Patent Application
20200226791
The invention relates to a method for compressing an input data set comprising a sequence of coefficients into a compressed data set and to a method for decompressing said compressed data set. The invention also relates to a compressed data set and to a device for performing this compression method and a device for performing said decompression method.
Data compression is required when one needs to transmit or store data that would require a larger bitrate or size than the available bandwidth of the communication channel or the capacity of the storage medium. This is possible when the data contain a significant amount of redundancy, and some amount of details that could be removed without compromising the purpose of the transmission. The compression is said to be lossless when the data are not modified by the compression—decompression cycle, and thus are identical at the emitter and the receiver. However, it is usually not possible to guarantee a lossless compression since it depends on the intrinsic characteristics of the data, in which the redundancy and the ability to remove it by a clever encoding alone may not always succeed in reducing the bitrate or data cost enough. In such case, a lossy compression scheme is required, and the encoder reduces the information contained in the data by quantization, in order to guarantee a required compressed data set size at the expense of quality.
Document SMPTE Registered Disclosure Document SMPTE RDD 35:2016, 24 Mar. 2016, pages 1-53, XP055366991, describes the TICO video compression scheme bitstream, the decoding process and the provisions for mapping bitstreams onto an IP network. In this process, residues are obtained and coded using unary encoding.
Document WO03/092286 describes an adaptive method and system for mapping parameter values to codeword indexes. This method requires the step of sorting the differences or parameters into at least a first group and a second group. This requires additional buffering and processing at the encoder and at the decoder. This also adds latency to the process.
Document EP2773122A1, also related to the TICO video compression scheme, describes a method and device for lossless data compression wherein one groups said data into groups of n words of m bits; one detects for each group the value of the Greatest Coded Line Index (GCLI), the GCLI being the index of the highest weight non-zero bit among the bits of the words in a group; one produces for each group an output data set comprising the GCLI bits of lowest weight of the words of the group, and meta-data comprising the value of the GCLI. This method is very simple in implementation, and is very efficient, especially in the case when many of the data have small values. The values of the GCLIs are comprised between zero, when all words of a group are zero, and m, when at least one word in the group has a bit of weight m equal to 1. Therefore, binary coding of the GCLIs requires ┌Log2(m)┐ bits, ┌x┐, being the smallest integer greater than x. This document also proposes an improvement wherein one replaces the GCLI in the output data set by the difference between the GCLI value and a predicted value of the GCLI. In doing so, the values to be coded are smaller, and a unary coding may require less space than binary coding of the GCLIs. However, it appeared that even when using said improvement, the coding of the GCLIs could represent a significant volume in the compressed data set, especially when many of the coefficients become zero. In addition, this document does not address the need of quantization, when the compressed data set size exceeds a data budget, and lossy compression must be performed. Therefore, there is a need for a method for compressing an input data set which is simple to implement, requires less space for coding the meta-data containing the GCLIs, and addresses the issue of quantization.
It is an object of the present invention to solve the above-mentioned problems. In particular, it is an object of the present invention to provide a compression and decompression method and device having a low complexity, wherein the data budget required for the meta-data is reduced.
The invention is defined by the independent claims. The dependent claims define advantageous embodiments.
According to a first aspect of the invention, there is provided a method for compressing, according to compression parameters, one or more input data sets, the or each input data set comprising a sequence of M coefficients, each coefficient having m bits coding a magnitude comprised between 0 and 2m−1, into one or more corresponding compressed data sets each comprising a magnitude compressed data set, and a meta-data compressed data set, wherein said compression parameters may comprise M, m, n, t, the type of quantization, the mapping mode, being either negative-first or positive-first; the number of rows and columns of a display image, if the input data set represents a display image; a number of rows of a subband, if the sequence of pixels is a decorrelative transform of a display image; the prediction mode, which may be horizontal or vertical prediction; the way the initial values of the predictors are determined; the entropy coding mode; the value of the parameter k of the Rice coding if the entropy coding is a Rice coding and the bounding mode, comprising the steps of, for the or each input data set:
r
i=max(GCLIi−t, 0)−max(predi−t, 0);
r
min=−max(predi−t, 0);
(2) computing the rmax, the largest value of the residues for all possible values of GCLIi, being
r
max=max(m−max(predi,t), 0);
Preferably, said entropy encoding is a Rice coding, with k=0, 1 or 2. More preferably, k=0.
Preferably n is smaller than or equal to 8 or is equal to 4.
Said quantization may advantageously be performed by removing the t lowest bit planes of the groups of coefficients.
Said input data set may be obtained by performing a decorrelative transform on a non-decorrelated input data set.
Said sequence of M coefficients may correspond to a sequence of pixels of one or more rows of a display image comprising rows and columns of pixels or one or more rows of a subband of a decorrelative transform of a display image comprising rows and columns of pixels.
According to an embodiment of the invention with horizontal prediction, in step B.e), said predictor predi of GCLIi is the GCLI of the previous groups of coefficients in a sequence of coefficients, being pred_init for the first group of coefficients in a sequence of coefficients.
According to an embodiment of the invention with vertical prediction, in step B.e), said predictor predi of GCLIi is the GCLI of the group of pixels in same column of previous row of pixels if said GCLI is larger than t, and zero if said GCLI is smaller than or equal to t, being pred_init for the groups of pixels of the first row of pixels.
In a preferred embodiment of the compression method of the invention with vertical prediction, said one or more input data sets comprising at least two input data sets, a first input data set having a quantization level ti, a second input data set having a quantization level t2, the last row of pixels of the first input data set being above the first row of pixels of the second input data set, in a display image, the predictor for a group of pixels of the first row of pixels of the second input data set is equal to the GCLI of the group of pixels of the last row of pixels in the same column of the first input data set if said GCLI>t1 and equal to zero if said GCLI≤t1.
According to a second aspect of the invention, there is provided a method for decompressing, according to compression parameters, one or more compressed data sets each comprising a meta-data compressed data set comprising a sequence of entropy encoded codes ci and a magnitude compressed data set comprising bit planes of coefficients, obtainable by a method of the first aspect of the invention, with same compression parameters, into one or more corresponding decompressed data sets each comprising a sequence of coefficients, each coefficient having m bits coding a magnitude comprising the step of:
pred′i=previ;
n
bp
=r
i+pred′i;
In step d), the residue r may be performed by actually performing steps d(1) and d(2) or by other equivalent methods such as a searching algorithm or a formula.
When compression has been performed according to said preferred embodiment of the invention with vertical prediction, in the method for decompressing these two or more input data sets, for the first row of the second data set, one may advantageously take as predictor pred′i, the corresponding value of previ obtained for the last row of the first data set.
In the compression method and int the decompression method of the invention, pred_init may be equal to zero or equal to int(m/2).
According to a third aspect of the invention, there is provided a compressed data set corresponding to an uncompressed data set, said uncompressed data set comprising a sequence of M coefficients, each coefficient having m bits coding a magnitude, said compressed data set being obtainable from said uncompressed data set by the compression method of the invention, comprising
According to a fourth aspect of the invention, there is provided a device for compressing an input data set comprising a sequence of M coefficients, each coefficient having m bits coding a magnitude into an compressed data set comprising a magnitude compressed data set, a meta-data compressed data set, comprising at least one of a logic circuit, an ASIC, an FPGA, a GPU and a CPU, configured for performing the steps of the compression method of the invention.
According to a fifth aspect of the invention, there is provided a device for decompressing a compressed data set having data representing compression parameters, said compression parameters comprising the values of M, m, n, t, the type of quantization, the mapping mode, being either negative-first of positive-first, the type of entropy encoding and the prediction mode, and comprising a meta-data compressed data set comprising a sequence of Rice-coded codes and a magnitude compressed data set comprising bit planes of coefficients, into a decompressed data set comprising a sequence of coefficients, each coefficient having m bits coding a magnitude, said compressed data set being obtainable by the compression method of the invention, into an decompressed data set, said decompressed data set comprising a sequence of M coefficients, each coefficient having m bits coding a magnitude, comprising at least one of a logic circuit, an ASIC, an FPGA, a GPU and a CPU, configured for performing the steps of the decompression method of the invention.
The drawings of the figures are neither drawn to scale nor proportioned.
Table 1 below is an example of a sequence of M=12 coefficients, each coefficient having m=15 magnitude bits, grouped in three groups of n=4 coefficients. For each group, the GCLI is represented, the GCLI being the Greatest Coded Line Index, i.e. the index of the highest weight non-zero bit among the bits of the magnitude of the coefficients in said group. A simple quantization, is represented, where the t=4 lowest bit planes of the coefficients are truncated, i.e. the quantization level t is 4.
The GCLIs of the three successive groups are 8, 7 and 3. If one uses a prediction method and replaces the GCLIs by a residue, given by equation 1
r
i=GCLI
i−predi (equ.1)
, one obtains smaller values to be coded. In addition, when a quantization is performed, the residue ri needed by the decompression device for decoding the data may be computed as
r
i=max(GCLIi−t, 0)−max(predi−t, 0) (equ.2)
The residue ri according to equation 2 provides the necessary information to the decoder, while keeping the values to be coded as small as possible. Assuming that the predictor is taken as the previous GCLI in the sequence, (horizontal prediction) and that the GCLI of the group preceding the first one of the example is 7, the residues ri are as follows:
As can be seen in this example, the residues ri are smaller values. The number of bit planes to be transmitted, nbp i.e. the number of bit planes above the dashed line in table I is equal to GCLI−t when GCLI>t and equal to zero if GCLI≤t.
Table III is an example showing the treatment of a sequence of coefficients grouped in seven groups producing seven GCLIs. The quantization level t is 6. At the encoder, an initial predictor for the first group is taken as pred_init=0. The subsequent predictors are taken as the GCLI of the previous group in the sequence GCLIi−1. This prediction mode applies in the general case, and when the input data set represents an image and horizontal prediction mode is used. The residues ri are computed according to equation 2. The nbp bit planes are copied from the quantized coefficients to the magnitude compressed data set. The sequence of residues and bit planes are provided to the decoder.
At the decoder, the initial predictor for the first group is taken in the same way as at the encoder (in the example, predi_init=0). The number of bit planes nbp is computed using equation 3.
n
bp
=r
i+pred′i (equ. 3)
The predictor for the subsequent group is taken as the nbp of the current group, which corresponds to the GCLI of the quantized coeffcients. The GCLI′ of the decompressed data set coefficients may be obtained by adding t to the nbp values when they are not null, and keeping a zero value when they are null. As one can see these GCLI′s are equal to the GCLIs if the input data set, if the GCLI of the input data set is larger than t, and are equal to zero if the GCLI≤t, since in that case all the coefficients bitplanes have been removed, yielding null coefficients at the decoder.
When the input data set corresponds to a sequence of pixels of one or more rows of a display image comprising rows and columns of pixels or one or more rows of a subband of a decorrelative transform of a display image comprising rows and columns of pixels, vertical prediction is possible. It has been observed that vertical prediction gives good results.
Table IV shows an example similar to the example of table III, but applying vertical prediction. Row I of GCLIs is processed, taking into account previous row 1-1. The predictor predi is determined as follows:
IF GCLIi(I−1)≤t THEN predi=0 ELSE predi=GCLIi(I−1)
The residues are computed according to equation 2.
At the decoder, the set of GCLI′s corresponding to the quantized coefficients of the previous row (row I−1) are kept in a buffer for use in processing row I. The predictor pred′i is computed with
pred′i=GCLI′i(I−1) (equ. 4)
and the number of bit planes is computed according to equation 3. The GCLI′s for the current row are equal to the number of bit planes nbp,i and stored for use by the subsequent row. For the first row of a data set, a row initialized to values pred_init is used.
When the input data represents images and comprises two or more input data sets, and when the last row of a first data set is immediately above the first row of a second data set, these two data sets may be processed independently, with the initializations of the predictors for the first rows as discussed in the previous paragraph. However, in a preferred embodiment of the invention, the last row of the first data set may be used as predictor for the first row of the second data set. A special situation occurs when the quantization level t1 of the first data set is different from the quantization level t2 of the second data set.
The differences with respect to the case of Table IV is that at the encoder, the predictor is determined as follows:
IF GCLIi(I−1)≤t1 THEN predi=0 ELSE predi=GCLIi=(I−1) (equ. 5)
where t1 is used instead of t. At the decoder, the predictor is taken from the prey; element of the last row of first data set.
The general expressions of the lower bounds rmin and upper bounds rmax of the residues are given by equations
r
min=−max(predi−t, 0) (equ.6)
and
r
max=max(m−max(predi,t), 0) (equ.7)
The range of values of the residues after quantization, is reduced with respect to the range of values without quantization.
According to the invention, the values of the residues may be mapped to non-negative integers C. This is illustrated on
According to a preferred embodiment of the invention, the mapping of the residues to a non-negative integer C may be improved by taking into account that when rmax is larger than |rmin| the range of values of C may be reduced by mapping the residues above rmin to successive values of C (and not stepping by 2). This is referred to as the “Bounded C” method and represented on
In the option “bounded by min”, the values of the codes C for the residues comprised between -trigger and rmax, if any, are smaller than without the “Bounded C” method.
In the option “bounded by max”, the values of the codes C for the residues comprised between -trigger and rmin, if any, are smaller than without the
“Bounded C” method.
In the option “bounded by min/max”, the values of the codes C for the residues comprised between -trigger and rmin, if any, and the residues comprised between trigger and rmax, if any, are smaller than without the “Bounded C” method.
The codes C corresponding to a range of residues, according to the negative-first or the positive-first mapping mode, may be obtained by performing the following steps of the mapping algorithm: We first define the two following parameters:
Cfirst=−1 if negative-first, +1 if positive-first
When using the “Bounded C” improvement of the invention, a bounding mode is selected, being one of
A trigger is determined as follows:
trigger=|rmin| if bounded by min,
trigger=MIN (|rmin|, rmax) if bounded by min/max,
trigger=rmax if bounded by max.
trigger=m if not bounded.
The code C can then be computed by performing the following steps:
The option “negative-first” is more interesting since it ensures shorter codes for all values below the predictor, which includes all the values that have been removed by quantization. This is shown in
The bounding mode “bounded by min/max” does not offer a significant advantage over the bounding mode “bounded by min” bonding method, and thus the method may safely be simplified should its implementation offer a gain by not taking rmaxinto account.
Also, the mapping may be obtained from predetermined lookup tables.
The inverse mapping is bijective and thus its inverse may be performed at de the decoder.
Table VI below provides an example of eight successive GCLI′s from 1 to 8, for groups of coefficients, coded in m=15 bits. The quantization level is t=4. The initial predictor, for i=1, has been selected as int(m/2), i.e. 7. For each subsequent GCLI, the predictor has been determined as being the previous GCLI. For each successive GCLIs, the residues ri have been computed according to equation 2, and the rmin and rmax, have been determined according to equations 6 and 7 respectively. The values of the triggers, according to the different options, have been computed for each GCLI. In the last line, the code
C has been in the option “bounded by min”, and negative first. Each GCLI is processed independently of the other GCLIs in the input data set, in dependence of the predictor predi of the current GCLI.
Table VII gives the number of bits required for coding the 8 codes C according to the different bounding modes and mapping modes.
47
The bold value corresponds to the example shown in table VI. It can be seen that negative first give better results. The “bounded by min/max” bounding mode gives the better results, but the “bounded by min” are also good.
The information that is prepared at the encoder,i.e. the bitplanes and the codes C coding the residues are such that the decoder can from the received codes C, and a reconstructed predictor, determine the number of bit planes nbp to be extracted from the magnitude compressed data set for reconstruction the original quantized data set. The trigger used at the encoder for coding the residue is not needed at the decoder and not transmitted in the meta-data compressed data set.
The decoder receives a value of C and has a value of pred, from the previous step or from pred_init. From these two values, the decoder can compute rmin and rmax, according to equations 6 and 7 respectively. Having a knowledge of the mapping mode and the bounding mode, the decoder computes the trigger. The decoder then may compute the code C corresponding to all values of r between rmin and rmax, according to the above steps of the mapping algorithm. This produces a table giving the correspondence between the values of r between rmin and rmax, and corresponding values of C. The decoder then obtains the value of r as the value corresponding to C in this table.
In an example of table VIII, a code C=5 has been received, with a predictor equal to 7, and a quantization level t=4. The values of rmin and rmax can be computed and a table for all possible residues between rmin and rmax built. The code C for each of these 11 values of r is computed and inserted in the table. The value of the received code C=5 is then search in the table for finding the corresponding value of r, r=−3. Other equivalent methods for performing the decoding of the codes C, such as a (binary) search in the table, or a formula.
The codes C obtained for the residues in the method of the invention are non-negative integers. If the predictor is an accurate predictor, the residues will be small values. An entropy coding is used for coding C in the meta-data. A preferred entropy coding for C is the Rice coding. Rice coding of a non-negative integer N in dependence of parameter k is as follows:
one divides N by 2k and the resulting quotient is unary coded;
the remainder of the division is binary coded in k bits.
Example values for N between 0 and 10, and k=0, 1 and 2 are given in the table below:
With k=0, the length of the code, for N=0, is one bit. Therefore, the value k=0 is the preferred value for the method of the invention, when the prediction is accurate, and the value 0 is frequent among the residuals to be coded. However, if the accuracy of the prediction is not perfect, larger values of N might occur, and a larger value of k, such a as k=1 or 2, where larger values of N require less bits, may be optimal. Rice coding is a prefix code. Therefore, no code in the set of possible codes is a prefix of another code in the set of possible codes. No special markers are needed between the codes, and the decoder can unambiguously extract successive codes from the meta-data compressed data set.
The quantization step applies the set of 2m values of the magnitudes of the coefficients, between 0 and 2m−1, to a set of 2m-t values between 0 and 2(m-t)−1. A simple way to perform this quantization is by removing the t lowest-weight bits of the unquantized coefficients. However, other quantization methods may be used in the invention.
These results show that using both |rmin| and rmax bounds provides a very small improvement with respect to using only the |rmin| bound. In both cases using the negative-first mapping mode is significantly better than using the positive-first mapping mode.
The present description addresses the processing of the magnitudes of the coefficients. The methods apply to unsigned coefficients. The methods apply as well when the input data comprise signed coefficients. These signed coefficients may be coded as sign+magnitude or transformed to sign+magnitude format. The sign bits of the coefficients of a group are grouped as a sign bit plane, and the sign bit plane is processed along with the magnitude bit planes. The invention relates to a method for compressing an input data set, wherein the coefficients in the input data set are grouped in groups of coefficients, a number of bit planes, GCLI, needed for representing each group is determined, a quantization is applied, keeping a limited number of bit planes, a prediction mechanism is applied to the GCLIs for obtaining residues, and an entropy encoding of the residues is performed. The entropy-encoded residues, and the bit planes kept allow the decoder to reconstruct the quantized data, at a minimal cost in meta-data.
| Number | Date | Country | Kind |
|---|---|---|---|
| 17180619.3 | Jul 2017 | EP | regional |
This is a U.S. national stage application, which claims the benefit of International Application No. PCT/EP2018/068716, filed Jul. 10, 2018, which claims the benefit of European Application No. 17180619.3, filed Jul. 10, 2017, each of which is hereby incorporated by reference in its entirety.
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/EP2018/068716 | 7/10/2018 | WO | 00 |
