Methods and systems for encoding/decoding signals, and computer program product therefor

Field of the invention

0001The present invention relates to encoding/decoding techniques, for instance for media signals.

0002The invention was developed by paying specific attention to the possible application to coding video signals by Multiple Description Coding (MDC) using polyphase downsampling (PDMD).

0003However, reference to this preferred application is in no way to be construed as limiting the scope of the invention.

Description of the related art

0004The goal of Multiple Description Coding (as described e.g. in <nplcit id="ncit0001" npl-type="s"><text>V.K. Goyal "Multiple Description Coding: Compression Meets the Network" IEEE Signal Proc. Mag. Sept. 2001 pp. 74-93</text></nplcit>, is to create several independent bitstreams using an existing video codec (i.e. coder-decoder). Bitstreams can be decoded independently or jointly. The larger the number of the bitstreams decoded, the larger the quality of the output video signal.

0005Multiple Description Coding (MDC) requires a pre-processing stage upstream of the encoder, in order to split the video sequence and control redundancy among subsequences. It also requires a post-processing stage downstream of the decoder, in order to merge the received and successfully decoded substreams.

0006Multiple Description Coding greatly improves error resiliency, because each bitstream can be decoded independently. Also, variable bandwidth/throughput requirements can be managed by transmitting a suitable number of descriptions. However, coding efficiency is somewhat reduced depending on the amount of redundancy left among subsequences.

0007Multiple Description Coding is essentially analogous to Scalable Coding (also known as Layered Coding). The difference lies in the dependency among bitstreams. The simplest case is when two bitstreams are created. In the case of scalable coding they are referred to as "base layer" and "enhancement layer", respectively. The latter layer depends on the former layer and cannot be decoded independently therefrom. On the other hand, in the case of Multiple Description Coding, each description can be individually decoded to get a base quality video. As for Scalable Coding, there can be spatial, temporal or SNR (Signal-to-Noise Ratio) Multiple Descriptions (MD).

0008Replicated headers/syntax and replicated motion vectors among bitstreams greatly impede coding efficiency in SNR MD. Replicated headers/syntax also hinder temporal MD, and motion compensation is less effective because of the increased temporal distance between frames. Spatial MD is hindered by headers/syntax as well. However, contrary to temporal MD, motion compensation is not affected, particularly when 8x8 blocks are split into smaller blocks, as in the latest H.264 codec. Because of this, spatial MD Coding is usually regarded as the best choice for video coding.

0009The underlying video codec can be either one of the traditional solutions based on DCT (Discrete Cosine Transform) transform and motion compensation (e.g. MPEG-x, H.26x), or one of the more recent codec based on the wavelet 3D transform (e.g. SPHIT).

0010Several schemes exists (see bibliography): overlapping quantization (MDSQ or MDVQ), correlated predictors, overlapped orthogonal transforms, correlating linear transforms (MDTC, e.g. PCT or pairwise correlating transform for 2 MD), correlating filter banks, interleaved spatial-temporal sampling (e.g. video redundancy coding in H.263/H.263+), spatial-temporal polyphase downsampling (PDMD, see below), domain based partitioning (in the signal domain or in a transform domain), FEC based MDC (e.g. using Reed-Solomon codes).

0011A simple scheme for SNR MD is coding of independent video fluxes created by means of MD quantizers, either scalar or vector (MDSQ, MDVQ). The structure of the MD quantizer controls redundancy.

0012A simple scheme for Spatial/Temporal MD is coding of independent video fluxes created by means of Spatial or Temporal Polyphase Downsampling (PDMD). A programmable Spatial or Temporal low-pass filter controls redundancy.

0013As an example, Temporal MD can be achieved by separating odd and even frames, creating two subsequences. Alternatively odd and even fields can be separated. Spatial MD is achieved by separating pixels of 2x1 blocks, so that two subsequences are created. Alternatively four subsequences can be created by separating pixels in 2x2 block. The two techniques can be combined. Each subsequence is then fed into a standard video encoder.

0014Polyphase downsampling (PDMD) for instance is based on a downsampling of pixels of a picture of the video signal. The number of pixels in a row is proportional to a horizontal sampling frequency Fsh, while the number of pixels in a column is proportional to a vertical sampling frequency Fsv.

0015With reference to fig. 1a, where a diagram of the power spectrum, i.e. power P as a function of frequency is shown, given a certain sampling frequency Fs, a spectrum S of the data will extend from the 0 frequency up to the Nyquist frequency, that is Fs/2. It must be understood that the spectrum S of figure 1a is simplified, since for pictures the spectrum will be two dimensional and will extend from 0 up to Fsh/2 and from 0 up to Fsv/2.

0016As can be seen from figure 1a, the spectrum S is subdivided in a high frequency part HS, corresponding to a high frequency range HR, i.e. the Fs/4...Fs/2 range, and a low frequency part LS, corresponding to a low frequency range LR, i.e. the Fs/2...0 range.

0017A downsampling operation DS is performed on spectrum S. A downsampling operation in general is performed by discarding some of the samples. Performing a N:1 downsampling means that only one sample out of N samples survive such a downsampling operation. In the frequency domain, the downsampling operation corresponds to an operation of folding the spectrum around a certain frequency.

0018By way of example, when a 2:1 downsampling operation DS is performed, as represented in figure 1b, a folded spectrum S_f is generated, where the sampling frequency Fs is reduced to its half, Fs/2. Therefore the Nyquist frequency is reduced from Fs/2 down to Fs/4. The high frequency part HS of the spectrum S that was in the high frequency range HR will be folded in the low frequency range LR, the 0 ... Fs/4 range. In particular, the frequencies located in proximity of the Nyquist frequency Fs/2 in the original spectrum S will be folded in proximity of the 0 frequency in the folded spectrum S_f.

0019It must be noted that in figure 1b, as in figure 1c, that will be described in the following, two folded spectrum S_f are shown, since the 2:1 downsampling operation DS originates two descriptions.

0020As another example, if a 3:1 downsampling is performed, the sampling frequency is reduced from Fs down to Fs/3. The Nyquist frequency will be reduced to Fs/6. Frequencies that were above Fs/6 will be folded in the following way: frequencies at Fs/2=3*Fs/6 will be folded at Fs/6, frequencies at 2*Fs/6 will be folded at 0 frequency.

0021In general, thus, when N: 1 downsampling is performed, the sampling frequency is reduced to Fs/N, the Nyquist frequency is reduced to Fs/2/N. The frequencies of the spectrum above Fs/2/N will be folded in the allowed range. Frequencies at n*Fs/2/N, where n is an odd integer index will thus be placed at Fs/2/N, frequencies at n*Fs/2/N, where n is an even integer index will be placed at 0.

0022From the above discussion turns out that, when the PDMD procedure is applied, the high frequencies of the spectrum of the picture are folded over the low frequencies. When standard video codecs are used to compress generated descriptions, such a compression operation CM originates a quantization error Qerr, as shown in figure 1c, that will affect the high portion HS of the folded spectrum S_f. In other words, the high portion HS of the spectrum S is highly quantized, to a higher degree with respect to the low portion LS, since the high portion HS is less important from a perceptive point of view. This means however that, when the folded spectrum S_f, as shown in figure 1d, will be unfolded in an unfolded spectrum Su, by a merging operation US on the decompressed descriptions at the receiver side, the quantization error Qerr will be located in the middle and the effects of such a quantization error Qerr will be therefore quite noticeable.

0023Further, it must be noted that each generated description, as a result of the downsampling operation, will have a folded spectrum with a relevant amount of energy in its high portion. This circumstance makes the task of standard video encoders more difficult, since high frequency coefficients will not be low after transform, probably the quantized coefficients will not be zero and, thus, the entropic coding of quantized coefficient will be inefficient. This means that the compression efficiency will be low, the quality for a given bitrate will be low.

0024A system like the one shown in Fig. 1 is disclosed in the publication by <nplcit id="ncit0002" npl-type="s"><text>Fumagalli et al. "Video transmission over IP using polyphase downsampling multiple description coding", IEEE Conference On Multimedia And Expo, 22 August 2001, pages 1095-1098</text></nplcit>.

0025The topics considered in the foregoing form the subject of extensive technical literature, as witnessed e.g. by: <ul id="ul0001" list-style="dash" compact="compact"><li><nplcit id="ncit0003" npl-type="s"><text>P. C. Cosman, R. M. Gray, M. Vetterli, "Vector Quantization of Image Subbands: a Survey", September 1995</text></nplcit>.</li><li><nplcit id="ncit0004" npl-type="s"><text>Robert Swann, "MPEG-2 Video Coding over Noisy Channels", Signal Processing and Communication Lab, University of Cambridge, March 1998</text></nplcit>.</li><li><nplcit id="ncit0005" npl-type="s"><text>Robert M. Gray "Quantization", IEEE Transactions on Information Theory, vol. 44, n. 6, October 1998</text></nplcit>.</li><li><nplcit id="ncit0006" npl-type="s"><text>Vivek K. Goyal, "Beyond Traditional Transform Coding", University of California, Berkeley, Fall 1998</text></nplcit>.</li><li>Jelena Kova<img file="EP1578133B1_D0001.tif" />ević, Vivek K. Goyal, "Multiple Descriptions - Source-Channel Coding Methods for Communications", Bell Labs, Innovation for Lucent Technologies, 1998.</li><li>Jelena Kova<img file="EP1578133B1_D0002.tif" />ević, Vivek K. Goyal, Ramon Arean, Martin Vetterli, "Multiple Description Transform Coding of Images", Proceedings of IEEE Conf. on Image Proc., Chicago, October 1998.</li><li><nplcit id="ncit0007" npl-type="s"><text>Sergio Daniel Servetto, "Compression and Reliable Transmission of Digital Image and Video Signals", University of Illinois at Urbana-Champaign, 1999</text></nplcit>.</li><li><nplcit id="ncit0008" npl-type="s"><text>Benjamin W. Wah, Xiao Su, Dong Lin, "A survey of error-concealment schemes for real-time audio and video transmission over internet", Proceedings of IEEE International Symposium on Multimedia Software Engineering, December 2000</text></nplcit>.</li><li><nplcit id="ncit0009" npl-type="s"><text>John Apostolopoulos, Susie Wee, "Unbalanced Multiple Description Video Communication using Path Diversity", IEEE International Conference on Image Processing (ICIP), Thessaloniki, Greece, October 2001</text></nplcit>.</li><li><nplcit id="ncit0010" npl-type="s"><text>John Apostolopoulos, Wai-Tian Tan, Suise Wee, Gregory W. Wornell, "Modeling Path Diversity for Multiple Description Video Communication", ICASSP, May 2002</text></nplcit>.</li><li><nplcit id="ncit0011" npl-type="s"><text>John Apostolopoulos, Tina Wong, Wai-Tian Tan, Susie Wee, "On Multiple Description Streaming with Content Delivery Networks", HP Labs, Palo Alto, February 2002</text></nplcit>.</li><li><nplcit id="ncit0012" npl-type="s"><text>John Apostolopoulos, Wai-Tian Tan, Susie J. Wee, "Video Streaming: Concepts, Algorithms and Systems", HP Labs, Palo Alto, September 2002</text></nplcit>.</li></ul>

Object and summary of the invention

0026The object of the present invention is thus to provide an improved arrangement dispensing with the drawbacks of the prior art arrangements considered in the foregoing, proposing a method for Multiple Description Coding that allows a high compression efficiency and high quality for a given bitrate when implemented in encoders and decoders.

0027According to the present invention, that object is achieved by means of the methods having the features set forth in the claims that follow. The invention also relates to a corresponding system as well as a computer program product loadable in the memory of at least one computer and including software code portions for performing the method of the invention.

0028A preferred embodiment of the invention is thus a method for encoding/decoding a video signal sequence that provides for scrambling the spectrum of the picture to be encoded using PDMD coding, in order to raise the efficiency of the compression for generated description, and also to enhance the visual appearance.

0029A particularly preferred embodiment of the invention is directed to a scrambling procedure in the frequency domain, although a scrambling procedure in the pixel domain is also provided.

Brief description of the annexed drawings

0030The invention will now be described, by way of example only, by referring to the enclosed figures of drawing, wherein: <ul id="ul0002" list-style="dash" compact="compact"><li>Figures 1a, 1b, 1c, 1d have been already discussed in the foregoing;</li><li>Figure 2 is a block diagram of an exemplary encoding-decoding system;</li><li>Figures 3a, 3b, 3c, 3d, 3e and 3f are exemplary of an encoding step within the framework of the arrangement described;</li><li>Figures 4a, 4b, 4c, 4d, 4e and 4f are exemplary of an encoding step within the framework of the arrangement described;</li><li>Figures 5, 6, 7, 8 are diagrams representative of test simulations of a system within the framework of the arrangement described.</li></ul>

Detailed description of a preferred embodiment of the invention

0031Figure 1 shows a block diagram of an encoding/decoding system adapted to operate according to the invention.

0032There, reference I indicates an input video sequence comprised of a digital video signal to be transmitted. The input signal I is fed to a pre-processing block 100 that creates a multiple descriptions by way of spatial sub-sampling.

0033The subsequences from the pre-processing block 100 are fed to a set of N encoder blocks, each indicated by the reference 102.

0034In the embodiment shown, coding of independent video flows is created by means of spatial polyphase downsampling (PDMD). A programmable spatial or temporal low-pass filter controls redundancy.

0035The encoder signals from the encoder blocks 102 are sent over a transmission channel C to the receiver side. On the receiver side a set of N H.264 decoder blocks are provided, each indicated 104.

0036The output signals of the decoder blocks 104 are fed to a synchronization block 106 and the signals from this block are sent to the decoder blocks. The synchronization block 106 is also able to effect error recovery.

0037The output signals from the decoder blocks 104 are also fed to a post-processing block 108 that merge the multiple descriptions. The output of the post-processing block 108 is the output sequence O.

0038According to the proposed method, in order to raise the efficiency of the compression for generated descriptions, and also to enhance the visual appearance, the spectrum of the picture to be encoded using PDMD undergoes a scrambling operation.

0039The proposed PDMD procedure will now be described with reference to figures 3a, 3b, 3c, 3d, 3e and 3f.

0040In figure 3a the spectrum S is shown, that is subdivided in a low frequency portion LS in a low frequency range LR, a middle frequency portion MS in a middle frequency range MR and a high frequency portion HS in a high frequency range HR. According to the proposed PDMD procedure, the spectrum S undergoes a scrambling operation, indicated with reference SC, so that the high frequency portion HS of the original spectrum S of the picture is displaced from the high frequency range HR to the middle frequency range MR, as shown in figure 3b. The high frequency portion HS thus swaps with the middle frequency portion MS, originating a scrambled spectrum S_s, so that when a folded spectrum S_f is generated through the downsampling operation DS, such a high frequency portion HS of the original spectrum S appears in the high frequency range HR of each folded spectrum S_f.

0041The example shown in figures 3a-3f refers to the case when two descriptions are generated by PDMD coding, through 2:1 downsampling, like the case already described with reference to figures 1a-1d. In the same way, thus, each spectrum S will be folded along the frequency Fs/4. Therefore, the high frequency portion HS of the original spectrum S is moved at such a frequency Fs/4, in a range where is supposed to undergo a lower quantization by the coder. As a consequence, the middle portion MS of the original spectrum S will be moved under Fs/2 in the folded spectrum S_f. It can be seen from figure 3c, that, after the folding due to the downsampling DS, the high portion HS of the original spectrum S is still located in the high frequency side of the folded spectrum S_f, whereas the frequency position of the low portion LS of the original spectrum S is not affected. This allows for preserving the visual meaning of the picture even after the scrambling operation SC has been performed on the spectrum S and makes the task of standard video encoders more easy, since high frequency coefficients will be low after transform, as can be seen in figure 3d, probably the quantized coefficients will be zero and the entropic coding of quantized coefficient will be quite efficient. Summarizing, the compression efficiency will be high, the quality for a given bitrate will also be high.

0042Also, it must be noticed that when the folded spectrum S_f will be unfolded by applying the merging operation on the decompressed descriptions, as shown in figure 3e, an unfolded spectrum S_u will be obtained, in which the quantization error Qerr will lay in the middle range MR of frequencies. Thus a descrambling operation DC is performed over the unfolded spectrum S_u, obtaining a descrambled spectrum S_d, that is shown in figure 3f. As can be seen, the quantization error Qerr is now moved to the high range HR of frequencies of the descrambled spectrum S_d, where it is less noticeable.

0043In figures 4a-4f the same procedure of figures 3a-3f is shown, where the spectrum S undergoes a scrambling operation SC', that is based on a mirror reflection instead of a swap. The high frequency portion HS of the original spectrum S of the picture is displaced as shown in figure 4b. The high frequency part HS and the middle frequency part MS, considered altogether, are mirror reflected. When a folded spectrum S_f is generated through downsampling DS, such a high frequency portion HS of the original spectrum S appears in the high frequency range HR of each folded spectrum S_f as well.

0044The easiest way to perform the scrambling operation SC on the spectrum S is to transform the corresponding picture into the frequency domain, through a suitable transform matrix having corresponding transform coefficients. Each coefficient will thus correspond to a frequency band. Moving coefficients, i.e. applying permutations to their matrix, is actually moving parts of the spectrum.

0045Any transform operation from the pixel domain to the frequency domain may be used. Of course, the larger the transform size, the smaller will be the subbands that can be moved in the given spectrum.

0046It must be noticed that the scrambling of the coefficients must match the downsampling factor so that high frequencies will be appear to be in the high portion of the folded spectrums.

0047The scrambling procedure just described for the frequency domain can be performed also in the pixel domain.

0048The scrambling procedure in the pixel domain involves isolating frequency bands by bandpass filtering, then such frequency bands can be remodulated to a given frequency by multiplication with a suitable co/sinusoid. Alternatively, frequency bands can be moved to the 0 frequency using a demodulator and a lowpass filter, then they can be moved to the desired frequency using a modulator and a lowpass filter.

0049The described method has been tested using a standard JPEG encoder for the compression of generated descriptions. Black-and-white images have been used to verifying the performance: a first picture PK1 used for testing has a rich spectrum with a lot of energy in the high portion of its spectrum, whereas a second picture PKlhas a relatively poor spectrum with a low amount of energy in the high portion of the spectrum.

0050The classical PDMD coding has been used so that two descriptions are generated by downsampling of a 2x1 pixel window and four descriptions are generated by downsampling 2x2 pixel window.

0051In figures 5 and 6 results for pictures PK1 and, PK2 respectively are shown in the form of Rate-Distortion diagrams, plotting the PSNR (Signal to Noise Ratio Power) in function of rate. The five curves, indicates respectively with C1, C2, C3, C4, C5, represents single description coding, PDMD with downsampling of 2x1 pixel window without scrambling, PDMD with downsampling of 2x1 pixel window with scrambling, PDMD with downsampling of 2x2 pixel window without scrambling, PDMD with downsampling of 2x2 pixel window with scrambling.

0052In figure 7 and 8 the PSNR loss in function of rate of curves C2, C3, C4, C5 with respect to single description coding is shown, that allows for appreciating the reduced PSNR loss for curves C3 and C5.

0053The proposed method and apparatus show several advantages.

0054The proposed method allows for raising the efficiency of the compression for generated description, and also to enhance the visual appearance.

0055The proposed method provides that the high portion of the spectrum is still located in the high frequency part of the folded spectrum, leaving not touched the low portion of the spectrum, preserving the visual meaning of the picture even after the scrambling operation and making easier the task of standard video encoders more easy by keeping low high frequency coefficients will be low after transform.

0056The proposed method can also be applied to the case of PDMD of filtered data.

0057The proposed method can also be applied to a method for encoding/decoding a video signal sequence that provides for generating therefrom multiple description subsequences by means of suitable equations systems, represented by corresponding encoding matrixes designed and optimised using mathematical techniques, and provides decoding matrixes that are the inverse or pseudoinverse of such encoding matrixes for decoding the video signal at the receiver side.

0058Such a method substantially provides for generating at the pre-processing block a multiple descriptions vector <o ostyle="rightarrow"><i>d</i></o>, containing a plurality of descriptions, from <i>d</i>₁ to <i>d_D,</i> where <i>D</i> indicates the number of descriptions. Such a descriptions vector <o ostyle="rightarrow"><i>d</i></o>, is generated by using a system of equations, whose coefficients are represented by a corresponding encoding matrix <i>M,</i> whereas a pixel values vector <o ostyle="rightarrow"><i>p</i></o>, related to a group G of P pixels in the input signal I, <i>p</i>₁...<i>p_p</i>, represent a variable vector. Descriptions vector <o ostyle="rightarrow"><i>d</i></o> is vector of constants generated by a weighted sum of such variables, i.e. by matrix multiplying such a pixel values vector <o ostyle="rightarrow"><i>p</i></o>by the encoding matrix <i>M</i>, according to the following relations-hip: <maths id="math0001" num=""><math display="block"><mover><mi>d</mi><mo>→</mo></mover><mo>=</mo><mi>M</mi><mo>⁢</mo><mover><mi>p</mi><mo>→</mo></mover></math><img file="EP1578133B1_D0003.tif" /></maths>

0059Such a method further provides for merging at the post-processing block an available descriptions vector <i><o ostyle="rightarrow">d</o>',</i> using a further system of equations, represented by a corresponding decoding matrix <i>M</i>' that is in an inversion relationship with such an encoding matrix <i>M</i>. In particular the decoding matrix <i>M</i>' is the inverse or pseudoinverse of such encoding matrix <i>M</i> . In this case pixel values vector <o ostyle="rightarrow"><i>p</i></o> represents the unknown quantity to be calculated, by matrix multiplying the available descriptions vector <i><o ostyle="rightarrow">d</o>',</i> by the decoding matrix M', according to the following relationship: <maths id="math0002" num=""><math display="block"><mover><mi>p</mi><mo>→</mo></mover><mo>=</mo><mi mathvariant="italic">Mʹ</mi><mo>⁢</mo><mover><mi>d</mi><mo>→</mo></mover><mo>⁢</mo><mi>ʹ</mi></math><img file="EP1578133B1_D0004.tif" /></maths>

0060Equations systems, represented by the encoding matrix <i>M</i> and by the decoding matrix <i>M'</i> can be designed and optimized using mathematical techniques and equivalence with a bank of filters can be further exploited for circuit implementation of such matrixes <i>M</i> and <i>M</i>'.

0061Scrambling according to the proposed method can be applied by a suitable permutation of the coefficients of the encoding matrix <i>M</i>.

0062It must be noted that the proposed method applies to audio signals as well.

0063The encoders the operate through lossy algorithms usually apply a highest degree of quantization to subbands that have less perceptive importance. In terms of video signal, this means the high frequency subbands. In terms of audio signal, on the other hand, this means the masked subbands, i.e. the subbands bearing a content that cannot be perceived by a user since adjacent subbands carry a higher perceived power.

0064Since spectrum scrambling aims to place the quantization error introduced by the encoders in subbands less important from a perceptive point of view, after unfolding of the spectrum, the proposed method can be applied also to audio signals, with respect to masked subbbands.

0065The spectrum scrambling must transfer in such subbands the subbands that would be normally quantized by an audio encoder operating on the single not downsampled description.

0066It must be noted that, for due to masked subbands disposition in audio signals, the scrambled and unfolded spectrum has masked subbands in final positions that are different from the original positions in the original spectrum.

0067A possible solution is to perform simultaneous coding of audio multiple descriptions, by coordinating and forcing the choice of the subbands to be quantized in the modified and folded spectra.

0068Another possible solution is to identify a scrambling pattern that does not change the distribution of masks as seen by the independent audio codecs.