Compression of signals using a perceptual model

This invention relates to methods for coding a signal.

The processing of signals for transmission often includes sampling of the input signal, quantizing the samples and generating a set of codes that represent the quantized samples. Most signals of interest (e.g., such as in speech or video signals) are highly correlated, which means that the signal can be thought of comprising a predictable component and an unpredictable component. Coding compression is achieved by encoding essentially only the unpredictable component. Moreover, since these signals are often destined to be received and perceived by humans, concepts that relate to the human perception of the information received have been employed to further compress the coding of such signals and, consequently, the rate of the transmitted signals.

In connection with both speech and video signals, the prior art coding approaches that most closely relate to this invention are transform coding and linear predictive coding.

In a communications system utilizing transform coding, the signal is divided into segments. The segments are sampled and the samples of a segment are transformed into a set of frequency domain transform coefficients. The coefficient signals are then quantized and applied to the transmission channel. In systems that account for noise perception characteristics, the quantization mode applied to the coefficients is made to depend on the signal characteristics and on the sensitivity of the recipient to the resulting quantization noise, achieving thereby coding efficiency. Superimposed on those considerations is the limited bandwidth that is available. Bit allocation is one approach for handling the available bandwidth. In this approach, bits are allocated to the encoding of the transform coefficients in a manner that attempts to achieve a constant bandwidth. Examples of transform coding are found, among others, in U.S. Patent 4,949,383, U.S. Patent 4,184,049, an article by J.D. Johnston titled "Transform Coding of Audio Signals Using Perceptual Noise Criteria", IEEE Journal on Selected Areas in Communications, Vol. 6, No. 2., February 1988, etc.

Linear predictive coding in the speech environment dates back to the mid 1960's. The article by B.S. Atal and M.R. Schroeder titled "Predictive Coding of Speech Signals", Proceedings of the 1967 Conference on Communications and Processing, Cambridge, MA, pp 360-361, is an early example of that. Later, it has been recognized that predictive coding may be improved by taking account of the not unlimited ability to perceive noise. For example, the article by M.R. Schroeder, B.S. Atal and J.L. Hall titled "Optimizing Digital Speech Coders by Exploiting Masking Properties of the Human Ear", Journal of the Acoustical Society of America, December 1979, pp 1647-1652, describes the benefits that may accrue from considering the perceptual characteristics of the human ear.

In Linear Predictive Coding (LPC) that accounts for the perception of noise, a signal segment is predicted from historical information, and an error signal is derived by subtracting the predicted signal from the actual signal. The error signal is typically transformed and weighted by a noise-perception frequency-sensitive function, to result in a modified transform. The modified transform is encoded and transmitted to the receiver.

In the area of video signals the situation is not dissimilar. For example, sub-band coding was applied to image signals by J. Woods and S.D. O'Neil in "Sub-Band Coding of Images", IEEE ASSP, Vol 34, No. 5, October 1986, pp 1278-1288. The arrangement proposed by Woods et al. divides the image into two-dimensional frequency bands and the signal of each band is compressed via DPCM. Two-dimensional frequency bands, in effect, measure the signal variability in the two dimensions that form the image. Vector quantization of video is described, for example, in "Sub-Band Coding of Images Using Vector Quantization" by P.H. Westerink et al., Proc. of Seventh Benelux Information Theory Symposium, pp. 143-150, 1986; and in U.S. Patent 4,811,112 issued to C.W. Rutledge on March 7, 1989. The "human visual system" (HVS) characteristics were incorporated by K.N. Ngan, et al. in an article titled "Cosine Transform Coding Incorporating Human Visual System Model", SPIE Vol 707, Visual Communications and Image Processing (1986) pp. 165-171. The system described by Ngan et al. basically executes a two-dimensional cosine transform on the source information and weights the derived coefficients in accordance with an HVS function. The weighted coefficients are then quantized, coded and sent to a buffer prior to being applied to the transmission medium. To insure a desired global bit rate, a buffer fullness indication is fed back to the quantizer to control the number of bits that are generated by the quantizer. More recently, is a co-pending application Serial No. 07/350435, filed May 4, 1989, publication number US-A-5 309 526, published 3 May 1994; J.D. Johnston and R.J. Safranek disclosed a sub-band analysis method where the quantization schema for each pixel is adapted so that the amount of quantizing noise that is produced is near, but below, the limit of perceptibility. By allowing the quantization noise to rise while still keeping it below perceptibility, greater compression of the signal is achieved.

The above-described coding approaches operate with sampled and quantized signals. To achieve a more compressed code, prior art approaches typically transform the signal to the frequency domain and thereafter operate in that domain. Given a fixed bandwidth, they allocate the available bits between the different frequency components to do as good a job as possible on all of the frequency components or on a prespecified number of them. In other words, the decision that is made is how well to encode the frequency coefficients; not whether to encode them in the first instance. The result is an encoding schema that is more complex than necessary and, when the total bit rate is constrained, is perceptually suboptimal.

EP-A-0 193 143 discloses a coding scheme in which the signal is first divided into bands. Next, the absolute maximum of the entire frequency spectrum is encoded, then the maxima of each band is encoded, relative to the first maximum. Thereafter, each band is divided into sub-bands. In encoding the amplitude values of the sub-bands, the amplitude is quantized into levels, down from the maximum of the band to which the sub-bands belong. Finally, a two bit code is transmitted to designate the range in which the amplitude of the sub-band resides. Those amplitudes that are in the lowest amplitude range get the 00 code, corresponding to a setting to zero, and are therefore effectively discarded.

WO-A-89/10661 discloses a coding scheme which selects for coding and transmission four bands from a plurality of bands, the four bands selected having the highest absolute amount of energy.

According to one aspect of this invention there is provided a method as claimed in claim 1.

According to another aspect of this invention there is provided a method as claimed in claim 10.

The underlying principle of this invention recognizes that good performance is achieved through a simple decision of whether or not to encode certain frequency bands; not how well to encode all or a prescribed number of the frequency bands. Recognizing that an analog signal has a baseband frequency spectrum, and that corresponding to the signal's frequency spectrum there is a "just noticeable difference" (jnd) noise spectrum, it is clear that since signal frequency bands that dip below the jnd spectrum cannot be perceived anyway there is no need to transmit those frequency bands. Furthermore, when the available bandwidth is limited, it makes sense to concentrate on transmitting only those of the signal's frequency bands that most exceed the jnd spectrum.

The jnd spectrum is applied to a decision circuit that selects a number of frequency bands of the signal to be encoded. Each of the selected bands is encoded and transmitted to the receiver.

In one embodiment, for example, the analog input signal is separated into bands through a transform circuit which, in effect, is a bank of bandpass filters. The jnd level within each band is evaluated and a quotient signal is developed for each band which corresponds to the input signal in the band, divided by the jnd signal in the band. A selection circuit identifies the n bands having the highest quotient, and when an analog signal transmission is desired, an inverse transform develops n analog signals which are modulated to form a baseband having a contiguous spectrum. The baseband signal is then directly modulated onto a carrier. When digital transmission is preferred, the outputs of the selection are formatted and applied to the transmission medium.

<u>Brief Description of the Drawing</u>

<ul id="ul0001" list-style="none" compact="compact"><li>FIG. 1 illustrates the spectrum of a signal and the spectrum of noise that might be barely perceived in the presence of the signal;</li><li>FIG. 2 presents the architecture of a coder/transmitter</li><li>FIG. 3 details the structure of selector 120 of FIG. 2;</li><li>FIG. 4 presents the schematic diagram of a switch used in selector 120;</li><li>FIG. 5 presents the structure of a decoder/receiver adapted to the signals developed by the coder/ transmitter of FIG. 2; and</li><li>FIG. 6 presents one embodiment for developing the jnd values in an video signal environment;</li></ul>

<u>Detailed Description</u>

A band-limited time varying signal can be represented by a finite frequency spectrum. Typically, the spectrum is very jagged within the band because information-laden signals don't generally contain the entire spectrum of frequencies. When bands of signals are considered a form of averaging occurs and the spectrum of such a signal is less jagged, so it may have a contour like the one depicted by curve 10 in FIG. 1 (also denoted by the letter S). Curve 10 may represent, for example, a speech signal. A noise signal (one that sounds like noise when it is an audio signal, or looks like a salt and pepper mixture when it is a video signal) typically contains all of the possible frequency components, and has a frequency spectrum that varies very slowly with frequency. A flat spectrum contour is typically referred to as "white noise".

Researchers have established that there exists a threshold below which a noise signal cannot be perceived by most people. This threshold varies with frequency. Because of masking properties, this threshold also varies with the spectrum of the information signal that exists in the presence of the noise signal. This masking phenomenon is well known even to lay people. Noise-like artifacts in an image of a blank wall are easily detected, whereas the same artifacts in an image of a jungle are not detected.

The exact relationship of the detectable noise to the information signal is not important to the principles of this invention, so only an illustrative relationship is depicted in FIG. 1. Curve 20 in FIG. 1, also denoted by jnd, illustrates the variation in the noise threshold as a function of frequency (the jnd spectrum), for the speech signal represented by curve 10. D in FIG. 1 denotes the energy difference between the speech signal and the perceived noise threshold. Actually, the ordinate of the FIG. 1 graph has a log scale, so the depicted difference signal D is really a log of the quotient S/jnd.

As indicated above, jnd spectrum is employed in the prior art merely as a means for modifying, or weighting the frequency coefficients of S, prior to coding of the modified spectrum. Through bit allocation and/or through quantization mode control (control of the number of bits used to quantize the signal) artisans have attempted to do the best job possible in encoding the modified frequency coefficients.

In contradistinction, the jnd spectrum enters into the decision of whether or not to encode the frequency bands; not how to encode them. Additionally, the encoding of this invention is adapted to achieve a constant perceived quality encoding process.

The spectrum of the signal to be encoded is divided into N frequency bands, and a jnd spectrum is computed and applied to a decision means that selects n out of the N bands. Both N and n are parameters that are under designer control. The resulting compression ratio is N/n, which means that for a given N, a smaller value of n yields a greater level of compression. Of course, a greater level of compression also results in lower fidelity at a receiver that is connected to the transmission medium. It can be shown that for a given ratio n/N, the fraction of the retained energy is greater for larger values of N (a finer and hence better selection is made of the peaks in the signal's spectrum). On the other hand, the value of N may be limited by the amount of hardware that the designer is willing to specify.

Whereas the principles of this invention can be implemented with analog circuits (or at least analog circuits and some switching, or sampling), in recognition of the fact that digital implementations predominate today's designs, the following discussion presents a digital embodiment.

In FIG. 2, the input signal is assumed to be a train of samples. This input signal is applied to analysis filter bank 100 via a serial to parallel register 105. Filter bank 100 receives a set of input signal samples with the arrival of every N samples at register 105 and develops therefrom N frequency coefficients. Filter bank 100 may be implemented in a number of ways, such as with a cosine transform circuit, or with generalized quadrature mirror filters (GQMF). See, for example, J. Makhoul, "A Fast Cosine Transform in One and Two Dimensions," IEEE Trans. Acoustics, Speech and Signal Processing, C. ASSP-28, No. 1, Feb. 1980, pp 27-34; and R.V. Cox "The Design of Uniformly and Nonuniformly Spaced Pseudo Quadrature Mirror Filters", IEEE Trans. ASSP, Vol. ASSP-34, No. 5, October 1986, pp. 1090-1096.

The output of filter bank 100 is applied to perceptual model block 110. The function of block 110 is to develop the jnd band signals of FIG. 1. The manner in which those signals are generated is strictly a function of the perceptual model selected and, in fact, while FIG. 2 shows block 110 to be responsive to the output of filter bank 100, it should be appreciated that some perceptual models may call for a connection between the input to bank 100 and the input to block 110, in addition to, or instead of, the connection of the output of bank 100 to block 110.

The concept of a perceptual sensitivity to noise has been studied by a number of researchers. See, for example, R. P. Hellmans, "Asymmetry of masking between noise and tone," Percept. and P Psychophys., Vol 11, pp. 241-246, 1972. Another article that is very informative is "Transform Coding of Audio Signals Using Perceptual Noise Criteria," by J. D. Johnston, IEEE Journal on Selected Areas in Communications, Vol 6, No. 2, Feb 1988, pp 314-323. Building on this information, in a copending application titled Perceptual Coding of Audio Signals, Serial No. 07/423,088, filed October 18, 1989 publication numbers US-A-5 040 217, published 13 Aug. 1991, one actual coder design is disclosed which includes a means for developing the threshold values; i.e., a perceptual motel 110. The teachings of this application are hereby identified by reference. For sake of completeness, however, a FORTRAN program is included herein and summarized below, which develops the outputs of perceptual coder 110 as taught by the incorporated application. A useful reference for understanding the FORTRAN program is FX/FORTRAN Programmer's Handbook, Alliant Computer Systems Corp., July 1988. Tables 1 and 2 present a list of constants used in connection with the illustrative program of Listing 1, such as the absolute thresholds used and the band definitions.

The program comprises primarily the "strt" routine and the "calcthri" routine. The "strt" routine is called initially to calculate various parameters. The "calcthri" routine is called thereafter, with every new block of sampled data, and it develops the output set labeled "ratio". This output set is actually representative of the ratio <maths id="math0001" num=""><math display="inline"><mrow><mfrac><mrow><mtext>S</mtext></mrow><mrow><mtext>jnd</mtext></mrow></mfrac></mrow></math><img file="EP0508581B1_D0001.tif" /></maths> for the different defined frequency bands and it eliminates the need for the separate dividers that are shown in FIG. 2. Noting that the listing's final step in developing the "ratio" is a division (as expected), and that divisions are time consuming, a better realization employs hardware dividers as shown in FIG. 2.

As indicated above, the FORTRAN program included herein computes the <maths id="math0002" num=""><math display="inline"><mrow><mfrac><mrow><mtext>S</mtext></mrow><mrow><mtext>jnd</mtext></mrow></mfrac></mrow></math><img file="EP0508581B1_D0002.tif" /></maths> measure of sensitivity to the perception of noise. Other measures can also be employed, such as<maths id="math0003" num=""><math display="block"><mrow><mfrac><mrow><msup><mrow><mtext>(S - jnd)</mtext></mrow><mrow><mtext>2</mtext></mrow></msup></mrow><mrow><mtext>jnd</mtext></mrow></mfrac><mtext> , </mtext><mfrac><mrow><mfenced open="|" close="|"><mrow><mtext>S - jnd</mtext></mrow></mfenced></mrow><mrow><mtext>jnd</mtext></mrow></mfrac><mtext>,</mtext></mrow></math><img file="EP0508581B1_D0003.tif" /></maths> or the logarithm of any of the above.

FIG. 2 illustrates an embodiment where, as suggested above, the ratio <maths id="math0004" num=""><math display="inline"><mrow><mfrac><mrow><mtext>S</mtext></mrow><mrow><mtext>jnd</mtext></mrow></mfrac></mrow></math><img file="EP0508581B1_D0004.tif" /></maths> in implemented with hardware dividers. Thus, in FIG. 2 the signal of coefficient i where i= 1,2...N, at the output of filter bank 100 (i.e., signal S_i) is divided in circuit 111-i by the jnd output of block 110 that corresponds to coefficient i (i.e., signal jnd_i), to develop thereby a quotient signal q_i. Circuits 111-i are identical read-only lookup tables that develop the quotient signaL With an 8 bit S_i and an 8 bit jnd_i, each lookup table is merely a 2¹⁶ memory. (If time permits, a single memory can be shared.) The output of each of the divider circuits and its associated S_i signal is applied to selector circuit 120. The identity, i, of each signal is also applied to circuit 120. Circuit 120 thus receives N triplets of signals, and its function is to select the n signals S_i (and their identities) that are associated with the n largest quotient signals q_i. The identity information is typically referred to as "side information". As will become apparent hereinafter, this information must be sent to the receiver.

In a digital environment, the signals chosen by selector 120 and their identities are formatted (and perhaps further encoded) in formatter 125 and transmitted to a remote receiver. This is depicted in FIG. 2 by the line marked "DIGITAL". The formatting may be done in a number of ways. One way is to send the information in the form of packets. That is, each signal identity is placed in a header field, and the corresponding coefficient signal is placed in the immediately following data field.

There are situations, however, where it is desirable to employ an analog transmission medium. In such a case, it is useful to convert the coefficients chosen by selector or 120 to a narrow band analog signal with a bandwidth that reflects the compression achieved by the system.

The latter is achieved in the FIG. 2 circuit with inverse transform circuit 130, a parallel to serial register pair 140 and 141, a low pass filter 150, and a formatter 126. Most advantageously, inverse transform circuit 130 is of the type used to realize the transform circuit of bank 100, except that it needs to respond only to sets of n inputs rather than to sets of N inputs. It develops n signals which correspond to time samples. Those time samples are shifted out serially through register 140 and filtered to excise the high frequency components with low pass filter 150. Of course, the clock of register 140 is <maths id="math0005" num=""><math display="inline"><mrow><mfrac><mrow><mtext>N</mtext></mrow><mrow><mtext>n</mtext></mrow></mfrac></mrow></math><img file="EP0508581B1_D0005.tif" /></maths> times slower than the clock of register 105. The developed baseband analog signal is applied to formatter 126 where it is modulated onto a carrier in preparation for transmission. Concurrently, the side information is shifted out to formatter 126 through register 141 where it is encoded, via pulse amplitude modulation, for example, to develop a second analog signal that is modulated onto a carrier in preparation for transmission. The analog transmission is depicted in FIG. 2 by the line marked "ANALOG". It is expected, of course, that either one of the two modes (analog or digital transmission) will be employed. Both are shown in FIG. 2 merely for illustrative purposes.

FIG. 3 depicts one implementation for selector circuit 120. It is based on the Batcher network. See, for example, U.S. Patent 3,428,946 issued February 18, 1969. In that patent, Batcher explicitly teaches how to sort arbitrarily arranged sets of four inputs and eight inputs. He also taught how to merge two bitonic sequences into a single sorted sequence. Applying these teachings directly, FIG. 3 includes four 4-input sort networks 127 and two bitonic merge networks 128. They are arranged to form a bitonic of sequence of length 16 where the top eight inputs are descending in the sort key (the q_i values) and the bottom eight inputs are ascending in the sort key. These networks are structured according to the teachings of Batcher, using switches 125 as described below. The bitonic sequence developed by networks 128 is applied to a modified bitonic merge network 129. Unlike bitonic merge network 128 which contains a log₂M number of stages, where M is the number of inputs of the network, the number of stages in merge network 129 is log₂(M/ n). In FIG. 3, where n is 4 and M= N= 16, the number of stages is log₂ (16/ 4) or 2. Also, because many of the outputs are not utilized, each of the switches 125 within merge network 129 can be simplified, compared to the canonic design of switches 125 within networks 127 and 128.

The structure of switch 125 is presented in FIG. 4. Each switching block 125 receives two input coefficient signals S_m and S_n, the corresponding two quotient signals q_m and q_n and the signal identity values m and n. The function of each switching block 125 is to ascertain which quotient signal is larger and to route that quotient signal, its corresponding coefficient signal, and the identity value, to a selected one of two outputs of three respective switch elements. This is accomplished with a subtractor 121 that is responsive to the quotient signals and with three "double pole - double throw" switch elements 122, 123 and 124 that are responsive to the output of subtractor 121. Switch 122 routes the quotient signals, switch 123 routes the coefficient signals and switch 124 routes the identity values. It may be noted that at the last stage of selector 120, the quotient signals need not be routed because no further decisions need to be made. Also, when log₂ N/n stages are employed in network 129, the n outputs of the network are not ordered; they are only guaranteed to be the ones that correspond to the signals with the largest quotient signals.

Needless to say, the timing of the selection process of circuit 120 must be controlled. It is not useful, for example, to allow circuit 120 to effect a different selection too frequently. Each selection provides a new set of side information that must be sent to the receiver, and that overhead should be kept to a minimum. Accordingly, selector circuit 120 includes a register 126 at the output of network 129 that is used to capture and store each selection. The question is what schema to apply to the clock of register 126. The simplest approach is to use a constant clock. This provides a constant bandwidth to the side information. The clock can be a N^th sub-multiple of the sampling clock of the input signal, or perhaps a KN^th sub-multiple of the sampling clock of the input signal, where K is an integer. A third alternative uses a small value of K (perhaps even K= 1) and at each appearance of the clock a decision is made as to whether or not a sufficient benefit results from changing the selection. This can be done with conventional circuitry (not shown in FIG. 3) that measures the ratio of the selected quotient signals to the unselected quotient signals. By comparing the ratio of the selections at the output of selector 120 to the ratio of the selections at the input of selector 120 a decision whether or not to allow a change in the selection can be made based on the observed difference in the ratios.

FIG. 5 depicts a receiver in conformance with the digital transmission approach of the FIG. 2 transmitter. Since it receives sets of n packets that are not necessarily sorted by frequency, the input signals are applied to a serial to parallel converter 200 which, while keeping each packet in serial form, applies the n packets, in parallel, to sorter 210. Sorter 210 sorts its input signals based on the header field, which identifies the frequency coefficients. The sorted outputs are applied to expander network 220 which routes the signals to appropriate outputs of the expander network. Sorter 210 may be a batcher network, as described above, except that the sorting is keyed to the identifier, i, and not on the quotient signals q_i. Expander network 220 may be a shuffle exchange network as described, for example, is U.S. Patent 4,516,238 issued to A. Huang and S.C. Knauer on May 7, 1985.

It may be noted in passing that the selection process carried out by selector circuit 120 is tantamount to a selection of frequency bands and a down-shifting to baseband of the selected frequency bands. The operation of sorter 210 is a sorting of the selected bands, and the operation of expander 220 is tantamount to an up-shifting of the sorted frequency bands to their appropriate place.

The output signals of expander 220 are applied to inverse transform circuit 230. Circuit 230 is akin to circuit 130, except that it is responsive to N inputs (although only n of them are non-zero) rather than to n inputs. The N outputs of inverse transform circuit 230 are serially shifted out via parallel to serial register 240 and filtered through low pass filter 250 to yield the final reconstructed signal.

For sake of simplicity the above description concentrates on "one dimensional" signals, such as speech. It should be clearly understood, however, that the principles of this invention apply quite well to higher dimensional signals. With video signals, for example (which can be thought to be two dimensional), the only change that needs be made is in filter bank 100, in inverse transform circuits 130 and 230, and in the perceptual model circuit 110. The modifications that relate to the two-dimensional aspects s of video signals are described in the aforementioned 07/ 350435 application. Both the perceptual model and the two-dimensional transform processes are described.

Again, for sake of completeness, the following describes in summary form one embodiment for perceptual model 100 in the video environment. In FIG. 6, N inputs are received from filter bank 100. One of them represents the band where the two dimensional "dc value" of the frame is found. All bands other than dc value band are applied to lookup tables 301. Each of the lookup tables develops a weighted measure of the power in the band. That is, each of the lookup tables develops a value of k_iS^{2<sub2>i</sub2>} where k_i may be different for different values of i. The computed weighted power measures represent an estimate of the visual "texture" of the image at that frequency band. Adder 302 sums the "texture" outputs of the N-1 tables 301 to produce an overall texture estimate. That estimate is applied to lookup table 303 which transforms the power domain output of adder 302 to an amplitude domain masking threshold. The mapping function within table 303 is a log-like function, which reduces the dynamic range at the output of table 303. A brightness correction is introduced in lookup table 304, which multiplies the input from table 303 by the "dc value" band. Lastly, lookup table 305 multiplies the masking threshold developed by table 304 by a set of constants, where each constant is related to the noise sensitivity of the human visual system in each of the frequency bands. The N outputs thus developed form the set of outputs that are applied by perceptual model 100 to dividers 111 in FIG. 2.

Extension to three dimensional cases, such as a time succession of video signal frames or a three-dimensional topographical map, is also straight forward in accordance with the well known principles by which the extension is made from one dimension to two dimensions.<img file="EP0508581B1_D0006.tif" /><img file="EP0508581B1_D0007.tif" /><img file="EP0508581B1_D0008.tif" /><img file="EP0508581B1_D0009.tif" /><img file="EP0508581B1_D0010.tif" /><img file="EP0508581B1_D0011.tif" /><img file="EP0508581B1_D0012.tif" /><img file="EP0508581B1_D0013.tif" /><img file="EP0508581B1_D0014.tif" /><img file="EP0508581B1_D0015.tif" /><img file="EP0508581B1_D0016.tif" /><img file="EP0508581B1_D0017.tif" /><img file="EP0508581B1_D0018.tif" /><img file="EP0508581B1_D0019.tif" /><img file="EP0508581B1_D0020.tif" /><img file="EP0508581B1_D0021.tif" /><img file="EP0508581B1_D0022.tif" /><tables id="tabl0001" num="0001"><img file="EP0508581B1_D0023.tif" /></tables><tables id="tabl0002" num="0002"><img file="EP0508581B1_D0024.tif" /></tables><tables id="tabl0003" num="0003"><img file="EP0508581B1_D0025.tif" /></tables><tables id="tabl0004" num="0004"><img file="EP0508581B1_D0026.tif" /></tables>