and CAPELLINI, v.: ‘DCT-based watermark recovering without resorting to the uncorrupted original image’. Proc. IEEE Int. Conf. Image Processing, Vol I, October 1997, pp. 520-523 HSU, c.-T., and wu, I.-L.: ‘Hidden signatures in images’. Proc. IEEE Int. Conf. Image Processing, Vol 111, September 1996, pp. 223-226 PIVA, A., BARNI, M., BARTOLINI, F , The threshold can be increased, according to practical requirements. Method for detecting all-zero DCT coefficients ahead of discrete cosine transformation and quantisation Simulation results: In our simulations, we have applied the proposed method to the H.263 software written by Telenor R&D. The frame frequency is ftxed to 10 frames per second. We tested the ‘Miss America’, ‘Claire’ and ‘Susie’ sequences, and found the proposed method to be very useful to these kinds of test sequences, i.e. head-and-shoulder type video sequences mainly used in video phone applications. When the quantisation level Q = 10, -40% of the blocks of the ‘Miss America’ and ‘Claire’ sequences and more than 20% of the blocks of the ‘Susie’ sequence can be determined to have all-zero DCT coefficients without any incorrectly determined blocks. Zhou Xuan, Yu Zhenghua and Yu Songyu Table 1: Simulation results for SQ The authors present a new algorithm for detecting all-zero DCT coefficient blocks before discrete cosine transformation and quantisation in very low bit rate video coding. The proposed algorithm uses the sum of absolute difference (SAD) of each motion compensation block as the criteria. Because the SAD can be obtained after motion estimation, no additional computation is required. Results show that the proposed algorithm can be used to reduce the amount of computation significantly without sacrificing video coding quality. Introduction: Motion estimation, motion compensation, discrete cosine transformation (DCT) and quantisation are major processing components of motion picture coding. In this entire process, DCT is applied to compress motion compensation data in the spatial domain, and a special case for the encoder occurs when all the coefficients from the DCT are quantised to zero. In this situation, instead of sending multiple zeros to the decoder, the encoder sends a special signal indicating the state. This makes its representation very efficient. However, the detection of this state is quite computationally expensive using traditional methods, because it requires the computation of one entire round of the DCT and quantisation, followed by a check to see if all the coefficients are zero. For very low bit-rate coding, especially in videophone applications, this all-zero state is quite common, so early detection of this state could significantly reduce the amount of computation necessary. In this Letter, a method for determining all-zero blocks before DCT and quantisation is proposed. When this method is used in H.263 [I], and the quantisation level is 10, -40% of the blocks of the Miss America and Claire sequences can be determined to have all-zero coefficients. r t 8Q 12Q 16Q 20Q determination threshold while Q=10 24Q 15211 Fig. 1 Graph of all-zero block ratio against determination threshold Proposed algorithm: The discrete cosine transform of a discrete functionAx,y), x, y = 0, 1, ..., N-1 is defined as [2] A Miss America Claire 1 Susie / where ku,k - 1 - u-fi foru,v=O =1 for u , v = 1 , 2,...,N In H.263, N = 8, so eqn. 1 gives c l 7 IF(% .)I < 4 a=O abs(f(z, 9 ) ) - I (2) y=o The condition for all-zero DCT coefficients is IF(u,v)l < 2Q (3) where U, v = 0, 1, ..., 7, and Q denotes the quantisation level. Thus 7 0Q 7 z=o y=o 16Q 20Q 24Q 45zizj The above inequality gives the conditions under which the DCT has all-zero coefficients. In eqn. 4, the left summation gives the sum of absolute difference (SAD) of the motion compensation block, which can be obtained during motion estimation. Therefore no additional computation is required. Eqn. 4 is the full condition. ELECTRONICS LETTERS 12Q determination threshold while Q=10 (4) 77th September 7998 Pig. 2 Graph of incorrect determined block ratio against determination threshold A Miss America Claire W Susie Vol. 34 No. 19 1839 Table 1 shows the simulation results for threshold 8Q. With increasing quantisation level Q, the number of all-zero blocks determined will increase very rapidly. From eqn. 4, we know that the threshold 8Q represents the sufficient condition for all-zero DCT coefficients but not the necessary condition. In our research work, we also tried to increase the determination threshold to reduce the amount of computation necessary. Fig. 1 shows the ratio of determined all-zero blocks against determination threshold, and Fig. 2 shows the ratio of incorrectly determined blocks against determination thieshold. In Figs. 1 and 2, the quantisation level Q = 10; from these Figures, we found that 16Q gives a good tradeoff between image distortion and computational burden. of Fig. 1, all cameras, including the ‘virtual’ ones, have optical centres on the X axis, and optical axes parallel to the 2 axis. In general, under these assumptions, points corresponding to different IVs can be accurately reconstructed from the two original images of the stereo pair, provided that the disparities have already been computed. Consequently, the correspondence problem should be solved prior to the IV reconstruction. We stress the importance of having reliable disparity maps to reconstruct IVs, which is supported by prior work [l]. The accuracy of the disparity map strongly affects the quality of the IV reconstruction. Hence, it is assumed here that the reason that points are marked as not matched is on account of the non-existence of homologous points in the other image due to occlusion. Conclusions: A new technique for the early determination of allzero DCT coefficient blocks has been presented. Using the technique, -40% of blocks in the ‘Miss America’ and ‘Claire’ sequences can be determined to be all-zero DCT coefficient blocks. Because the technique uses the SAD as a criterion, no additional computation is required. Moreover, the determination threshold can be changed with the quantisation level self-adaptively. This new technique is very useful for very low bit rate video coding, and can be applied to videophone and video conferencing systems. 0 IEE 1998 Electronics Letters Online No: 19981308 Zhou Xuan, Yu Zhenghua and Yu Songyu (Institute Communication and Information Processing, Shanghai University, Shanghai, 200030, People’s Republic of China) Image Jiaotong of References 1 Draft ITU Rec. H.263: ‘Line transmission of non-telephone signals: Video coding for low bit rate communication’. 5 December 1995 2 MAKHOUL, J.: ‘A fast cosine transform in one and two dimensions’, ZEEE Trans., 1980, ASSP-28, (1), pp. 27-34 Stereo-based intermediate view synthesis with realistic ’look around’ capability M.M. Perez, C.L. Pagliari and T.J. Dennis An algorithm is presented for synthesising intermediate views from a stereo image pair that attempts to produce a realistic 3D ‘look around’ effect. The task is accomplished by suitably incorporatinga term proportional to the corresponding disparities to the co-ordinatesof the ‘virtual’ intermediate view points to be reconstructed. Introduction: In this Letter, we propose an image synthesis algorithm which generates views between those of the two camera views that originally captured the scene. A special treatment for occluded areas is also embedded in the algorithm. In previous work [l, 21 the intermediate views (IVs) are synthesised by mapping the intensities of the points over the 2D image co-ordinates of one of the two camera views that constitute the stereo pair. The key-point of our approach is that the 2D co-ordinates of the new intermediate view points generated are established as functions of both the disparity data for the original stereo pair and the desired position for the ‘virtual’ intermediate camera. A basic motivation for the use of stereoscopic data is to give the feeling of telepresence, which we try to enhance by allowing a realistic ‘look around’ capability. A further benefit occurs in scenes shot by a stereo camera with a large baseline which may cause discomfort to viewers using stereoscopic displays [2]. This can be avoided by generating more tolerable intermediate views. Problem dejnition: As depicted in Fig. 1, the task of IV reconstruction is to generate images that would have been acquired from ‘virtual’cameras, depicted in grey, located anywhere between the original left and right extremes. According to the convention 1840 \U 7 July 1998 Fig. 1 Stereo cameras showing left view (L), intermediate views (pictured in grey) and right view (R) Synthesis of intermediate views: A setup where the cameras have parallel optical and vertical axes and coplanar image planes is assumed, thus implying a linear relationship between the disparity and the displacement, aB,of a virtual camera optical centre along the real camera baseline, length B. Although the algorithm has been developed assuming a parallel camera setup, IV synthesis using the proposed algorithm is also feasible for scenes shot with a camera convergence angle small enough for the induced vertical disparity to be neglected. The reconstructed image Z,, is a linear combination of the left (IL)and right (IR)images and the disparity (d) map. The ‘virtual’ intermediate camera (corresponding to the IV)is related to the left image plane by a translation aB,0 I a < 1. It can be shown, through manipulations over the expressions for the projective transform, that under such a translation the resulting disparity map between the original left stereo view and the new IV is a scaled version (by a factor a)of the original disparity map for the left image of the stereo pair. As a consequence, the intensity of a point in the new IV should not simply be mapped over the coordinates of one of its two homologous points, in one of the original stereo views. Instead, the new co-ordinates of each IV point can be precisely computed, based on the co-ordinates of its homologous point in one of the original stereo views, and the scaled disparity map corresponding to this new IV. Eqn. 1 represents the proposed interpolator applied to synthesise the IV for all matched points of an original M x N pixel size stereo pair. The expression for the resulting co-ordinates of the IV points is the left side of eqn. 1, whereas the right side corresponds to the synthesis of the intensity of the new ‘virtual‘ point from a weighted average of the intensities of its original homologous points. Eqn. 1 takes the left image of the stereo pair as the reference view. This is consistent with the fact that the IV corresponds to the left image for a = 0. In this case, d (which is computed using [3]) is the disparity of the right image in relation to the left image. The remaining unmatched points receive a special treatment. We assume that unmatched regions have near-constant depth which is similar to that of adjacent background areas. This is not an unreasonable assumption, considering that the depth variation along these images patches is unpredictable, and also that they cannot have the same depth as the foreground region, because they could not then be occluded in one of the stereo views. For unmatched points in the left image, each IV intensity point is ELECTRONICS LETTERS 17th September 1998 Vol. 34 No. 19

1/--страниц