FAST AND EFFICIENT SIDE INFORMATION GENERATION IN DISTRIBUTED VIDEO CODING BY USING DENSE MOTION REPRESENTATIONS

Transcription

1 18th Euroean Signal Processing Conference (EUSIPCO-2010) Aalborg, Denmark, August 23-27, 2010 FAST AND EFFICIENT SIDE INFORMATION GENERATION IN DISTRIBUTED VIDEO CODING BY USING DENSE MOTION REPRESENTATIONS Giovanni Petrazzuoli, Marco Cagnazzo, Béatrice Pesuet-Poescu Institut Telecom Telecom ParisTech CNRS LTCI 46 rue Barrault, F Paris Cedex 13, FRANCE hone: +33 (0) , fax: +33 (0) {etrazzu, cagnazzo, ABSTRACT Distributed video coding (DVC) does not demand motion estimation (ME) and comensation (MC) at the encoder, but only at the decoder and so it is more suitable for alications that reuire a simle encoder, like wireless sensor networks. In DVC the video seuence is slit into Key Frames (KFs) and Wyner-Ziv Frames (WZFs): the first are intra-coded and the latter are coded by a channel code and only the arity bits are sent to the decoder. The KFs are available at the decoder, while we need to estimate the WZF and correct this estimation with arity bits. One critical ste is the estimation of the WZF. The method of the state of the art, with which we comare, is given by DISCOVER. It estimates the WZF by linear interolating the two adjacent KFs. We roose a higher order motion interolation for WZF estimation by using four KFs. Due to the high comutational efforts, we roose also a fast algorithm that halves the comlexity of the revious method. We note that the results of the fast method are comarable with the original one. An other roosal is to increase the density of the motion vector field in order to imrove the estimation of the WZF. Keywords: Distributed video coding, image interolation 1. INTRODUCTION Let X and Y be two correlated sources. If we encode them jointly, we can decode them without loss of information if R X +R Y H(X,Y), 1 but if we encode searately X and Y, we exect that we can decode them without loss of information if R X H(X) and R Y H(Y), according to the first Shannon theorem. Indeed, according to the Sleian-Wolf Theorem [5], a total rate R X +R Y H(X,Y) is sufficient, even for searated encoding of deendent sources, rovided that we decode them jointly. A articular case of distributed source coding is the source coding with side information: the variable X, generated by the source, is statistically deendent of Y. The variable Y is available at the decoder, but not at the encoder. Y is called side information. Then, if R Y = H(Y), a rate R X H(X Y) is sufficient for recovering X without loss of information, according to the Sleian-Wolf theorem. While Sleian-Wolf theorem is about lossless coding, the roblem of lossy coding was solved by Wyner and Ziv [6]. They roved that there is no rate increase in the source coding with side information with resect to the joint coding for a given uality, and conversely, there is no uality loss for a given rate, subjected to some mild constraints. 1 Let H(X) be the entroy of X and H(X,Y) the joint entroy of X and Y. These results can be alied to video coding in order to simlify the encoder structure. In Distributed Video Coding (DVC) the video seuence is divided in Grou of Pictures (GOP). Each GOP consists of one Key Frame (KF) (usually the first one), that is intra-coded (i.e. it is coded indeendently of the other frames) and Wyner-Ziv Frames (WZFs) that are coded by a systematic channel code. We send to the decoder only the redundance bits (i.e. the arity bits). Even if the KFs and the WZFs are correlated sources, we do not exloit this deendence at the encoder, but only at the decoder. We roduce an estimation of WZF by interolating the adjacent decoded KFs. This ste is called Image Interolation and the WZF estimation is called Side Information (SI). The SI can be considered a noisy version of the WZF. Then, we correct the errors of the WZF estimation with the arity bits sent by the channel encoder. The weak oint of this scheme is a feedback channel is needed to adjust the number of arity bits made by the encoder. On the other hand, the advantage of this structure is that we move the comlexity in terms of comutation (the motion estimation) and in terms of memory (the exigence to storage the revious frames) from the encoder to the decoder. This is desirable if we need a low-comlexity encoder, as in wireless sensor network, but it is not well suited for broadcast transmission. The reference techniue for WZF estimation is given by DISCOVER, that erforms a linear interolation between the two more adjacent KFs. In [4] we roose a higher order motion interolation (HOMI) method by using four KFs, instead of two such as in DISCOVER. In this aer we imrove those results by increasing the density of the motion vector field (MVF) for the motion estimation, and we roose also a new method (Fast HOMI) in order to reduce the comlexity of our algorithm. The rest of the aer is organized as follows. In Section 2 we describe the DISCOVER motion interolation method. In Section 3 we illustrate the method roosed in [4] and after we roose a variant to our method in order to reduce the comlexity. Exerimental results are reorted in Section 4, while conclusions and future work are in Section STATE OF THE ART: DISCOVER MOTION INTERPOLATION ALGORITHM One of the most oular methods for image interolation is the method roosed in the DISCOVER roject [3]: it consists in a linear interolation between two adjacent KFs. For examle, if the GOP size is eual to 2, we use the KFs I k 1 and I k+1 for the estimation of the WZF I k. The DISCOVER method consists of four stes: EURASIP, 2010 ISSN

2 k 1 k k+ 1 KF u WZF w KF Figure 1: Bidirectional motion estimation in DISCOVER. The distance between and is small, such that the motion incan be aroximated with the motion in. 1. Low Pass Filter. The two KFs, I k 1 and I k+1, are satially filtered in order to reduce noise. 2. Forward Motion Estimation. A block matching is erformed from I k+1 to I k 1. Letvbe the motion vector field (MVF) calculated at this ste (green arrows in Fig. 1). 3. Bidirectional motion estimation. For each macroblock (letbe its center) we search the vectorvthat intercets the frame I k in the oint closest to. Let be this intersection. Then, the movement in can be aroximated as the movement in (black vector in Fig.1). Afterwards, the vector is slit intow(backward MVF) andu(forward MVF) (light-blue vectors in Fig. 1) 4. Refinement and median filter. The vectorswanduare refined around their initial ositions in order to minimize the SAD (or SSD) between the MB ointed by w in I k+1 and the one ointed by u in I k 1. Afterwards, a median filter is alied to the two MVFs in order to smooth them. Finally, I k+1 is motion-comensated with w and I k 1 with u. The average of these two redictions is the estimation of the WZF. 3. PROPOSED METHODS In [4] we roosed a higher order motion interolation method in order to increase the results given by DISCOVER. For the sake of clarity, we reort here the basic ideas of this algorithm and after we roose some variants to this algorithm. 3.1 Higher order motion interolation (HOMI) While DISCOVER motion interolation method erforms a linear interolation between the two adjacent KFs I k 1 and I k+1 (by suosing the GOP size eual to 2), we roosed in [4] a higher motion interolation by using four KFs: I k 3, I k 1, I k+1 and I k+3. Our method consists into four stes: 1. Initialization. By using only I k+1 and I k 1 we comute the backward and the forward MVF, resectively u and w, as in DISCOVER algorithm (black dashed vectors in Fig. 2(a)). 2. Motion estimation from I k±1 to I k±3. We erform a block matching motion estimation from I k±1 to I k±3. Let be the center of the MB in the frame I k that we want to estimate. Let B k be the MB in the frame I k, centered in. Then, we search for the vector ũ such that the following functional is minimized: J(ũ)= B +u() k 1 () B +ũ +λ ũ 3u, k 3 () (1) where λ 0 is a regularization constant. The regularization term is added for enalizing too large deviations ofũ from 3u, i.e. the solution that we would have suosing a linear motion along the four frames. 3. Interolation. Now, we can trace the trajectory of the object along the various KFs by interolating the ositions +ũ(), +u(), +w(), and + w(), resectively at the instants k 3, k 1, k+1 and k+3. Then, by samling it in k, we obtain the estimated osition of the object in the frame I k. Let ˆ be this osition. We can then estimate the motion vectors centered in ˆ (red vectors in Fig. 2(a)). 4. Vector adjustment. We suose that the distance between ˆ and is so small, such that the motion in is the same as in ˆ, that we can aroximate the vectors in with the vectors estimated in ˆ (green dashed vectors in Fig. 2(a)). Afterward, we reeat this for each macroblock. The average of the two comensations will be the WZF estimation. We reeat this rocedure for each macroblock. Afterward, we motion-comensate the frame I k+1 by the backward MVF and I k 1 by the forward MVF. The average of the two comensations will be the WZF estimation. 3.2 Fast HOMI The comlexity of the interolation rocedure described in the revious section can be reduced, because at the instant k, we have already estimated the forward motion vector field v, from the frame I k 2 to the frame I k 3 and the backward motion vector fieldzfrom I k 2 to I k 1. In this way, it is only necessary the motion vector field from I k+1 to I k+3 while it is not necessary to erform the motion estimation from I k 1 to I k 3 : we can exloit the MVF comuted at instant k 2. This method is less robust than the first one, because instead of the motion estimation from I k 1 to I k 3, we just exloit the estimation of two MVFs v and z. This gives us a less accurate motion estimation. On the other hand, thanks tovandz, we can find another macroblock in the frame I k 2 : then we can interolate the data by using five oints (not regularly saced). The rocedure consists therefore of the following five stes (see Fig. 2(b)): 1. Initialization. - We estimate u from I k to I k 1 and w from I k to I k+1 by DISCOVER method. 2. Motion estimation from I k+1 to I k+3. We erform a block matching motion estimation from I k+1 to I k+3 and we find the osition + w. 3. Motion estimation from I k 1 to I k 2 and from I k 2 to I k 3. We search for the vector z() that oints in I k 1 to the osition closest to +u(). Then, we estimate the intersection of the trajectory in the frame I k 2 as the 2157

3 k 3 k 1 k k+ 1 k+ 3 +ũ() B +u() k 1 +u() û ŵ B +w() k+1 ˆ +w() + w() Discover vector Interolated vector Vector comuted Interolated trajectory with our method (a) k 3 k 2 k 1 k k+ 1 k+ 3 +ũ v +t z +z() +u() ˆ +w() + w() Discover vector Interolated vector Vector comuted with the roosed method Previous vectors Adjusted revious vectors Interolated trajectory (b) Figure 2: (a) HOMI method for motion estimation. (b) Proosed interolation method (Fast HOMI) for WZF estimation by exloiting the revious estimated MVF. oint +t, with t = u() z(). For the estimation of the intersection of the trajectory in I k 3, we use the vector v(). The intersection oint will be +ũ, with ũ()=t+v(). 4. Interolation. Finally, we interolate a vector function with the five values+ũ,+t,+u(),+w() and + w() resectively, at the instants k 3, k 2, k 1, k+1 and k+3, in order to find its value at the instant k, 2158

4 GOP size λ Table 1: Values of λ for different GOP sizes which will be denoted by ˆ. 5. Vector adjustment. It is done as in the revious section. We observe that the comlexity of Fast HOMI is about the half of the original algorithm. 3.3 Increasing the density of the MVF When we erform the block matching, we consider a block B k, centered in and whose size is N N. Then, we let vary in {(nm,mm)} (n,m) C, where C Z 2 is such that we comute a vector for each M M block in the image. Although the condition M= N is usually chosen, in order to increase the SI uality we can use denser MVF by selecting a value for M smaller than N. We observe that, while we can increase the density by reducing M without articular constraint (exceted for the comutational comlexity), we cannot allow arbitrary variations of N (the block size) since if the blocks are too small, the matching may suffer from noise. On the contrary, thanks to the block overla, we can densify the MVF without sacrificing robustness. So, we will consider N = 8 in the following, but we will use denser MVF with M < N. We can therefore modify both the HOMI and the Fast HOMI techniue, ending u with 4 SI generation techniue: HOMI with N = M = 8 (that is, the one roosed in [4]), referred to as HOMI8 from now on; HOMI with N = 8 and M = 4, which we will indicate as HOMI4; and the fast version of these techniues, referred to as FastHOMI8 and FastHOMI4. The three new techniues (along with HOMI8 for comleteness), will be comared with the reference method, DISCOVER, which does not use the overla. 4. EXPERIMENTAL RESULTS In order to use the roosed SI generation techniues, we need to tune the regularization arameter λ for different GOP sizes. At this end we can use the results reorted in [4] for HOMI8, since we consider that λ deends mainly on the block size N. For the sake of comleteness, we reort the otimal values of λ as a function of the GOP size in Tab. 1. We found that the otimal value of λ decreases when the KFs are farther aart. This is reasonable since in this case we must allow larger vector deviations to take into account the movement. Now we can comare the different methods. We used the test seuences book arrival, ballet, jungle and breakdancer at a resolution of ixels. We encoded the KFs by the INTRA mode of H.264, using four uantization ste values, namely 31, 34, 37 and 40. At this stage, we use as evaluation metric the PSNR of the SI with resect to the original WZF. More recisely for each of the four methods, we comute the PSNR difference (averaged along each seuence) with resect to DISCOVER. This uantity is called PSNR. The results of this test are reorted in Tab. 2 to 5. We observe that in almost all cases, the uality of the side information is imroved w.r.t. DISCOVER. The only excetion is for GOP size eual to 8, when a good SI estimation is QP Table 2: PSNR [db] for HOMI 8 difficult, and all methods are almost euivalent (differences usually under 0.1 db). Then, we observe that denser MVFs imrove the SI uality for GOP size eual to 2, while they do not hel in the case of long-term estimation. We ascribe this behavior to the difficulty of estimating images that are uite far from the references. Finally we observe that the fast versions of HOMI have fairly good erformances, since the uality of the SI is almost unchanged in many cases, while the comutational comlexity is halved. The last exeriment consisted in comuting end-to-end erformances (i.e. rate reduction and PSNR imrovement) of the roosed techniues when inserted into a comlete DVC coder like the one in [1], by using the the Bjontegaard metric [2]. The results are shown from Tab. 7 to Tab. 9, and they are not surrising: the roosed method are in general better than the reference DISCOVER, exceted for GOP size eual to 8, where they are ractically euivalent. Moreover, even from the oint of view of RD erformances, denser MVF are better than sarser ones, and the fast version of HOMI are as effective as the original algorithms. We remark that globally, the better techniues is HOMI4, which allows rate reductions u to 8% w.r.t. the reference. As a final observation, we note that increasing the uality of the side information does not mean always an increasing of the RD erformances. For examle, the HOMI8 method has a better side information than DISCOVER for GOP size eual to 8, but worser RD erformances. This confirms the intuition that the PSNR with resect to the original WZF is not necessarily an accurate method for evaluating the SI uality, even though for the moment is the most common, since in most cases the RD erformances are well correlated to the side information PSNR. 5. CONCLUSIONS AND FUTURE WORK Based on the revious work in [4], where we roose a higher order motion interolation, we continue to exlore higher order motion interolation techniues in order to increase the SI uality. We try to increase the density of the estimated MVF and at the same time to reduce the comlexity by reusing the already estimated MVF. The techniue using dense MVFs is successful both in increasing the SI uality (u to 0.47 db better than the reference method DISCOVER), and in imroving the end-to- 2159

5 QP Table 3: PSNR [db] for HOMI 4 QP Table 4: PSNR [db] for Fast HOMI 8 end RD erformances, with a rate reduction attaining 8.2% in the best case. Moreover we show that fast version of the HOMI algorithms have almost the same erformance as the original one, but with an halved comlexity. These good results encourage us to kee looking for efficient SI methods. A techniue we intend to investigate will exloit the revious WZF which has been fully reconstructed (i.e. by using the arity bits to correct it) to roduce more accurate motion vector fields. REFERENCES [1] A. Aaron, R. Zhang, and B. Girod. Wyner-Ziv coding of motion video. In Asilomar Conference on Signals and Systems, Pacific Grove, California, Nov [2] G. Bjontegaard. Calculation of average PSNR differences between RDcurves. In VCEG Meeting, Austin, USA, Ar [3] C. Guillemot, F. Pereira, L. Torres, T. Ebrahimi, R. Leonardi, and J. Ostermann. Distributed monoview and multiview video coding: Basics, roblems and recent advances. IEEE Signal Processing Mag., ages 67 76, Set [4] G. Petrazzuoli, M. Cagnazzo, and B. Pesuet-Poescu. High order motion interolation for side information imrovement in DVC. In International Conference on Acoustics, Seech and Signal Processing, Dallas, TX, [5] D. Sleian and J. K. Wolf. Noiseless coding of correlated information sources. IEEE Trans. Inform. Theory, 19: , July [6] A. Wyner and J. Ziv. The rate-distortion function for source coding with side information at the receiver. IEEE Trans. Inform. Theory, 22:1 11, Jan QP Table 5: PSNR [db] for Fast HOMI 4 R (%) PSNR [db] R (%) PSNR [db] R (%) PSNR [db] Table 6: Rate-distortion erformance for HOMI 8 R (%) PSNR [db] R (%) PSNR [db] R (%) PSNR [db] Table 7: Rate-distortion erformance for HOMI 4 R (%) PSNR [db] R (%) PSNR [db] R (%) PSNR [db] Table 8: Rate-distortion erformance for Fast HOMI 8 R (%) PSNR [db] R (%) PSNR [db] R (%) PSNR [db] Table 9: Rate-distortion erformance for Fast HOMI