MSE or VAR of the MB with a constant threshold. In line 7, workload of the current MB, W m (i) is allocated as * * W ( i)! ( W ) W ( j)) * MSE( i) / MSE( j) " W ( i) m frame MC MC j! i j! i where W frame is the remaining workload of the current frame; MSE(i) is the MSE of the i th MB (or VAR(i), if the i th MB is an I-MB); is the numer of MBs of the frame. W MC (i) is the workload of MC of the i th MB. The rationales ehind this equation are: 1) as mentioned earlier, we do not change motion vectors recision or MB tye after motion estimation, the workload of MC can e regarded as fixed; 2) we allocate more workload to VLD and IDCT tasks of the MB which has larger MSE/VAR. A MB with larger MSE/VAR imlies more residual error. Therefore, it requires more Huffman codes for the encoding. And to decode a MB with more Huffman codes, more decoding workload is required in VLD and IDCT tasks. In line 9, we use the quantization scale otained from line 8 for the actual encoding (generating the itstream). If the encoder also emloys a rate control scheme, we will get another quantization scale for the rate control. In this case, oth the decoding workload constraint and it rate constraint can e satisfied y selecting the larger quantization scale. In line 10, we estimate the decoding workload using the arameters extracted from the encoded MB and udate status of the scheme. To summarize, in the frame level, we allocate the workload ased on statistical ratio etween different comonents which is udated with a moving average of recent history; in the MB level, we allocate the workload ased on the image comlexity. The exeriment results in Section 4 show that these two strategies imrove the video quality consideraly. 3. FRAME RATE SELECTIO SCHEME This section resents our fast frame rate selection scheme: it enumerates all the frame rate candidates; for each candidate, the distortion of the target video itstream is estimated. The candidate with the smallest distortion is selected. The rolem is how to erform a fast distortion estimation of all target video itstreams with all ossile frame rates efore actual encoding. (4) Before going to the detail of the algorithm, we introduce some notations first. Assume we have a raw video sequence containing frames: P(0), P(1), P(2) P(-1) (see Figure 2). For each frame rate candidate f, we evenly select M=*f/f max frames from the original sequence for actual encoding, where f max is the maximum frame rate. In our imlementation, f max is set as 25fs. We denote P (0),P (1),P (2),,P (M-1) are the frames decoded at the client end. In Figure 2, f is equal to 12 fs. Relacing a droed frame y its revious frame, we get the frame sequence P (0,0),P (0,1)..P (0,f max /f-1),p (1,0),P (1,1) P (1,fmax/f-1) P ( M-1,0),P (M-1,1) P (M-1, f max /f-1), where P (i,j) is exactly the same as P (i,0). And P (i,j) is corresonding to the frame P(i*f max /f+j) in the original video sequence. Original f=12 fs After relacement P(0) P(1) P(2) P(3) P(4) P(-2) P(-1) P (0) P (1) P (2) P (M-1) P (0,0) P (0,1) P (1,0) P (1,1) P (2,0) P (M-1,0) P (M-1,1) Figure 2: An examle for illustrating frame rate selection scheme The distortion of the video sequence with frame rate f is calculated y the distortion of the video sequence after frame relacement, which is then calculated y the average distortion of the corresonding frames. Here are the asic ideas to calculate the distortion etween two corresonding frames: For the frames P (i,0) According to the frame rate and GOP structure, we know the frame tye of P (i,0). Using a simle version of workload control scheme in Section 2, we can also estimate the numer of Huffman codes in this frame. If P (i,0 )is an I-frame, The variance of the frame descries its image comlexity. The distortion etween P (i,0) and P(i* f max /f) is estimated as the image comlexity lost when eing encoded into the target itstream due to workload constraint, i.e., a MB will e coded using just a art of the total 64 Huffman codes: whuff ( ) D( i,0)! Var( i * fmax / f )*(1 ) ) (5) w (64) huff where is the numer of Huffman codes, W huff () is the weight of the first Huffman codes, Var(i) is the variance of the original frame, which can e calculated efore the actual encoding. If P (i,0) is a P-frame, A P-frame is deendent on its reference frame. Assuming the reference frame is P (k), its distortions have two arts: the distortion roagated from its reference frame and the residual error lost due to the workload constraint: whuff ( ) D( i, 0)! W ro * D( k, 0) " Res'( k, i) * (1 ) ) (6) w (64) huff 47

4 where W ro is the weight reresenting the error roagation effect, D(k, 0) is the distortion of P (k) which can e calculated y Eq. 5 (if P (k) is an I-frame) or Eq.6 (if P (k) is a P-frame). Res (k,i) is the residual error etween P (k) and P (i). The residual error is calculated in the motion comensation rocedure. Since running the motion comensation for all frame rate candidates is comutationally exensive, we estimate Res (k,i) y Re s'( k, i)! Re s(, q)! Re s(, " 1) " W *(Re s( " 1, q)) (7) where =k* f max /f, q=i* f max /f, W res is a arameter. Thus, we can estimate the residual error etween any two frames y a linear comination of the residual error etween two adjacent original frames, which needs to e calculated only once efore the actual encoding. If P (i,0) is a B-frame It deends on two frames P (k) and P (t). Similar to the P-frame, its distortion can e calculated as: D( k,0) " D( t,0) Re s'( k, i) " Re s'( i, t) whuff ( ) D( i,0)! Wro * " *(1 ) ) (8) 2 2 w (64) For the frames P (i,j), where j>0 res huff The detail of the distortion estimation for each frame rate candidate is shown as Algorithm 2. 1) Select the frames, P (i,0),i=1 M, from original sequence ased on the frame rate. 2) For all the selected frames: 3) Allocate the workload to the current frame according to the frame tye. 4) Estimate the sum of the workload of MC, W MC of all the MBs in the current frame 5) Estimate the sum of the workload of IDCT+VLD, W VLD_IDCT =W- W MC, where W is the workload of the frame. 6) Estimate the average numer of Huffman coefficients of MB in the frame. 7) Estimate the distortion D(i,0) etween P (i,0) and P(i* f max /f). 8) Relace the droed frame y its revious frame. For all the relaced frames, P (i, j),i=1 M, estimate the distortion D(i,j) etween P (i,j) and P(i* f max /f+j), j=1.. f max /f. 9) Calculate Avg(D(i,j)) as the overall distortion of the target video sequence. P(i* f max /f) MSE P(i* f max /f+j) Algorithm 2: Frame Rate Selection Scheme D(i,0) P (i,0) (1-W)*D(i,0)+W*MSE P (i,j) Figure 3:The distortion calculation for P (i,j) The distortion etween P (i,j) and its corresonding frame P(i* f max /f+j) equals to the distortion etween P (i,0) and P(i*f max /f+j), since P (i,j) is a direct coy of P (i,0). Using the frame P(i*f max /f) as a ridge (see Figure 3), the distortion can e estimated as a weighted sum of the satial distortion etween P (i,0) and P(i* f max /f), and the temoral distortion etween P(i* f max /f+j) and P(i* f max /f). D(i,j) = (1-W tem )*D(i,0)+W tem*mse(i*f max/f, i*f max/f+j) (9) where W tem is the weight of the temoral distortion and MSE(i*fmax/f, i*fmax/f+j) is the MSE etween P (i*fmax/f) and P(i*fmax/f+j), which is used to reresent the temoral distortion caused y the frame relacement in the dislay sequence. Again, we do not want to calculate MSE for all the ossile candidates. Let =i*f max /f and q=i*f max /f+j, we estimate MSE(,q) y MSE(, q)! MSE(, " 1) " W * MSE( " 1, q) (10) mse Thus, we estimate the MSE etween any two frames y a linear comination of the MSE etween two adjacent original frames, which needs to e calculated only once efore the actual encoding. The details are as follows: In Line 3, frames with different tye are allocated using different ratio. We kee the ratio the same as that in Algorithm 1: 2:2:1 for I-, P- and B-frame. In Line 4, we assume all MBs in I-frame are I-MBs and 1/3 MBs in P-frame are I-MBs and another 2/3 MBs are P-MBs. We also assume 1/2 MBs in B-frame are P-MB and another 1/2 MBs are B-MBs. Based on this ratio, we estimate the sum of the workload of MC of the MBs in the current frame. It should e noted that aove aroach is not the most accurate one. A more accurate aroach can emloy the residual error to estimate the numer of I-, P- and B-MBs of the frame. However, our simle scheme is designed to select the est frame rate. Exerimental results show that this simle aroach works sufficiently well. For I-, P- and B-MB, we use a constant value to estimate the MC workload. The constant value is otained from statistical analysis. Again, this is not the most accurate aroach, ut is sufficient for our urose. In Line 6, the numer of Huffman codes is estimated using the decoding workload model in [1]. In Line 7, the distortion D(i,0) is calculated as Eq 5, 6 and 8. In Line 8, the distortion D(i,j) is calculated as Eq 9. In the roosed scheme we have many arameters such as W huff (), W tem, W ro, W mse and W res. They are all otained from the statistical analysis: W huff () is otained from the exeriment where we select a 8*8 lock from a raw icture, erforming the DCT oeration, setting the coefficients after osition as zero and finally calculating the difference etween the original lock and the lock after IDCT. For W tem, W ro, W mse and W res, we use a set of video as the training set. We enumerate the four arameters from 0~10 with a ste of 0.1 and select the values with est estimation result. 48

5 4. EVALUATIO 4.1 Workload Control Scheme Evaluation Exerimental Setu For roof of concet of the roosed decoding-workload-aware video encoding, we emloy the MPEG-2 as the video format. We modify MPEG-2 reference encoder to a decoding-workload-aware encoder. In our exeriments, we select 12 raw video sequences which are shown in Tale 1. Each of them is encoded under 11 workload constraints: 20 MHz, 30 MHz, 40 MHz, 50 MHz, 60 MHz, 80 MHz, 100 MHz, 120 MHz, 150 MHz and 200 MHz. We use MPEG-2 decoder of TCPMP roject [4] as the target decoder. We use SimleScalar [5] to simulate the decoding rocedure and record the actual decoding workload, which is then comared with the workload constraints. o Video ame Descrition 1 akiyo Still ackground and a foreground oject with very low 2 ridgeclose Still ackground and some small ojects with random 3 ridgefar Almost a still image. 4 coastguard Still ackground and two foreground ojects with contrary 5 container Still ackground and two foreground ojects with same 6 foreman Background and foreground have moderate 7 hall Still ackground and two ojects with moderate 8 highway Background with very fast 9 mother-daughter Still Background and two ojects with very slow 10 news Still ackground, an oject with fast movements and two ojects with very low 11 silent Still ackground and an oject with moderate movements 12 walk Both ackground and two foreground ojects are with very fast movements Tale 1: 12 raw video sequences Exerimental Results Figure 4 shows the comarison etween the constraint and the actual decoding workload for the sequence akiyo. We run 10 exeriments, where we set the workload constraint as 20, 30, 40, 50, 60, 80, 100, 120, 150 and 200MHz, resectively. The results of the other sequences also show similar matches. Two curves in the figures reresent the constraint and the actual decoding workload, resectively. It can e oserved that, in most cases, the actual workload is very close to the constraint. However, when the constraint is very low (20MHz), the actual decoding workload is more than the constraint as seen in Figure 4. It is ecause each sequence has a minimum decoding workload requirement which deends on the video content. For examle, when the motion of the video is large, the residual error in P- and B-frame is large as well, which demands more decoding workload. Thus the minimum decoding workload requirement will e large, and vice visa. The average difference etween the constraint and actual decoding workload is less than 1.8 %. This indicates that the workload control scheme controls the decoding workload very well. >%;"?%@25AB8CD )'& )&& ('& (&& '& &!"#$% E%1=7;@#17!F7G@?5H%;"?%@2 ( ) * + ', -. / (& % :3;#<317= Figure 4: The comarison etween the constraint and actual decoding workload for sequence akiyo. MSE Hall Workload Constraint (Mhz) fix_fix his_fix his_mse Figure 5: The comarison etween the video distortions etween different workload control schemes for the sequence hall. ext, we evaluate the strategies we emloyed in the workload control scheme. In Figure 5 we comare the video quality of the itstream generated y our scheme with the itstreams generated without emloying the strategies for the video sequence hall. The x-axis reresents the workload constraint and the y-axis reresents the MSE of the encoded video itstream. The results of the other video sequences show similar atterns. In Figure 5, the curve his_mse reresents MSE value of the itstream generated y the scheme using oth strategies descried earlier. The curve his_fix reresents the MSE value of the itstream generated y the scheme only using the strategy in frame level; in the MB level, we allocate workload according to a fixed ratio. And the curve fix_fix reresents the MSE value of the itstream generated without using any strategy, i.e., we allocate workload ased on fixed ratios in oth frame and MB level. It is oserved that, under the same constraint, the itstream generated y using oth strategies has etter quality than the itstream generated y using only one strategy on the frame level, which is still etter than the itstream generated without using any strategy. Our exerimental results show that oth workload allocation strategies in the frame and MB levels are effective. 4.2 Frame Rate Selection Scheme Evaluation Exerimental Setu In the exeriment, given a raw video sequence and the workload constraint, we select the est frame rate from the candidates using the frame rate selection scheme. To evaluate the result, we encode 49

6 and decode the sequence under the same workload constraint for all the frame rate candidates. Then, the droed frames are relaced with the revious un-droed frames and the average MSE is calculated. We evaluate if the frame rate selected y our scheme has the smallest MSE. In the exeriment, we use 12 different video sequences and 14 workload constraints: 10, 20, 30, 40, 50, 60, 80, 100, 120, 150, 180, 200, 250 and 300 MHz; and frame rate candidates are 5, 10, 15, 20 and 25 fs. I3=75;@<35L@735A6:=D *& )' )& (' (& ' & I;#2J3E?%=3 MG;5F83<3 B9 (& )& *& +& '&,&.& (&& ()& ('& (.& )&& )'& *&& >%;"?%@25E%1=7;@#175AB8CD Figure 6: The comarison etween our scheme and MSE for the sequence ridgeclose Exeriment Results Figure 6 shows the comarison etween our scheme and MSE for the sequence ridgeclose. The results for the other sequences show similar atterns. It is oserved that our scheme and MSE match well in most cases. The ercentage that the frame rate selected y our scheme has the smallest MSE value is 74.4%. The ercentage that the frame rate selected y our scheme has the smallest or second smallest MSE value is 90.4%. Furthermore, the cases our scheme does not match the MSE, are ossily ecause MSE does not reflect the video quality accurately. For examle in Figure 6, when the workload constraint is 100 and 120 MHz, MSE selects the est frame as 25fs. However, when the workload constraint increases to 150, MSE selects the est frame as 15fs which is counter-intuitive. The est frame rate should not decrease with the increase of workload constraint. The case shown in Figure 6 illustrates that our scheme is more reliale than the conventional MSE. 943FG735P#<35A=3F%12D )&& (.& (,& (+& ()& (&&.&,& +& )& & B9 MG;5F83<3 ( ) * + ', -. O5%65;@<35L@735E@12#2@73= conventional aroach, we have to encode, decode and calculate MSE for n times, where n is the numer of the frame rate candidates; while in our scheme, we run the motion estimation (a art of the encoding rocess), calculate MSE and variance only once. A comarison of time comlexity of the two schemes is shown in Figure 7. The test was run on a deskto with Pentium 4 CPU and 1G RAM running Windows P. As shown in Figure 7, the execution time increases with the numer of frame rate candidates for the conventional aroach, while the execution time for the roosed scheme is almost constant. When the numer of frame rate candidates is 8, our scheme is aout 25 times faster than the conventional aroach. 5. COCLUSIO In this aer, we have resented a new decoding-workload-aware video encoding scheme with two main contriutions: a decoding workload control scheme and a fast frame rate selection scheme. The workload control scheme can control the decoding workload accurately when the generated video itstream using the roosed scheme is decoded in a target client. The fast frame rate selection scheme can select out the most suitale target frame rate, alancing the satial and temoral distortions, efore the actual encoding. We elieve that the roosed fast frame rate selection scheme is not only useful for workload control ut also for rate control. On the other hand, our workload control scheme still has a lot of room for imrovement. For examle, the workload allocation in the task level is an imortant and interesting rolem to study in the future. Another exciting future work lies on the relationshi etween the decoding workload and rocessor energy consumtion. We can use the same rincile resented in this aer to control the energy consumtion level of the video decoding devices via dynamic voltage/frequency scaling for the urose of extending attery life or simly matching the current attery level. 6. REFERECES [1] Y. Huang, V. Tran, Y. Wang, "A Workload Predication Model for Decoding MPEG Video and its Alication to Workload-scalale Transcoding", ACM Multimedia Conference, , Setemer, [2] M. Bonuccelli, F. Lonetti, F. Martelli, Temoral Transcoding for Moile Video Communication, the second Annual International Conference on Moile and Uiquitous Systems: etworking and Services, , March, [3]. gan, T. Meier, Z. Cheng, Imroved Single-video Oject Rate Control for MPEG-4, Circuits and Systems for Video Technology, IEEE Transactions on, , May, [4] htt://tcm.corecodec.org. [5] T. Austin, E. Larson, D. Ernst, Simlescalar:An infrastructure for comuter system modeling, IEEE Comuter, , Figure 7: The comlexity comarison etween the two schemes In comarison with the conventional aroach, such as MSE, our scheme has a much lower comutational comlexity. If we use the 50