In this chapter we study the problem of dependent bit allocation. addressed in the literature so far has been conæned to coding environments where

Transcription

1 105 Chapter 4 Optimal Bit Allocation for Dependent Quantization 1 Contents 4.1 Introduction : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : Dependent coding problem formulation : : : : : : : : : : : : : : : : : Temporal dependency: the MPEG case : : : : : : : : : : : : : : : : : Buæer constrained allocation for dependent coders : : : : : : : : : : : Appendix: Proof of Theorem 1 : : : : : : : : : : : : : : : : : : : : : Introduction In this chapter we study the problem of dependent bit allocation. All the work addressed in the literature so far has been conæned to coding environments where the input signal units èe.g. image blocks or subbandsè have been coded independently. However, many popular schemes èe.g. DPCMè involve dependent coding frameworks, i.e. where the set of available RíD operating points for some coding units èfig. 1 Part of this work has been done jointly with Kannan Ramchandran. For related publications see ë85, 72,84,86ë

2 èbèè depends on the particular choice of RíD point for other coding units èfig. 4-1èaèè. Other dependent coding examples include multiresolution èmrè coders like the Laplacian spatial pyramid ë11ë and the closedíloop spatio-temporal pyramid video coder ë108ë, as well as the MPEG ë58ë coder. Refer to Fig. 4-2 for an overview of where dependent coding schemes æt into typical source coding environments. In this chapter ë85, 84, 86ë, we generalize the bit allocation problem to include temporally dependent coding blocks èsee ë85, 86ë for the case when dependent allocation in multiresolution scheme is consideredè. As in ë101ë, we make no assumptions about the input or quantizer characteristics, and deal with arbitrary input signals and arbitrary discrete admissible quantizer sets. An alternative, model-based, solution of the dependent bit allocation problem has been proposed in ë107ë. As seen from Fig. 4-1 for the twoíframe case, the number of possible operating points for each frame grows exponentially in the dependency tree depth, making the problem very diæcult in the general case èe.g. DPCMè. However, when the dependency tree is structured, as in MPEG, with several ëleaves" or ëterminal nodes" èe.g. B-frames of MPEG, as will be seen in Section 4.3è, then it is possible to solve the diæcult dependent problem elegantly, by describing how to formulate intelligent pruning rules to eliminate suboptimal operating points. Moreover, a number of important applications èlike CD-ROM storageè can aæord considerable oæ-line complexity,thus making our proposed approach relevant. This chapter is organized as follows: in Section 4.2, we provide a Lagrangianí based solution for an arbitrary set of quantizers for each coding block. We point out the complexity of this approach, and introduce a certain monotonicity property of the operational RíD curves of the signal blocks to help reduce the complexity of the search for the optimal solution. We address the temporally dependent coding scheme in detail in Section 4.3 using MPEG as an example, and show how the monotonicity

3 107 property introduced earlier can be used to formulate fast heuristic solutions. In Section 4.4, we brieæy outline how these techniques could be extended to include buæer constraints as well. 4.2 Dependent coding problem formulation In this section, we deæne a general dependent allocation problem, show how this general formulation is applicable to MPEG and multiresolution coders, and give a solution based on Lagrange multipliers. Before we introduce the dependent coding problem, let us review the optimal independent allocation case which has been studied in the literature ë50, 101, 113ë and which was used in Chapter 2. The formulation is reviewed here for convenience and also to introduce a notation that is more adequate for the speciæc problems discussed in this chapter Optimal independent allocation í the constant slope condition The classical rateídistortion optimal bit allocation problem consists of minimizing the average distortion D of a collection of signal elements or blocks subject to a total bit rate constraint R budget for all blocks. For the two-block case, where fq1;r1èq1è;d1èq1èg and fq2;r2èq2è;d2èq2èg refer to the quantizer, distortion and bit rate of each block respectively, the independent allocation problem is: min ëd1èq1è+d2èq2èë Q 1 ;Q 2 è4.1è such that R1èQ1è+R2èQ2èçR budget : è4.2è The ëhard" constrained optimization problem of è4.1è,è4.2è can be solved by being converted to an ëeasy" equivalent unconstrained problem. This is done by ëmerging"

4 108 rate and distortion through the Lagrange multiplier ç ç 0 ë101ë, and ænding the minimum Lagrangian cost J i èçè = min Qi ëd i èq i è+çr i èq i èë for i =1;2. The search for the optimal RíD operating points for each signal block can be done independently, for the æxed quality ëslope" ç èwhich trades oæ distortion for rateè because at RíD optimality, all blocks must operate at a constant slope point ç on their RíD curves ë101, 87ë. The desired optimal constant slope value ç æ is not known a priori and depends on the particular target budget or quality constraint, but can be obtained via a fast convex search ë87ë. See Section for more details General formulation In the more general case, signal units may not be independently coded. Without loss of generality,we ærst consider a 2-level dependency as in Fig Shown are the RíD characteristics for a given discrete set of quantization choices for the ærst independent frame èr1èq1è;d1èq1èè and the second dependent frame èr2èq1;q2è;d2èq1;q2èè. Our constrained optimization problem is: what quantization choice do we use for each frame such that the total èor averageè distortion is minimized subject to a maximum total bit budget constraint? We model the total distortion as a weighted average of the individual distortions D1 and D2 in our general formulation. We will show how diæerent choices of the weights lead to diæerent problems of interest. Our problem can be formulated as: min ëw1d1èq1è+w2d2èq1;q2èë è4.3è Q 1 ;Q 2 such that R1èQ1è+R2èQ1;Q2èç R budget : è4.4è Note that although Q1 and Q2 here represent frameílevel quantization choices, it does not imply that all blocks within the frame have the same quantization scale.

5 109 Thus Q1 could consist of a vector choice of diæerent quantization scales for each block of frame 1. Also note that for arbitrary choices of w1 and w2, wehaveaweighted mean-squared-error èmseè criterion, which reduces to the conventional èunweightedè MSE measure when w1 = w2 =1. D 1 3 D 2 Slope =λ J 2 (3,2) J 2 (2,2) Slope =λ J 1 (2) 2 J 2 (1,2) 1 R 1 R 2 (a) (b) Figure 4-1: Operational R-D characteristics of 2 frames in a dependent coding framework, where frame 2 depends on frame 1. èaè Independent frame's R-D curve. èbè Dependent frame's R-D curves. Note how each quantizer choice for frame 1 leads to a diæerent R2, D2 curve. The Lagrangian costs shown are J = D + çr for each frame Examples We now provide examples of problems of interest which follow as special cases of è4.3è,è4.4è. See Fig. 4-2.

6 Figure 4-2: Overview of typical source coding environments. Example 1. Independent coding: The independent case seen in Section is a special case of è4.3è,è4.4è, where frame 2 does not depend on frame 1, i.e. R2èQ1;Q2è = R2èQ2è and D2èQ1;Q2è = D2èQ2è. The independent coding case arises for intraframe coding as well as pyramidal coding without quantization feedback ë11, 85, 86ë. This was also the situation considered throughout Chapter 2. Example 2. Spatially dependent pyramid coding with quantization feedback: When w1 = 0;w2 = 1,wehave the case of a two-layer closed loop èquantization feedbackè pyramid, where the bottom layer èlayer 2è depends on the quantization choice of the top layer èlayer 1è ë85, 86ë. Note that è4.4è refers to a total bit rate constraint, and we solve the fullíresolution quality optimization problem only. In

7 111 addition, for a multiresolution coding environment, it may be necessary to throw in additional constraints on the top layer bit rate ë11, 85, 86ë: R1èQ1è ç R 0 1 ; è4.5è or quality D1èQ1è ç D 0 1 : è4.6è Example 3. Temporally dependent coding: This is the most general case where è4.3è,è4.4è apply without any restrictions. DPCM and MPEG come under this class Solution based on Lagrange multipliers The problem of è4.3è,è4.4è can be solved by introducing the Lagrangian cost J corresponding to the Lagrange multiplier ç ç 0 as in ë101ë as follows: J1èQ1è = w1d1èq1è+çr1èq1è; è4.7è J2èQ1;Q2è = w2d2èq1;q2è+çr2èq1;q2è; è4.8è and considering the following unconstrained minimization problem: min Q 1 ;Q 2 ëj1èq1è+j2èq1;q2èë: è4.9è Then, by a direct extension of Theorem 1 of Shoham and Gersho in ë101ë, the following result follows: Theorem 4.1 If èq æ 1 ;Qæ 2è solves the unconstrained problem of Eq. è4.9è, then it also solves the constrained problem of Eqs. è4.3è, è4.4è for the particular case of R budget = ër1èq æ 1è+R2èQ æ 1 ;Qæ 2èë.

8 112 Proof: See Appendix 4.5. The above result implies that if we solve the unconstrained problem of è4.9è for a æxed value of ç, and if the total bit rate happens to be R budget, then we have also optimally solved the constrained optimization problem of è4.3è, è4.4è. Further, as ç is swept from 0 to 1, one traces out the convex hull of the composite RíD curve of the dependent allocation problem. The monotonic relationship between ç and the expended bit rate ë101ë makes it easy to search for the ëcorrect" value of ç, sayç æ, for a desired R budget. Note how for the independent case èj2èq1;q2è=j2èq2èè, where there is a single R2, D2 curve in Fig. 4-1, è4.9è becomes the familiar result of ë101ë èsee also Chapter 2è. Here, each frame is minimized independently, aswas shown in Section For the general dependent case, the 2-frame problem becomes the search for Q æ 1 ;Qæ 2 that solve: J1èQ æ 1è+J2èQ æ 1 ;Qæ 2è = min Q 1 ;Q 2 ëj1èq1è+j2èq1;q2èë è4.10è = min Q 1 ëj1èq1è+j2èq1;q æ 2èQ1èèë; è4.11è where J2èQ1;Q æ 2èQ1èè = min Q2 ëw2d2èq1;q2è+çr2èq1;q2èë is the minimum Lagrangian cost èfor quality condition çè associated with the dependent frame when the independent frame is quantized with Q1. See Fig Thus, for the desired operating quality ç, we ænd the optimal solution by ænding, for all choices of Q1 for the independent layer, the optimal èq æ 2èQ1èè which ëlives" at absolute slope ç on the èdependentè R2íD2 curve associated with Q1. The important point to note is that here we cannot treat each of the blocks separately as in Section and thus minimizing è4.9è actually requires, if no further assumptions are made, an exhaustive search for the minimal lagrangian for

9 113 all possible combinations of Q1;Q2. Fig. 4-3 shows an example where a greedy choice of quantizer for the ærst frame, i.e. choosing the quantizer that minimizes J1èQ1è would have resulted in overall suboptimality. In general we will have that: min ëj1èq1è+j2èq1;q2èë 6= minëj1èq1èë + minëj2èq æ Q 1 ;Q 2 1 ;Q 2èë Q 1 Q 2 where Q æ 1 = arg min Q1 ëj1èq1èë. D 1 D 2 J 1 (1) Slope =λ Slope =λ 3 J 2 (2,2) J 1 (2) 2 1 J 2 (1,2) R 1 R 2 J 1 (2) < J 1 (1) BUT J 1 (1) + J 2 (1,2) < J 1 (2) + J 2 (2,2) Figure 4-3: Example of search for the minimal ç. Choosing quantizer 2 would be optimal if only the ærst frame were considered. However choosing quantizer 1 for frame 1 is better overall because the gain for frame 2 compensates the suboptimality for frame 1. By a simple extension of Theorem 4.1, it follows that the optimal solution to our general Níframe dependency problem consists in introducing J i èq1;q2;:::;q i è= w i D i èq1;q2;:::;q i è+çr i èq1;q2;:::;q i è for i =1;2;:::;N and solving the fol-

10 114 lowing unconstrained problem for the ëcorrect" value of ç which meets the given R budget : min Q 1 ;Q 2 ;:::;Q N ëj1èq1è+j2èq1;q2è+:::j N èq1;q2;:::;q N èë: è4.12è Complexity The optimal solution, as shown by è4.12è is obviously exponentially complex in the dependency-tree depth N. Moreover, it has to be pointed out that the computational complexity is dominated by the data generation phase, i.e. ænding all the èr i èq1;q2;:::;q i è;d i èq1;q2;:::;q i èè points for the problem is much more complex than ænding the optimal solution, given all the possible RíD operating points. The complexity comes from the fact that, contrary to the situation in the independent case, the RíD characteristic for a given block is a function of choices of quantizers for previous blocks and thus an exponentially growing èin the depth of the dependencyè number of possible solutions has to be considered. In order to ease this computational burden, we are therefore interested in methods which will avoid the need to grow all the RíD data, while retaining optimality. We now examine an important property which enables us to do this, and which will be used in Section 4.3 to formulate pruning conditions to eliminate suboptimal choices in the MPEG allocation problem. Note that while our methods rely on generating ëreal" RíD data, a model-based approach may be used within our framework to reduce the complexity. For instance, one could measure some of the R-D points and the extrapolate the R-D values for the others.

11 Monotonicity The key to obtaining a fast solution to the complex dependent allocation problem of è4.3è,è4.4è is the monotonicity property of the RíD curves of the dependent components èframesè. Consider the example of 2 frames, with the operational RíD curve of the second frame depending on that of the ærst, as in Fig Assume that the quantizer grades are ordered as monotonically increasing from ænest to coarsest. Let us use i and j to denote the quantization choices for the independent and dependent frames, respectively. Thus, using our convention, iéi 0 denotes that quantizer i is æner than quantizer i 0. Deænition 4.1 A dependent coding system has the monotonicity property if, for any ç ç 0: J2èi; jè ç J2èi 0 ;jè; for i ç i 0 : è4.13è For example, for ç = 0, this means that D2èi; jè ç D2èi 0 ;jè; for i ç i 0 : è4.14è Stated in words, the monotonicity condition simply implies that a ëbetter" èi.e. æner quantizedè predictor will lead to more eæcient coding, in the rate-distortion sense, of the residue èwhose energy decreases as the predictor quality gets betterè. That is, the dependent frame's family of R-D curves will be monotonic in the æneness of the quantizer choice associated with the parent frame from which they are derived. As can be seen, the æner the quantization for frame 1 èfig. 4-1èaèè, the closer to the origin of the R2, D2 graph will be the corresponding curve for the dependent frame 2 èfig. 4-1èbèè. Experimental results involving MPEG verify this monotonicity property for all the cases that we studied. Thus, monotonicity appears to be a realistic

12 116 property, which has favorable theoretical implications as well, and it can be used to formulate fast pruning conditions for the MPEG allocation problem in Section Temporal dependency: the MPEG case We now address the general temporal dependency quantization problem of which MPEG ë58ë is an example. The MPEG coding format, a CCITT video compression standard, shown in Fig. 4-4, segments the video sequence into groups-of-pictures ègop'sè. Each GOP consists of three types of frames using the intraframe èiè, prediction èp è and bidirectionallyíinterpolated èbè modes of operation. The I frames are coded independently, the P frames are predicted from the previous I or P, and the B frames are interpolated from the previous and next I andèor P frames. For the 2-frame dependency illustrated in Fig. 4-1, the problem we are trying to solve èfor an MSE criterionè is the problem of è4.3è,è4.4è with w1 = w2 =1: min Q 1 ;Q 2 ëd1èq1è+d2èq1;q2èë s:t: R1èQ1è+R2èQ1;Q2èçR budget : The solution to this problem was shown in Section as being exponentially complex in the dependency tree depth. Here, we will show how to reduce this complexity for the MPEG coding case èsee Fig. 4-4è. Before we tackle the general MPEG problem èwith I,P and B framesè, we begin with a simpler special case of MPEG that is easier to analyze and which provides the intuition for the more complex general problem A particular case of MPEG: I-B-I We consider a special case of MPEG having only I and B frames èsee Fig. 4-5è, i.e. the predicted P frames of the more general MPEG format are omitted. The

13 Figure 4-4: Typical MPEG coding framework. èaè The MPEG frames: the I frames are independently coded, the P frames are predicted from previous I or P frames, and the B frames are interpolated from adjacent I andèor P frame pairs. èbè Temporal dependency in the MPEG framework. Note that the B frames are leaves in the dependency tree, since no prediction is generated using them. dependency tree is shown in the more compact form of a trellis. The ëstates" of the trellis represent the quantization choices for the independently coded I frames èordered from top to bottom in the direction of ænest to coarsestè, while the ëbranches" denote the quantizer choices associated with the two B frames. The trellis is populated with Lagrangian costs èfor a æxed çè associated with the quantizers for each frame. Let us focus on the I1, B1, B2, I2 stage of the trellis. The state nodes are populated with the costs of the respective I frame quantizers J èqè = èdèqè + çrèqèè. Each èi; jè branch connecting quantizer state i of I1 to quantizer state j of I2 is populated with the sum of the minimum Lagrangian costs

14 Figure 4-5: The I-B-I special case of MPEG. Finding an R-D convex hull point corresponding to a ç is equivalent to ænding the smallest cost path through the trellis. Each trellis node corresponds to a quantizer choice for the I frames, monotonically ordered from ænest to coarsest, and is populated with the associated Lagrangian cost èj èiè = Dèqè + çrèqèè. The branches correspond to the B frame pairs, and are populated with their minimum Lagrangian costs èj èbè = minëdèqè+çrèqèëè for the particular I frame quantizer choices given by each branch's end nodes. For quality slope ç, the optimal total cost path is obtained with the Viterbi algorithm. The ëdark line" path joins the smallest cost I frame nodes. Monotonicity implies that all dashed line paths can be pruned out. of the B1 and B2 frames, i.e. with JèB1è+JèB2è, where: JèB l è = min Q Bl ëdèq Bl è+çrèq Bl èë for l =1;2 è4.15è where the R-D curves for B1; B2, are generated from the i; j quantizer choices for I1;I2 respectively. From è4.12è, it is clear that the optimal path is that which has the minimum total cost across all trellis paths. Since the independent I frames ëdecouple" the B frame pairs from one another, it is obvious that the popular Viterbi

15 119 algorithm èvaè ë34ë will provide the minimum cost path through the trellis, i.e. we need to keep a single path èthe minimum cost oneè arriving at each node. More formally,wehave the following algorithm for the I1, B1, B2, I2 stage of the trellis: Algorithm 4.1 Step 1 : Generate J èi1è and J èi2è for I-frames I1 and I2 for all quantizers in Q I1 ;Q I2 respectively, i.e. populate all the nodes of the trellis with the Lagrangian costs. Step 2 :For every pair of nodes èi; jè, where i and j are the quantizer choices for I1 and I2 respectively, assign branch cost JèB1è+JèB2è where JèB l è for l =1;2 are obtained from è4.15è, i.e. populate all the branches of the trellis with the minimum Lagrangian costs. Step 3 èva pruning ruleè: At every node of I2, keep only that branch which minimizes J èi1è+j èb1è+j èb2è. As noted in Section 4.2.5, the computational complexity is dominated by the data generation phase, i.e. in the trellis population phase Pruning conditions implied by monotonicity The monotonicity condition stated earlier in Section will now be used to formulate pruning conditions to eliminate suboptimal operating points in the temporal dependency coding problem. The ærst lemma is associated with Fig. 4-6èaè. As a reminder, the quantizer states are ranked in a monotonically increasing order from ænest to coarsest. Lemma 4.1 If J1èiè+J2èi; jè éj1èi 0 è+j2èi 0 ;jè for any iéi 0 ; è4.16è

16 Figure 4-6: Pruning conditions obtained from monotonicity. èaè J1èi2è + J2èi2;jè is the minimum Lagrangian cost of all branches terminating in node j. Therefore èsee Lemma 1è, the èi3;jè branch can be pruned. èbè J2èi; j1èè is the minimum Lagrangian cost of all branches originating from node i. Therefore èsee Lemma 2è, the èi; j2è and the èi; j3è branches can be pruned. ècè Diagram used for the proof of Lemma 1.

17 121 then the èi 0 ;jè branch cannot be part of the optimal path and can be pruned out. Proof: We prove the lemma by contradiction. Assume that èi 0 ;jè for any iéi 0 is part of the optimal path èsee Fig. 4-6ècèè and let the optimal quantizer sequence path be èi 0 ;j;k;:::;lè. But, by monotonicity, wehave: J3èi; j; kè ç J3èi 0 ;j;kè è4.17è æææ J L èi;j;k;:::;lè ç J L èi 0 ;j;k;:::;lè è4.18è Summing up Eqs. è4.16è, è4.17è, :::, è4.18è, we get the contradiction that the total Lagrangian cost of the path èi;j;k;:::;lè is smaller than that of the optimal path èi 0 ;j;k;:::;lè. 2 The above lemma is associated with pruning branches that merge into a common destination state. A dual result holds for the pruning of branches that originate from a common source state èsee Fig. 4-6èbèè leading to the following companion Lemma 2, whose proof is omitted as it is similar to that of Lemma 1: Lemma 4.2 If J2èi; jè éj2èi; j 0 è for any j é j 0 ;then the èi; j 0 è branch cannot be part of the optimal path and can be pruned out. Note that a consequence of the above lemma is that if J1èiè éj1èi 0 è for iéi 0, then the state node i 0 èand all branches from itè can be pruned out. The two pruning conditions of Lemmas 1 and 2 can be used to lower the complexity of the VA-based search. In the special case of MPEG of Section èrefer to Fig. 4-5è, Lemmas 1 and 2 eliminate the need to consider the full trellis on which to run the VA, making

18 Figure 4-7: General MPEG ëtrellis" diagram extension of Fig. 3. Here, the inclusion of the P frames prevents the decoupling of the B frame pairs, and the entire tree has to be grown. Note that each stage of the trellis is represented by ëvector" branches whose dimension grows exponentially with the dependency tree depth General MPEG bit allocation Having established the intuition behind dependent allocation and the power of monotonicity,wenowevolve to the more complex ègeneralè MPEG format of Fig The presence of the P frames extends the dependency tree depth, and the decoupling between successive stages of the trellis is lost. We can thus no longer resort to the

19 123 Viterbi algorithm, but must instead retain the entire tree, which grows exponentially with the number of dependent levels èalthough only until a new I frame is transmitted, so that the complexity may still be manageable.è The good news, however, is that the monotonicity conditions still apply, and the pruning conditions of Lemmas 1 and 2 can aid in reducing complexity dramatically. As an example, see Fig. 4-8 where we consider an I-B-P -B-P sequence of MPEG frames ènote that for simplicity, we use only one B-frame between I-P pairsè and a choice of 3 quantizer grades for each frame. More formally, the algorithm used is the following èrefer to Figs. 4-7 and 4-8è: Algorithm 4.2 Step 1 : Generate JèIè for all quantizers q 2 Q I.See Fig. 4-8èaè. Step 2 : èmonotonicityè Prune out all I-nodes lying below minimum cost node q æ 2 Q I in Step 1. Step 3 : Grow JèI;P1è for all combinations of q 2 Q P1 and all remaining q 2 Q I after Step 2. See Fig. 4-8èbè. Step 4 : èmonotonicityè Use pruning conditions of Lemmas 1 and 2 to eliminate suboptimal I, P1 combinations. See Fig. 4-8èbè. Step 5 : For every surviving I, P1 combination, ænd the B1;B2 quantizer pair that minimizes J èb1è +J èb2è, i.e. populate the branch costs of the trellis of Fig See Fig. 4-8ècè. Step 6 : èmonotonicityè Use pruning conditions of Lemmas 1 and 2 to eliminate suboptimal I, B1, B2, P1 combinations. See Fig. 4-8ècè. Step 7 : For all remaining paths, repeat Steps 3 to 6 for the èp1, B3, B4, P2è and the èp2, B5, B6, Iè sets.

20 124 The smallest cost path after running Algorithm 4.2 is the optimal solution corresponding to the chosen ç for the group-of-pictures considered. While the exhaustive search would have us grow as many as 363 Lagrangian costs, in our example, only 36 costs need to be grown, an order of magnitude reduction in complexity with no loss of optimality if the monotonicity conditions apply èin our example, application of Algorithm 4.2 for ç = 10 provides an optimal RíD operating point í db at 1bppíaswas veriæed through exhaustive searchè. The complexity reduction due to monotonicity is dependent on the desired quality slope ç, with higher quality targets achieving better reduction. In the limit, as ç goes to 0, the minimum cost path is always the one corresponding to the ænest quantizers and thus only a single ëhighest quality" path has to be grown. Conversely, ifçgoes to 1 the monotonicity property provides no gain. Thus one should not resort to the convex search techniques described in Section and would have to use instead search strategies which start with smaller values for ç and increase it until a solution is reached Suboptimal heuristics As pointed out, the amount to which the monotonicity property can be exploited is ç-dependent, and may not suæce for some applications. To this end, it is advisable to come up with fast heuristics, which, used in combination with monotonicity, can approach the optimal performance at a fraction of the complexity. In trying to formulate a fast MPEG heuristic, it is necessary to consider some important points: èiè the ëanchor" I-frame is the most important of the group of pictures and must not be compromised, èiiè most signal sequences enjoy a ænite memory property, where the inæuence of a parent frame diminishes with the level of its dependency. Thus, it may paytochoose èonlyè the lowest cost nodes for all frames except the I-frame, for which we retain all nodes remaining after applying monotonicity-based

21 125 pruning. Thus, a single path is grown from each of these admissible I-frame nodes, whereas in the general case, a whole tree evolves from each such node. Based on this, we propose the following heuristic: èiè retain all paths that originate from each of the I-frame quantization states remaining after the monotonicity-based pruning, i.e. it is not prudent to be greedy for the I-frame, as a greedy error aæects all dependent frames derived from it; èiiè use a ëgreedy" pruning condition èin combination with the monotonicity propertyè to keep only the lowest cost branch thus far at all other stages in the trellis. That is, we follow Algorithm 4.2, except that we add an extra pruning condition in Steps 4 and 6, where we retain only a single èminimum costè path corresponding to every surviving I-frame node. This heuristic, as shown in Fig. 4-9, leads to near optimal performance at a fraction of the computational cost. For the example of Figs. 4-8 and 4-9, the optimal solution for a particular ç gives db at 1 bpp, while the heuristic achieves db at 0.97 bpp, certainly very close to optimality. Unlike Algorithm 4.2, which relies on monotonicity pruning conditions only and which works best when ç is low èhigh qualityè, the fast heuristic retains low complexityeven at high values of ç. Note how choosing the greedy solution for the I-frame èbottom-most nodeè would have given a much worse performance Discussion We have shown a method to ænd the optimal bit allocation strategy for an MPEG coding framework, assuming arbitrary quantizer sets for each MPEG frame. Although our scheme can be computationally complex, it can serve as an optimal benchmark to evaluate more practical allocation strategies. Also, model-based approaches èe.g. measuring one RíD point and using a model-based extrapolation to ænd other pointsè can be combined with our techniques to ease the computational burden. Note that since MPEG coders are typically buæered with a buæer size of the order of a group-

22 Figure 4-8: Tree pruning using the monotonicity property èlemmas 1,2è. The numbers are the cumulative Lagrangian costs for a typical example for ç = 10. Branches pruned at each stage are shown with dashed lines. In this example, the number of RíD points generated is cut down from 363 èexhaustiveè to only 36 with no loss of optimality. 4.4 Buæer constrained allocation for dependent coders In Chapter 2 we dealt with the problem of allocating bits among diæerent signal blocks when the sequence of blocks had to be transmitted with ænite end-to-end delay, i.e. we had an allocation problem with the constraint that the encoder buæer did not overæow. In Chapter 2, however, we had assumed that the blocks to be quantized were coded independently èin an MPEG environment this would mean

23 Figure 4-9: Tree pruning using monotonicity as well as a ëgreedy" heuristic for the same conditions as those of Fig. 6. The number of R-D points generated is now 24, at a slight loss of optimality ètotal Lagrangian cost is versus optimal cost of 77.24è. only I frames are usedè so that our solution there would not be applicable to the dependent environments discussed in this chapter. Here we outline how the solution of Section can be modiæed to account for the buæer constraint. Given the rate and distortion èr and D, resp.è for each frame and calling Bèiè the buæer occupancy after coding frame i, our buæer control problem is that of ænding the quantizer choice èq1;:::;q N è to: NX NX minè D i èq1;:::;q i èè s:t: èiè R i èq1;:::;q i èçn ær = R budget è4.19è i=1 i=1 èiiè Bèiè ç B max ; 8i: è4.20è Obviously, without the buæer constraint of èiiè, our problem reduces to the dependent bit allocation studied in Section There, the minimization of the corresponding Lagrangian cost P J i èq1;:::;q i è for each block was formulated as the

24 128 search for the minimum Lagrangian cost path in a trellis, where the states represent the possible quantizers for each frame èsee Fig. 4-7è. The search can be sped up by resorting to heuristics based on the monotonicity property to prune out suboptimal branches. When constraint èiiè is considered, not all solutions obtained using the algorithms of Section will be feasible due to overæow. However, if we add to the trellis pruning rules a new rule which eliminates paths that overæow, the resulting algorithm will provide the optimal solution to è4.19è and è4.20è, within a convex hull approximation. More formally, we grow all the solutions as in Section where each branch, associated with P a given quantizer selection, èq1;:::;q i è, will have associated a Lagrangian cost J i èq1;:::;q i è and a buæer state, Bèiè =maxèbèi, 1è + R i èq1;:::;q i è,r;0è: Thus, on top of the pruning rules described in Section 4.3.2, we will prune out any branch such that Bèiè é èb max, rè. The one parameter that remains to be set is the initial buæer state. Typically, since regardless of the coding environment the ærst frame in the sequence will be coded in Intra mode èand thus will very likely require a number of bits above the channel rateè, we can assume that Bè0è = 0 so that the buæer is initially empty. If we focus speciæcally in an MPEG environment, we can consider a group of pictures ègopè and compute the allocation so that the initial and ænal buæer states are both zero. In this manner we can decouple the buæer constrained allocation for consecutive GOPs. The P above described method requires, in order to solve the problem, that a ç such that R i èq æ 1 ;Qæ 2 iè=nærbe found. The pruning rules based on monotonicity ;:::;Qæ are more eæcient for small values of ç èi.e. high bit rateè and, similarly, smaller ç's will result in more overæowing paths. Thus it is reasonable to initialize ç to a ësmall"

25 129 value and then, to ænd the optimal solution, increase ç until the target rate is met. Solutions that were discarded because of overæow for a given ç do not have tobe reconsidered for a diæerent ç. This algorithm leaves also room for fast approximations based on heuristics. For example, assume that Q1 = q is æxed and consider the resulting R-D characteristic, i.e. the set of all points è P R i ; P D i è for which Q1 = q. Then, if the point corresponding to èq; Q2;:::;Q N è lies on the convex hull of the the R-D characteristic generated with Q1 = q we have found that typically èq 0 ;Q2;:::;Q N è will also lie on the convex hull of the R-D characteristic obtained when setting Q1 = q 0.Thus after studying the R-D characteristic for a particular value Q1 = q one can extrapolate which combinations for èq2;:::;q N è are likely to be suboptimal for any Q1 and therefore reduce the search to those that are likely to be ëgood". To summarize, in this section we have described some initial results for a deterministic rate control for MPEG encoders. The optimal solution can be found with slight modiæcations of the algorithm that solves the unconstrained problem. Further work should focus on exploring more ways of reducing the complexity of the search by exploiting the structure of the R-D data through the use of heuristics. 4.5 Appendix: Proof of Theorem 1 Note that in the proof, in the interest of notation brevity, we omit the explicit dependence of D1;R1;J1 and D2;R2;J2 on Q1 and Q2; i.e. we use J1 = J1èQ1è;D1 = D1èQ1è;R1 = R1èQ1è and J2 = J2èQ1;Q2è;D2 = D2èQ1;Q2è;R2 = R2èQ1;Q2è, and similarly for the optimal quantizers Q æ 1 and Q æ 2. Then, for all Q1;Q2,wehave from the unconstrained minimization: J æ 1 + J æ 2 ç J 1 + J2: è4.21è

26 130 i.e., D æ 1 + çræ 1 + Dæ 2 + çræ 2 ç D 1 + çr1 + D2 + çr2: è4.22è or, ëd æ 1 + D æ 2ë, ëd1 + D2ë ç çèër1 + R2ë, ër æ 1 + R2æëè: è4.23è Since è4.23è holds for all admissible fq1;q2g, it certainly holds for the subset of fq1;q2g for which ër1+r2ëçr budget, where R budget =ër æ 1+R æ 2ë. Therefore, from è4.23è, since ç ç 0, we have that: ëd1èq æ 1è+D2èQ æ 1 ;Qæ 2èë, ëd1èq1è+d2èq1;q2èë ç 0; è4.24è i.e. over all fq1;q2g which meet the rate budget, èq æ 1 ;Qæ 2è gives the minimum distortion. 2