METHOD AND DEVICE FOR COMPATIBLE ENCODING AND DECODING OF TELEVISION IMAGES OF DIFFERENT RESOLUTIONS.

Abstract

Dans ce procédé les signaux représentatifs des images sont transmis sur un canal de transmission entre au moins un codeur d'émission et un décodeur de réception lié à un récepteur. Il consiste: - à découper l'image à coder en sous-bandes suivant une même structure arborescente par filtrage et décimation la bande de fréquence spatiale des signaux à transmettre quelque soit leur famille de résolution - à coder indépendamment les signaux à transmettre dans chaque sous-bande suivant un mode de codage intra-image ou un mode inter-images (95). Application: standardisation des systèmes de télévision.

Description

Encoding and decoding method and device compatible with television images of different resolutions This certificate of addition is an improvement to the invention described in main patent application 89 15175 filed in France on November 20, 1989 in the name of the Applicant. The encoding and decoding process as described in claim 1 of the main patent application consists of - to divide each image to be transmitted into sub-bands following the same tree structure by filtering and decimating the spatial frequency band of the signals to be transmitted, regardless of their family of belonging, - to independently encode the signals to be transmitted in each sub-band - to multiplex the coded signals of the sub-bands before transmitting them on the transmission channel and to be performed in the receiving decoder - demultiplexing of the coded signals received in each sub-band, - to decode the received signals relating to each sub-band according to the resolution of the transmitted signals and the receiver's own resolution. When applied directly to encoding points located solely within the same image, this method unfortunately loses its effectiveness rapidly as the transmission channel bit rate decreases. In fact, for bit rates below 1.4 bits per pixel, the visual quality of the observed images is no longer sufficient. The object of the invention described in this application for a certificate of addition is to take into account the natural temporal correlation of image sequences to reduce the bit rate to about 0.8 bits per pixel without apparent degradation of image quality, while retaining compatible coding properties. To this end, the invention relates to a compatible encoding and decoding method for television images of different resolutions according to claim 1 of the main patent, characterized in that it consists of performing inter-image coding based on motion compensation in each of the sub-bands. Other features and advantages of the invention will become apparent from the following description, taken with reference to the accompanying drawings which depict - Figures 1A and 1B show different levels of resolution in spatial-temporal Fourier spaces - Figure 2 shows a one-dimensional, two-channel cell - Figure 3A shows a two-dimensional, four-band separable decomposition cell - Figure 3b shows a segmentation of the spectrum obtained using the decomposition cell shown in Figure 3A - Figure 4A shows an optimal decomposition cell - Figure 4B shows a segmentation of the spectrum obtained using the decomposition cell shown in Figure 4A - Figure 5 shows a frequency segmentation corresponding to a seven-branch tree - Figure 6 shows an embodiment of a FIS (INTRA IMAGE) compatible encoder - Figure 7 shows an embodiment of a compatible decoder for a VT format receiver - Figure 8A shows a simplified embodiment of an inter-image encoder - Figure 8B is a graph illustrating the operation of the quantifier used in the embodiment of Figure 8A - Figure 9 illustrates the positioning of a search window relative to a block of images whose movement is estimated using a line - Figure 10 shows a quantification device implemented by the invention to encode an inter-intra image with motion compensation - Figure 11A shows an embodiment of a sub-band HDP encoder according to a configuration where inter-frame coding with motion compensation is selected - Figure 11B shows an embodiment of an EDP-compatible decoder with an inter-frame configuration and motion compensation - Figure 12 shows an inter-image compatible encoder - Figure 13 shows an embodiment of an inter-image decoder for an arbitrary sub-band. The method according to the invention which is described below is based on the same principle as the method already described in the main patent application, however it differs from it in that the tool takes into account the temporal correlation of the image sequences by introducing an additional inter-image coding. As already indicated in the main patent application, a decomposition into sub-bands of the two-dimensional spectrum of the image, considering only progressive formats, can be envisaged in many ways depending on the desired shape of the sub-bands obtained by rectangular, quincunx, hexagonal, etc. decimations, depending on the hierarchical or parallel structure used of the filter bank and finally depending on the decomposition filters used, whether these filters have a finite or infinite impulse response, separable or not.Nevertheless, for both practical and theoretical reasons—particularly the need to easily design sub-band filters that ensure perfect, or near-perfect, reconstruction of the original signal in the absence of any sub-band signal quantization, and the desire for a modular system that facilitates finding the optimal sub-band decomposition based on the desired bitrate—it is preferable to choose a hierarchical system. This is the only system that allows for nested resolution levels. This enables the image spectrum to be segmented as desired, starting from a single specification of an elementary analysis / synthesis cell. The advantage is that several elementary cells can then be arranged in a tree structure, allowing certain parts of the spectrum to be decomposed to varying degrees depending on their informational content.Finally, to create the rectangular frequency bands required by the coding compatibility constraints between the various video standards currently under consideration (HDP, HDI, EDP, TV, and VT), whose spectra are shown in Figures 1A and 1B, an approach using separable filters is preferable to any other because it allows the rows and columns of the image to be processed separately using one-dimensional filters. As already described in the main patent application, interlaced images can be associated with equivalent progressive images having a rectangular band spectrum, and the spatial spectrum of the image can then be divided according to a four-band frequency tree structure. The design of the analysis-synthesis system can then be based on the significantly simpler design of a two-channel, one-dimensional analysis-synthesis system.A one-dimensional cell comprises, as shown in Figure 2, two filters 11 and 12 respectively low-pass and high-pass with impulse responses HOn and Hln and two interpolation filters 17 and 18 with impulse responses GOn and Gln coupled together by devices 13 and 14 for subsampling and 15 and 16 for oversampling. The two-dimensional separable four-band decomposition cell shown in Figure 3A comprises a set formed by cascading three one-dimensional cells in a two-stage tree. The first stage, formed by filters 20 and 21 with transfer functions HOn and Hln, filters the columns of an image, and the second stage, formed by filters 240 and 243 with transfer functions HOn and Hln, filters the rows of two sub-band images obtained from the first stage. Cascading such cells thus allows the image spatial spectrum to be decomposed according to a "quadtree" structure of a four-band frequency tree, in which the hierarchy of the initial image is first decomposed into four sub-bands, then some of the sub-bands are further decomposed, and so on. It is possible to use truncated trees to, for example, decompose lower frequencies more than higher frequencies.In the embodiment shown in Figure 2, filters 11 and 12 must provide a perfect reconstruction, or at least one better than 50 decibels, of the signal Xn applied to their input. For their implementation, the choice of filters with a finite impulse response is preferable because they can have a strictly linear phase and avoid numerical problems (particularly stability) in finite arithmetic. Under such conditions, a good choice of filters 17 and 18 eliminates inter-band aliasing, and any pair of solutions for the choice of filters 11 and 12 can be obtained by factoring a half-band filter.However, only a portion of this set of solutions is truly usable for an application such as coding, which imposes a crucial condition regarding the quantization of the sub-band signals y0 and Y1 obtained from the outputs of the subsampling devices 13 and 14: namely, that the variance of the reconstruction error must equal the sum of the variances of the quantization noise for each channel. This condition, in addition to the requirement for perfect reconstruction of the original signal without quantization, is only met with filters known by the abbreviation "CQF" (conjugate quadrature filters), in which the impulse response of filter H1 is obtained from the impulse response of filter HO by time inversion and sign reversal of every other coefficient. Furthermore, the condition for perfect reconstruction requires that filter HO have an autocorrelation function that is a half-band filter.These filters necessarily have an even length and can be determined for practically any spectral template. Their main drawback is that they cannot have a linear phase once their length exceeds 2; however, good zero matching during the factorization of the resulting filter product can yield an almost linear phase. Another, perhaps less fundamental, condition to verify when implementing the filters is the decorrelation of the sub-band signals YO and Y1 obtained at the output of subsamplers 13 and 14, so that the sub-bands can be encoded independently of each other with near-optimal performance.However, this condition is too stringent to satisfy. If the transfer functions HO and H1 of filters 11 and 12 form a pair of CQF filters—that is, if the transfer function H1 is obtained from the transfer function HO by time inversion and changing the sign of every other coefficient—and if filters 11 and 12 have even lengths and are linear in phase, then a point-to-point decorrelation of the YO and Y1 signals is obtained. The resulting filters are known by the abbreviation QMF, "quadrature mirror filter," but can only perform an approximate reconstruction of the original signal. Nevertheless, the signal-to-noise ratios obtained during reconstruction are greater than 50 decibels as soon as the filters have sufficient lengths. Furthermore, no phase distortion is introduced since the transfer function of the overall analysis-synthesis system is then linear in phase.The choice of one type of filter or the other can only result from a compromise between good reconstruction and inter-band decorrelation. Intra-band decorrelation, on the other hand, is determined by the choice of the decomposition tree, an arrangement of two-dimensional, four-band cells. Naturally, the more a band is segmented, the lower the correlation within it, but conversely, the longer the corresponding reconstruction filter, the greater the risk of quantization noise propagating into the reconstructed image. A good compromise is represented by the 16-band decomposition tree described in Figure 4. Regardless of the image to be segmented, this tree provides good intra-band decorrelation.Once decomposed, each strip is linearly quantized. The quantization step for each strip must be a function of the number of cuts it has undergone to obtain optimal quantization in the sense of the minimum variance of reconstruction terror: more precisely, the quantization step of strip i must be proportional to 4-ki, where ki is the number of times strip i has been cut into four strips. Thus, strips located at the same level of the decomposition tree are quantized with the same step, and an additional cut into four strips halves the step. After quantization, each strip is encoded using variable-length codes defined as described in patent application 2,627,337 filed in the name of the Applicant, with a description by zero range.However, in the baseband (the one that has always been filtered by a low-pass filter) the use of fixed-length codes is preferable since its statistics are very dependent on the original image. In the intra-compatible coding scheme described in the main patent application, interlaced formats are handled by using their equivalent progressive signal. Every interlaced signal is first transformed into an equivalent progressive signal by a deinterlacer and then decomposed like any progressive signal, except for the initial decomposition into two horizontal bands, which is not performed. Therefore, in what follows, interlaced signals are no longer considered, and only the progressive formats HDP, EDP, and VT are taken into account. However, the same process can easily be extended to the case of interlaced HDI and TV signals by considering their associated progressive signal and proceeding as in the main patent application, incorporating a deinterlacer into the encoder and an interlacer into the decoder in the case of an interlaced receiver.The intra-compatible encoding used is a scheme known as FIS (format independent splitting), in which the signal to be encoded is always decomposed in the same way regardless of its resolution. Therefore, for example, to ensure three-level compatibility, a split into at least seven bands, as shown in Figure 5, is necessary. The corresponding tree also ensures compatibility with interlaced HDI formats. TV. Naturally, independent coding and transmission must be ensured, that is to say, separated by unique synchronization words, groups of bands (1, 2, 3) [part HFC], (5, 6, 7) [part IFCj and band 4 [part BB]. An embodiment of a corresponding compatible encoder is shown in Figure 6A. This encoder comprises a BBC encoding block 29, an IFCC encoding block 30, and an HFCC encoding block 31. The outputs of the encoding blocks 29, 30, and 31 are connected respectively to the corresponding inputs of a multiplexer circuit 32. Sub-band segmentation is achieved by means of low-pass filters referenced 32 to 37 respectively, coupled to subsampling circuits (not shown), and by means of high-pass filters referenced 38 to 43 respectively, also associated with subsampling operators (not shown). In addition to these good performance characteristics, the encoding device of Figure 6A ensures complete transparency between the different HDP sampling formats. HDI, EDP, TV, and VT, as described previously, are available regardless of the BBC, IFCC, and HFCC coding blocks. I1 also allows for a sixth format corresponding to the definition of the interlaced telephone video signal. The signal to be encoded, regardless of its format (HDP, EDP, or VT), is applied to the device's input marked IP and is divided into seven sub-bands, referenced from S1 to S7, according to a decomposition tree similar to the diagram in Figure 5. Sub-band S1 is obtained through filters 42 and 43. Sub-band S2 is obtained through filters 37 and 43, sub-band S3 through filters 41 and 32, sub-band S4 through low-pass filters 32 to 35, sub-band S5 through filters 36, 40, 33, and 32, and sub-band S6. S6 is obtained through the sequence of filters 38, 34, 33 and 32, and sub-band S7 is obtained through filters 39, 40, 33 and 32. Regardless of the type of BBC, IFCC and HFCC, the compatibility properties are verified. In fact, they constitute additional cuts of the sub-bands S1 to S7 in order to obtain the optimal 16-band tree represented in Figure 4. Each of the bands that is then obtained is followed by a linear quantizer Qi and a variable-length encoder VLC. as shown in Figure 6B, which depicts one embodiment of the BBC 29 encoder. In Figure 6B, the sub-band S4 is again divided into two sub-bands by a low-pass filter 45 and a high-pass filter 46. The sub-band obtained at the output of the low-pass filter 45 is further divided into two sub-bands by a low-pass filter 47 and a high-pass filter 48. The resulting sub-bands are then quantized by quantization circuits 49 and 50, respectively, and the quantized signals are subsequently encoded using variable-length encoding devices 51 and 52.Similarly, the sub-band obtained at the output of filter 46 is itself divided into two sub-bands by a low-pass filter 53 and a high-pass filter 54. The sub-bands thus obtained are also quantized by two quantization circuits 55 and 56, and the resulting quantized signals are encoded by variable-length encoding devices 57 and 58. The outputs of the variable-length encoding devices 51, 52, 57, and 58 are connected respectively to the inputs of a multiplexing circuit 59, the output of which is connected to a first input of the multiplexer circuit 32 in Figure 6A. Likewise, the IFCC encoder 30 performs only quantization and variable-length encoding of the S5 bands. S6 and S7. Finally, the HFCC 31 encoder cuts each of the bands S3 and S2 are split into four additional bands. Band S1 remains unchanged. For interlaced signals, the decomposition procedure is not exactly the same. To avoid performing a three-dimensional decomposition involving a staggered decomposition in the temporal and vertical frequency planes, which could seriously compromise compatibility, an interlaced signal is first transformed into an equivalent progressive signal. The equivalent progressive signal is obtained by placing the interlaced signal applied to the IE input of the device into a deinterlacer 44 in Figure 6A. The aim is to vertically phase-shift each odd frame by half a pixel; this is achieved simply using a one-dimensional vertical interpolating filter.The resulting output is a progressive signal with half the vertical dots of a progressive signal with the same horizontal resolution, corresponding to the low-frequency band of the first encoder split into two bands (split into two horizontal bands by filtering along the columns). The signal obtained from the deinterlacer 44 is then applied to the common connection point of the low-pass filters 32 and 33. This signal is subsequently decomposed in the same way as any progressive signal, but naturally, in this case, it only yields five sub-bands, as sub-bands S1 and S2 cannot be formed. In the subsequent embodiments, the optimal 16-band encoder just described will be used. An embodiment of a corresponding decoder is shown in Figure 7 for a VT receiver. This decoder performs the inverse functions of the encoder in Figure 6A. The complexity of this decoder depends primarily on the associated processing speed. It is similarly constructed to the decoding device corresponding to the decoding tables in Figure 13 of the main patent, by a set of decoding blocks numbered 60 to 65 that receive the digital frame of the video signals from a transmission channel 66 via a multiplexer circuit 67 and a series-connected logic switching block 68. The decoding blocks 60 to 62 are blocks BBC 1, decoding blocks 63 and 65 are IFCC 1 decoding blocks and block 64 is an HFCC 1 decoding block. A set of summing circuits and low-pass and high-pass filters, referenced from 69 to 86, ensures, in the manner shown in the tables of Figure 13 of the main patent application, the reconstruction of the telephone video signal and its application to the corresponding receiver, referenced as 87 in Figure 7. In Figure 7, logic block 68 applies the useful portion of the binary stream received on channel 66 to the appropriate inputs of decoders 60 to 65. Under these conditions, the decoders associated with the receiver are in EDP and HDP formats and can be easily deduced. In the case of an EDP decoder, decoder 60, for example, must be removed, and in the case of an HDP decoder... For HDP, the IFCC-1 63 and 65 decoders must also be removed. For interlaced decoders, the principle remains the same, and in this case, the inputs that are never used must be removed from the VT decoder, according to Table 13 of the main patent application. Finally, at the decoder level, for a given receiver format, some of the hardware is not used if the transmitted image has a higher resolution; however, the decoders will always run at the same speed regardless of the resolution of the transmitted image. This speed is that of the receiver associated with the last reconstruction stage and decreases towards the input stages, as it is halved at each stage. The intra-image coding scheme just described unfortunately loses its effectiveness as soon as the transmitted bit rate decreases, since the visual quality obtained with this type of coding is no longer sufficient. The device described below takes into account the temporal correlation of image sequences, allowing for even lower bit rates on the order of 0.8 bits per pixel. The problem is solved by using a hybrid scheme through temporal prediction combined with compensation for estimated motion within the image. Although this scheme is classic and is already applied, for example, to cosine transform encoders, the originality here lies in its application to the sub-band switching scheme described previously, while ensuring compatible coding properties. The simplest possible temporal predictive scheme (INTER PUR) is shown in Figure 8A.It consists, according to a differential coding method, of subtracting from the pixel to be coded x(k,l) of the image at time n the coded value of the opposite pixel in the previous image at time nl and then coding this difference. As shown in Figure 8A, it comprises a quantizer 88 coupled at its input to the output of a subtractor circuit 89. The output of the quantizer circuit 88 is coupled to the transmission channel and to a dequantizer circuit 90, whose output is connected to an adder circuit 91. A delay circuit 92 connects the output of circuit 91 to a first input of the subtractor circuit 89 and to a second input of the adder circuit 91. According to this scheme, the pixel x(k, 1, n) with address k,l in the image n is applied to the encoder input on the second input of the subtractor circuit 89. The circuit 89 subtracts from the value of the pixel x(k,l,n) in the image at time n the decoded value of the pixel x(k,l,n-1) at time nl.The difference d(k,l,n) obtained is, under these conditions, equal to. d(k,l,n) = x(k,l,n)-x(k,l,nl) In this scheme, it is the coded value of the corresponding pixel in the previous image that is taken into account to avoid any drift phenomenon of the decoder, that is to say, in order to do exactly the same thing temporally to the encoder as to the decoder. Quantizer 88 associates a value d(k, 1, n) with the difference d(k, l, n), which is actually a quantization level number. Dequantization is performed using dequantizer 90, which associates its corresponding reconstruction value to this level number. This reconstruction value is an approximation of the difference d(k, 1, n) affected by quantization noise. The reconstruction value d(k, l, n) obtained at the output of dequantizer 90 is shown in Figure 8B as a function of the difference d(k, l, n) calculated by subtractor 89. Under these conditions, it is clear that the predicted value x(k, l, nl) must be used, and not the value x(k, l, n-1), to recalculate the predicted value x(k, l, n) at the decoder. In this case at the decoder level the recalculated value x(k,l,n) is equal to x(k,l,nl)+d(k,l,n) and only the quantization error committed on the difference d(k,l,n) intervenes.Indeed, the coding error is equal to x(k,l,n)-x(k,l,n) or e(k,l, n) =d(k, 1, n) +x(k,l, n-1) -x(k,l, n) or d(k,l,n) -d(k,l,n-1). Using the pixel value x(k,l,n-1) as a predictor, previous coding errors of pixels x(k,l,i) for i < ns would accumulate, creating a drift phenomenon. In fact, we obtain here a local decoder at the encoder level itself. Since the scheme just described is very simple, there is a risk that it will lose its effectiveness when there is any movement in the image sequences, the prediction becoming very poor and thus resulting in a very large prediction error that is much more difficult to encode. This risk can, however, be minimized by improving the prediction through the introduction of motion estimation and image registration. Thus, to improve the prediction, the method according to the invention consists of introducing motion estimation and image registration. This method consists, instead of removing the decoded pixel opposite the previous image (x(k,l,n-1)), of removing the pixel x(k',l',n-1) with coordinates k' and l' calculated based on the local motion. A corresponding device for obtaining this result is shown in Figure 10. In addition to the elements of Figure 8A, which are shown with the same reference numerals, it comprises a block-forming device 93, a motion estimator 94, an inter-intra-image switching device, and a motion compensation device 96. The device allows for simple quantification of an image associated with an inter- or intra-image pattern with motion compensation. The motion estimator device 94 estimates the motion between two successive images by executing the well-known "block-matching" algorithm, which calculates an intercorrelation between two successive images and estimates the corresponding motion vector from the resulting correlation peak.Each image is divided by the block-forming device 93 into disjoint NxN blocks that completely cover the image, and a motion vector is estimated for each block. In this calculation, the number N is assumed to be even. For each block B(k, 1, n) as shown in Figure 9, the nearest block B(k', l', nl) is found within a search window in the previous image, according to a certain criterion. The estimated motion vector for each point in block B(k, i, n) is given the coordinates k'-k, l'-l. The search window used has an extent of 2M × 2L and is centered on block B(k, i, n). The criterion used is an energy criterion consisting of calculating a sum of squared differences or a sum of absolute differences.The motion registration performed by block 96 is achieved by searching for the corresponding block in the previous decoded image for each block in the current image. When the components of the motion vector are integers, meaning they correspond to an integer number of pixels, the problem is reduced to a simple memory addressing issue. However, when the motion vector is not an integer, its values ​​must be interpolated based on neighboring pixels (bilinear interpolations can be used). The choice between intra- and inter-mode at each block is made by block 95. This choice is based on an energy criterion. It involves calculating the energy of the intra-block to be encoded and the energy of the prediction error block (inter-block), provided by subtractor 89. This choice consists of selecting the mode (inter- or intra-) that yields the minimum energy.In the embodiment shown in Figure 10, the delay element 92 is an image memory. This memory contains the decoded previous image, and it is this image that is generally used to perform motion estimation. However, it is also possible to use the uncoded previous image. Subband decomposition is very well suited to the context of motion estimation and compensation because the geometric properties are preserved in the subband images, and motion estimation and compensation can be performed both at the full-band image level (i.e., before subband decomposition) and at the subband level. However, compatible coding constraints dictate that motion compensation must be performed at the subband level, not at the undecomposed image level. Indeed, if motion compensation were performed at the undecomposed image level, the processing at the encoder and associated compatible decoder levels (i.e., decoding only a portion of the transmitted image) would differ, inevitably leading to drift.This case of motion compensation at the level of the undecomposed image is illustrated in Figures 11A and 11B, and the drift phenomenon that appears will be highlighted below. To simplify the explanation, the operation of the encoder and decoder shown in Figures 11A and 11E is explained below, assuming that the decoder shown in Figure 11E is the associated EDP-compatible decoder that decodes the EDP portion of the HDP signal provided by the encoder in Figure 11A. In Figure 11A, the components analogous to the encoding device shown in Figure 8A are represented with the same reference numerals. Unlike the components shown in Figure 8A, it includes two quantizers 88a and 88b to quantize, respectively, the portions relating to the EDP signal and the complement of the HDP signal, and two corresponding dequantizers 90a and 90b.They also include an HDP 97 sub-band decomposition device and an HDP 98 sub-band reconstruction device. The delay device 92 consists of an image memory. A motion compensation device 96 is also introduced, as shown in Figure 10, between the image memory 92 and the subtraction circuit 89. The motion compensation device 96 receives motion vectors vx and vy from a motion estimation device shown in Figure 11A, which has characteristics identical to the device 94 shown in Figure 10. The HDP subband decomposition device 97 is interposed between the output of the subtraction circuit 89 and the inputs of the quantization circuits 88a and 88b. The HDP subband reconstruction device 98 is interposed between the outputs of the dequantization circuits 90a and 90b and the input of the summing circuit 91.According to this embodiment, the signal d(k,l,n) obtained at the output of the subtractor circuit 89 is equal to the difference between the signal x(k,l,u) applied to the "+" input of the subtractor circuit 89 and the reconstructed signal x(k+vx,l+vy,n-1) provided by the motion compensation block 96, where vx and vy are the components of the estimated motion vector provided by the motion estimator 94. After sub-band reconstruction performed by the reconstruction block 98, the signal d(k,l,n) is recovered, which in fact represents the signal d(k,l,n) affected by the quantization errors of the sub-band signals. The reconstructed signal x(k,l,n), equal to the signal x(k+vx,l+vy,nl) + d(k,l,n), is applied to the input of memory 92 by the output of the adder circuit 91. The overall quantization error is given by the relation e(k,l,n)=x(k,l,n)-x(k,l,n) or also e(k,l, n)=d(k,l, n) +x(k+vx,l+vy, n-1) -x(k, l, n) or e(k,l,n) =d(k, 1, n) -d(k,l, n) The preceding relationship shows that the overall quantization error is identical to the quantization error applied to the prediction error and that there is no drift phenomenon. The corresponding EDP-compatible decoder shown in Figure 11B comprises, connected to the transmission channel 99, a dequantizer 100 coupled at its output to an EDP subband reconstruction device 101. It also includes a motion compensation device 102 and an image memory 104. An adder circuit 103 is coupled via a first operand input to the output of the EDP subband reconstruction device 101 and via a second operand input to the output of the motion compensation device 102. The output of the adder circuit 103 is connected to the input of the image memory 104 and provides the reconstructed EDP-compatible signal xc(k, 1, n). In the decoder of Figure 11B, the sub-band reconstruction device 101 EDP ​​provides the filtered, undersampled, and quantized portion dc(k,l,n) of the signal d(k,l,n). The reconstructed signal xc(k,l,n) is obtained at the output of the adder circuit 103 by adding the signal dc(k,l,n) and the signal xc(k+vx / 2,1+vy / 2,n-1) provided by the output of the motion compensation device 102. In this example, it is assumed that the motion vectors vx / 2 and vy / 2 are integers in order to avoid taking into account the interpolation that would be necessary otherwise. The coding error made with respect to the signal xc(k,l,n) (which represents the PDE part of the signal x(l,l, s)) is equal under these conditions to ec(k, i, n) = xc(k, l, n) ~ XC(k,l,n) or ec(k,l,n) = Xc (k,l,n)-x c (k+vx / 2, l+vy / 2, n- 1) -dc (k, I, n) As soon as the motion vector is non-zero, the previous relationship shows that the error is not solely due to the quantization of the prediction error of the signal dc(k, i, n), since the motion estimation is not performed at the signal level. PDE. And so there is an introduction of a recursive error term which creates the drift. The previous example gives a major reason to perform motion compensation at the subband level rather than at the fullband image level. A second reason stems from the fact that an inter / intra mode choice is made for each block. Thus, if motion compensation and the inter / intra choice were performed on the undecomposed image, a set of disparate inter and intra blocks would have to be divided into sub-bands. However, the artificial high frequencies introduced at the boundaries between inter and intra blocks would create bounces in the sub-band images, making them much more difficult to encode, and under these conditions, the scheme would lose its effectiveness. One possible solution would be to perform the inter-intra choice at the subband level, while simultaneously performing motion compensation on the full-band image. This would involve compensating the entire image for motion and decomposing both the intra and compensated images into subbands. The inter-intra decision would then be made within the subbands, using homothetic blocks corresponding to the motion estimation blocks in the full-band image. However, such an approach is very costly in terms of the hardware implementation of the encoders and decoders, as two subband decompositions must be performed. Furthermore, block-based motion compensation risks creating artificial boundaries in the compensated image, particularly due to the non-uniformity of the estimated motion field. This could lead to bouncing problems in the subbands, and the inter-intra choice might thus be biased in favor of the intra.It is therefore preferable to perform motion compensation at the sub-band level, and this is the approach adopted in the invention. In this case, motion estimation can be performed at both the full-band image level and the sub-band level. However, since the physical motion is the same in each sub-band image, it is preferable to use a single set of motion vectors rather than estimating the motion in each band. The latter approach also introduces redundancy between the motion vectors of each band, and fairly sophisticated encoding of these vectors is necessary to avoid excessive data loss during transmission. By performing motion estimation at the full-band image level, this estimation can be done between two successive source images or between a source image and the previous decoded image. The first solution provides a more physical representation of the motion but may be less suitable for encoding than the second. Furthermore, full-band estimation allows for greater accuracy in the motion vectors. An embodiment of the corresponding encoders is shown in Figure 12.This encoder contains as many analogous encoding devices to the one shown in Figure 10 as there are sub-bands. In Figure 12, each analogous encoding device has the same elements as those in Figure 10, assigned the same references with an index i corresponding to the sub-band. After decomposition of the original image into sub-bands (filter bank 93), the sub-bands are transmitted to each encoding device using a block-forming device B. A motion estimation device 94 is coupled to a motion vector scaling device 105 to apply the motion parameters respectively to each motion compensation device 96. The reconstruction device 106 is connected to the outputs of the memories 92 to reconstruct a starting image and apply it to the motion estimation device 94. Variable-length encoding devices 107. are coupled at the output of quantization devices 88. and a multiplexer circuit 108 coupled at the output of variable length coding devices 107. to transmit on the transmission channel the result of the coding through a buffer memory 109. An embodiment of the corresponding decoder for any of the bands is shown in Figure 13; it comprises, connected in this order, a variable length code decoding device 110, a decoding device 111, and an intra- or inter-image mode selection device. In intra-mode, the decoded information is directly transmitted to the reconstruction filter bank (not shown). In contrast, in inter-mode, the information is transmitted to the first input of an adder circuit 113. The second input of the adder circuit 113 is connected to the output of a motion compensation device 114, which is connected to the output of an image memory 115 and controlled by the motion information received from the transmitting encoder. The memory 115 is connected via its input to the common output of the intra-inter image selection device 112 and to the output of the adder circuits 113. Depending on the encoder and decoder embodiments just described, the images are divided into NN blocks. A motion vector is estimated for each block and transmitted to the receiving decoder. Variable-length coding can be used for the motion vectors.These vectors are rescaled according to the subsampling factor by dividing each component by this factor, so that they can be used in each band where motion compensation is performed on the blocks corresponding to the estimation. These "corresponding blocks" were scaled homothetically in the same way as the vectors (size divided by the subsampling factor). It should be noted, however, that the phase shifts introduced by the filtering must be taken into account when locating the corresponding blocks. In the case of using linear-phase filters of length L, a shift of length (L-1) / 2 is introduced into the filtered image. This offset must be compensated. However, since the length L is even, there is always a phase shift of + pixel, which is negligible, especially as it is reduced to i pixels with subsampling. This compensation is easily achieved using a delay line. If the accuracy of the motion vectors estimated at the full-band image level is 1 pixel (components estimated as an integer number of pixels), then the accuracy of the corresponding vectors is i in band 1, i in bands 2 through 12, and 1 / 8 in bands 13 through 16, assuming of course the use of the 16-band decomposition tree shown in Figure 4. A generally bilinear interpolation is therefore necessary to perform the compensation. Furthermore, N, the size of the motion estimation blocks at the full-band image level, can be of the form k.21 and 1 must be at least equal to 4 to obtain compensation blocks larger than 2x2 at the lowest bands. Generally, k can be set equal to 1. Finally, for each block and for each band, an inter-intra decision is made according to an energy criterion. However, it is primarily in the low-frequency bands that the interpolation mode appears most useful. In Figure 12, the memories 92 contain decoded sub-bands from previous moments and can therefore, via the reconstruction filter bank 106, create the previous decoded image (local decoder to the encoder), which can then be used for motion estimation performed by block 94. Separating the motion compensation blocks 96 by band is very advantageous from a hardware perspective, as it allows for parallel processing. Smaller memories running at lower speeds are required compared to motion compensation at the full-band image level. Similarly, interpolations can be performed at reduced speeds.As for the compatible encoding-decoding process, everything happens as in the case of the intra-schemes described in the main patent application and in the table shown in Figure 13 of the main patent application.

Claims

DEMANDS 1. A method for coding and decoding compatible television images of different resolutions according to claim 1 of the main patent characterized in that it also consists of performing inter-image coding (95i' 88i, 67i) in each of the sub-bands.

2. Method according to claim 1 characterized in that it consists, in order to perform the inter-image coding, of performing an estimation of the moving points (94) at the level of the full-band image, before performing a compensation of this movement in each sub-band.

3. A method according to any one of claims 1 and 2 characterized in that it consists of performing a differential coding (89i) of the points located opposite the current image and the previous image.

4. A method according to any one of claims 1 to 3, characterized in that it consists, before decomposing the image into sub-bands - to cut the image into blocks (93) and to determine for each current block a motion vector by means of a prediction block obtained by comparing each current block to its counterpart inside a window in the previous image.

5. Method according to claim 4 characterized in that the image is cut into sub-bands by means of filters (20, 21, 24i) associated with sub-sampling circuits (22, 23, 25 ... 27) and in that the phase shift introduced by the filters is compensated to allow the localization of the blocks before performing the compensation of the motion in each sub-band.

6. A method according to any one of claims 4 and 5 characterized in that it consists of matching each block of the fullband image with a homothetic block in each subband image.

7. A method according to any one of claims 2 to 6 characterized in that the estimation of moving points (94) at the full-band image level provides motion vectors for each block of the full-band image and in that the motion vectors are divided according to the sampling factors used to obtain the sub-bands considered, before being applied to the blocks homothetic to the blocks of the full-band image.

8. A method according to any one of claims 4 to 7 characterized in that it consists of choosing (95) the intra-image coding mode or the inter-image coding mode by calculating the energy of the current block and the energy difference between the current block and that of the predicted block in order to choose the coding mode whose energy is minimal.

9. A method according to any one of claims 1 to 8 characterized in that the coding takes place on interlaced or progressive signals, the coding of the interlaced signals taking place after they have been transformed into equivalent progressive signals.

10. Method according to claim 9 characterized in that it consists of cutting each progressive signal into the same determined number of sub-bands regardless of its format.

11. A method according to any one of claims 1 to 10 characterized in that the estimated motion vectors are scaled (105) before performing motion compensation in each sub-band.

12. Device for implementing the method according to any one of claims 1 to 11 characterized in that it comprises for each sub-band an inter-intra (88il 107i) image encoder controlled from a motion estimation device (94).

13. Device according to claim 12 characterized in that each encoder of a subband comprises a quantizer device (88) coupled to a differential coding device via a switching device (95) to quantify each pixel supplied by a block directly when the intra-image coding mode has been chosen or to quantify its difference with respect to its motion-vector corrected counterpart in the previous image when the inter-image coding mode has been chosen.

14. Device according to claim 13 characterized in that each differential coding device is coupled to a motion estimation device (94) of the image points via a motion compensation device (96i).

15. Device according to claim 14 characterized in that each motion compensation device (96i) is coupled to the motion estimation device (94) via a scaling device (105) for the motion vectors provided by the motion estimation device (94), so that the motion estimation is done at the level of the undecomposed image while the motion compensation takes place on the sub-bands.