WO2001063934A1

WO2001063934A1 - Method for compressing data

Info

Publication number: WO2001063934A1
Application number: PCT/EP2001/001713
Authority: WO
Inventors: Joachim Buhmann; Wolfgang Raab; Hansjoerg Klock-Werneboldt; Ulrich Ramacher; Mathias Richter
Original assignee: Infineon Technologies Ag
Priority date: 2000-02-22
Filing date: 2001-02-15
Publication date: 2001-08-30
Also published as: DE10008055A1

Abstract

The invention relates to a method for compressing image data, comprising a differentiation between important and less important data. According to the invention, said method is characterised in that a calculation of a so-called pyramid of significance of wavelet-transformed data takes place in a separate step of the method. Said pyramid contains the wavelet coefficients, together with the corresponding maximum values of the trees of all the descendants or of all of the indirect descendants.

Description

Description:

Data compression method

The invention relates to a method for compressing data.

In various applications for the storage or transmission of digital images, compression of the image data is necessary in order to reduce storage space or transmission bandwidth. A distinction must be made between moving pictures (for example video) and single or still pictures. With the former, the fact that successive images differ only slightly in many regions enables particularly high compression factors to be achieved by merely storing or transmitting the image differences. In the case of single images, on the other hand, each image is to be viewed in isolation, a distinction being made between lossless and lossy methods of data reduction.

Lossless methods rely on the greatest possible elimination of redundancy by means of suitable transformations (for example discrete cosine transformation DCT or wavelet transformation) and coding (for example entropy coding). The compression factor can be further increased by - with more or less high quality losses - suppressing "unimportant" information (e.g. quantization of the data or omission of middle-value data bits). At the price of a higher computational effort, transformations that allow the exploit subsequent coding of higher order statistical relationships (these include the wavelet transformations), superior to the simpler transformations (such as DCT) with regard to the ratio of image quality to the remaining amount of data.

A large number of compression methods are known in the prior art. drive known. The known compression methods either allow high compression factors or a good image quality.

The invention has for its object to provide a compression method that combines high compression factors with a good image quality. The compression process should preferably be able to be carried out as quickly as possible.

According to the invention, this object is achieved in that the significance of data, in particular image data, is calculated in a separate method step.

From the international patent application WO 97/26603, a matrix arrangement of processor units is known, wherein the processor unit in addition to an arithmetic-logic unit and a result register bank, yet another arithmetic-logic unit, a multiplier / adder unit, a partial memory of an advantageous screen section memory and has a general purpose local storage. A processor containing the matrix arrangement of the processor units is characterized by a high processing speed

From the international patent application WO 99/42923, a device with a large number of control modules is known, the control modules being assigned a large number of processing modules for actuation. Here, the separate control modules are controlled by a higher-level controller and are synchronized by a common synchronization unit. This device reduces a transmission bandwidth between an external command memory and a highly integrated system.

The present invention serves to compress single images, whereby, by means of a coordinated combination of hardware and software, wise image quality and throughput, or time required for compression, better performance data can be achieved than with conventional implementations. Another aspect of the method is the subsequent - lossy - reduction of the amount of data already compressed images, for example in order to make space for further images on a storage medium that has been used up in its capacity.

The implemented method allows a very simple and quick subsequent data reduction.

A further improvement in the quality of the reconstructed data, in particular the image quality, can be achieved through a refined evaluation of the importance of individual data components. The latter can advantageously be achieved by coding in wavelet coefficients.

Acceleration of the compression method can advantageously be achieved by calculating significance parameters of the wavelet coefficients in a separate method step.

It is particularly expedient to carry out the method in such a way that the separate method step takes place in a separate processor element.

The invention can be used in a wide range of technology areas. In a variety of applications, there is a need to compress large amounts of data.

The invention is explained below using the example of image data. As a result, the advantages of the data compression according to the invention can be represented particularly clearly. However, the invention is not limited to the compression of image data. Rather, the invention is suitable for all areas of application in which particularly important information is obtained from a large amount of continuous data should. Possible further applications are therefore compression of audio data, speech recognition or evaluation of measurement results which are to be transmitted over data transmission channels with a bandwidth which is small in relation to the amount of data.

In order for a digital camera to have a resolution that is comparable to a conventional analog camera, in particular with regard to the resolution reserve for subsequent enlargements, more than one million image points (pixels) must be provided per single image. When storing data in an uncompressed state, this requires between 1 and 3 bytes per pixel. A data reduction is necessary in order to accommodate as many pictures as possible in the memory in the camera, for example a flash memory, and on the other hand to reduce the transfer time per picture, for example to an electronic "photo album" on the hard disk of the home PC to hold: take pictures during the day (requires storage capacity), at night - when the communication fees are low - remote transmission of the pictures home, for example with a mobile data transmission terminal, for example a cell phone.

Conventional analog cameras often offer the possibility of image series. In order to compete with this, a digital camera can only compress a picture in a fraction of a second (order of magnitude 5 pictures per second).

Various methods are known for data compression.

One of the methods used for single image compression is the "JPEG" standard with the sequence DCT, quantization (lossy) and Huffman coding. Because of the relatively low computational effort, JPEG, like other DCT-based methods, can be implemented well on standard processors become.

Wavelet-based implementations on standard processors are also known, but the real-time condition for image series is not met. For example, a 333 MHz Pentium processor using the method described below requires approximately 1/3 second per image with lossy compression by factors greater than 16 for a 512 x 512 image with 256 gray levels. For 1024 x 1024 images, it is approximately 1.5 seconds.

The invention is based on the object of providing a method for compressing data, in particular image data, the method being to be carried out with the least possible computational effort in the shortest possible time and with the highest possible reliability, in particular with the highest possible similarity to reproduction.

In order to achieve high compression factors and good image quality at the same time, a method based on wallets is used. The coding of the wavelet coefficients generates a so-called “embedded code”, that is to say the information is coded and transmitted or stored in the order from “important” to “unimportant”. The data stream can thus be transmitted at any time be canceled, or the last part of a stored compressed image are simply deleted, sure to suppress only ^'information that are unimportant or remaining as the already transferred, in memory.

It is particularly expedient to carry out the method on suitable processors. The term processor is not to be understood in any way restrictive. This can be any unit suitable for performing calculations. It is possible for the processor to consist of several individual processors, preferably one suitable processor unit are summarized. For the purposes of the present invention, any circuit suitable for performing calculations can also be used. The circuit is expediently installed in a computer or in a logic module.

One variant of the scalable processor architecture with a single processor element (PE) is sufficient, the processor preferably being a vision instruction processor (VIP). It allows the compression of several frames per second.

The analysis of wavelet-based compression methods shows that the computing effort is dominated by the wavelet transformation. This bottleneck is eliminated by the implementation on the VIP, since the architecture of the VIP is optimized for algorithms such as 1- or 2-dimensional convolution on which the wavelet transformations are based.

The embedded coding of the wavelet coefficients could be done by directly modeling the statistical dependencies between the wavelet coefficients. This requires arithmetic coding, which, however, requires scalar computing power, ie cannot be accelerated by parallelization. Therefore, according to the invention, coding methods that require arithmetic coding are used in particular.

The coding process is adapted so that the computationally intensive part can be parallelized and thus accelerated by the VIP.

A wavelet algorithm is preferably used. It belongs to the group of the "Zerotree method" and represents a variant of the method proposed by Said & Pearlman (Article Amir Said and William A. Pearlman: A new fast and efficient image codec based on set partitioning in hierarchical cal trees, IEEE Transactions on Circuits an Systems for Video Technlology, 6: 243-250, June 1996). New are the shift in the calculation of the so-called significance pyramid to a preprocessing step, which is particularly favorable for an implementation on the VIP, and an efficient scanning (re-sorting) of the wavelet pyramid during data transmission.

In a first preprocessing step, the image data is wavelet-transformed:

C = W (B).

Here B = {bij} are the image data and

the coefficients of an L-level wavelet pyramid. 1 shows a 3-level wavelet pyramid as an example. The assignment of father coefficients to son coefficients at lower levels is also shown.

In general, let S (l) be all coefficients of a stage 1. For the index (i, j) of a coefficient ci _j e8 (l) the following is defined:

• The set D (i, j) of the indices of all agreements at levels 0 <k <1 - 1.

• The set 0 (i, j) of the indices of all direct agreements at level k = 1 - 1.

• The set L (i, j) of the indices of all indirect agreements at levels 0 <k <1 - 2.

To achieve high compression rates, it is important to order the wavelet coefficients according to their size in order to obtain sequences of insignificant values that are as long as possible. The statistical dependency of the coefficients should be exploited: if a coefficient Ci _{j is} insignificant, it is probably also its agreements (with index in D (i, j)). "Insignificant" means: smaller than a barrier θ, which is gradually refined until the desired bit rate is reached. Insignificant coefficients are coded efficiently by means of so-called "zerotrees". C _x - means the root of one

• Type A zerotrees if c _ι and all agreements from D (i, j) are insignificant

• Type B zerotrees if c _XD and all agreements from L (i, j) are insignificant.

A second preprocessing step is carried out to determine zerotrees: the calculation of the so-called significance pyramid Z, which contains a value z _XD for each position (i, j). This consists of three parts:

• the wavelet coefficient c _1D

• the value a _x - _,: = max {| c _kx |; (k, 1) eD (i, j)}

• the value ß _x3 : = max {| c _kx |; (k, 1) eL (i, j)}

If auxiliary variables γ _x -, are introduced, the calculation of the values a _X j and ß _{Xj can be formulated} as a series of (non-linear) filtering over the different levels of the wavelet pyramid. In pseudo programming language notation:

for 1 = 1, ..., L for all (i, j) with c ₁₃ eS (l) if 1 = 1 a _x - ,: = max {| c _k |; (k, l) eO (i, j)} if 1> 1 γ _1D : = max {| c _k ι |; (k, l) eO (i, j)}

max {a _kx ; (k, l) eO (i, j)} a _X] : = max {γ _x - ,, ß _{x: ι} }

The creation of the significance pyramid, including the calculation of the wavelet coefficients, is by far the most computationally intensive part of the entire process. But you can implement very efficiently on the VIP, as detailed below.

The coefficients are treated according to their significance - only as much can be taken into account - as the set bit rate allows. The binary representations of the coefficients are considered and initially only examined for a most significant bit, that is to say those coefficients are determined whose magnitude is greater than a predetermined suitable threshold value θ ₀ : = θ _max . Then the threshold value is successively halved (θ _{t +} ι: = θ _t / 2) and the process is repeated (t = 0, l, ...). This corresponds to a step-by-step refinement of the quantization of the wavelet coefficients and is known as the bitplane technique. The scanning of the coefficients, which corresponds to a sorting, is important for the efficiency of the method. For each bitplane it is important to detect insignificant coefficients that are en bloc as possible (those that are smaller than the current threshold value θ _t ). For this purpose, the following lists of nodes are introduced, which are expediently modified as the method progresses (with increasing fineness of quantization).

LIP, the list of coefficients that are insignificant but not part of a zero tree,

• LZ, the list of coefficients that are the root of a zero tree and

• LSP, the list of coefficients that are (already) significant.

The following abbreviations simplify the notation:

• Z _t (ij): = true if and only if ß _x -, <θ _t , that is, if c _x: root of a zero tree. • A _t (Z _1D ): = true if and only if a _x -, <θ _t , that is, if c _x -, is the root of a type A zerotree.

• S _t (Z _X] ): = true if and only if | c _x -, | <θ _t , that is if Ci _{j is} significant.

• P (Zi _j ): = true then if c _X j> 0

• G _t (Zi _j ): = true if and only if c _Xj > c _n (0, where

• c ⁽⁾ = {C--}, the reconstructed pyramid into which the coefficients transferred up to time t have been inserted.

The entire coding process consists of preprocessing, initialization and a coding loop:

Preprocessing:

Calculation of the significance pyramid Z including the wavelet coefficients

Initialization:

c ⁽⁰⁾ : = 0, t: = 0, θo: = θmax

Initialize LSP, LZ, LIP as empty lists

insert (i, j) in LIP and in LZ as the root of one

Type A Zero Trees

Coding loop:

Continue until the set bit rate is reached Sorting phase (scanning the significance pyramid) Refinement phase t: = t + l, θ _t : = θ _t -ι / 2.

The refinement phase serves to refine the quantization of the transmitted wavelet coefficients:

Refinement phase: for all (i, j) in LSP if S-ι (z _Xj ) = true transmit G _t (z _Xj ) if G _t (z _Xj ) = true otherwise ω. (tl) θt y • Cy + y "~ y

The sorting phase is more complex:

Sorting phase: for everyone (i, j) in LIP

Transfer of S _t (z _XD ) if S _t (z _x -,) = true Insert (i, j) into LSP

Transfer sign (i, j) for all (i, j) in LZ if A _t -ι (z _{x: ι} ) = true Check whether (i, j) is the root of a type A zero tree

Check whether (i, j) is the root of a type B zero tree

with the sub-steps:

Transfer of the signs (i, j): Transfer of P (z _x -,) if P (z _x] ) = true

and:

Check whether (i, j) is the root of a type A zerotree: transfer A _t (z _x -) if A _t (z _XJ ) = false for all (k, l) eO (i, j) transfer of S _t (z _kx ) if S _t (z _k ι) = false

Insert (k, l) in LIP otherwise

Insert (k, l) in LSP transmit sign (k, l) if c _Xj belongs to the part of the wavelet pyramid with level 1> 1

Check that (k, l) is the root of a type B zero tree

Check whether (i, j) is the root of a type B zerotree: transfer Z _t (z _xj ) if Z _t (z _Xj ) = false for all (k, l) eO (i, j)

Insert (k, l) in LZ Remove (i, j) from LZ.

For the coding of n x n pictures, performance estimates give the following values for the number of arithmetic operations required:

• Wavelet pyramid: - 8 n2 (k _τ + k _H ), whereby half of these operations are attributed to multiplications, the other half to accumulations

• Rest of the significance pyramid: - 5 n ₂

Coding loop: 8n ² / c (without entropy coding of the transmitted coefficients)

A preferably biorthogonal wavelet transformation with separable conjugate mirror filter cores (Conjugate Mirror Filters) is expediently carried out here. Here, k _{τ is} the length of the core for low-pass filtering, k _{H is} the length of the core for high-pass filtering and C is the compression factor to be achieved (corresponding to the bit rate). Typical are cores of length 7 and 9, with which a computational effort of approximately 43n ² operations results for the parallelizable preprocessing. The ratio of parallelisable and sequential part in terms of computational effort is theoretically about 6 C to 1, if the Wavelet coefficients to be stored are not entropy encoded.

For the implementation on the VIP it is assumed that the image values are given with 256 gray levels (i.e. 8-bit

Data), the core elements, however, with 16 bit accuracy. However, since the calculation of the wavelet transformation after the first horizontal (or vertical) low-pass and high-pass filtering only continues "in the pyramid" and this has to be represented with 16-bit precision, an accuracy of 16-bit from the beginning accepted for all dates.

The processing pipeline of the VIP is outlined in FIG. Data sources of the processor element (PE) are the memories ISB (image section buffer), GPM (general purpose memory) and TMP (temporal data register), data sink is one of the RER (result register). From there, the data is written to an output buffer.

With the given data accuracy of 16 bits per cycle, the ALU 1 of the PE element can take 4 picture elements and 4 core elements from the ISB and GPM memories of the VIP, multiply them element by element and accumulate the results in a register. The calculation of 2 result pixels can be summarized, and it takes 4 cycles for cores of length 7 to be produced. For cores with a length of 9, this takes 8 cycles, on average 6 cycles. 6 cycles are to be used for the final accumulation of the segments of the result register and 4 cycles for the removal of the result pixels. The core elements in the GPM can be reused, the ISB allows the PE element to be supplied with image data precisely on time. Accordingly, the average is 16n ² (2x2) Bars to complete a one-dimensional (1D) fold. For the bandpass and lowpass filtering device, a total of 8n ² clocks are required. If both folds in the other (horizontal or vertical) direction are also taken into account, the result is 16n ² bars.

Continuing the process across any number of levels of resolution of the pyramid brings an additional factor 4/3 into play, so that a total of ² bars can be used to create the Wavelet pyramid 22n. With a clock rate of 200 MHz and for n = 1024 (photo quality), this results in a time requirement of approximately 115 ms for the calculation of the wavelet transformation.

In addition there is the calculation of the values a _X j and ßij of the significance pyramid according to the scheme outlined above. The intermediate _values γ _Xj are temporarily stored in a register. The following are used to calculate a value ßi _j : 4 cycles for supplying the 4 values a _kx from one of the two ISB memories (in the last of these cycles, the ISB is moved to the next position by means of a double shift, so that the data supply to the PEs is ensured) and 10 cycles for calculating the value ßi _j (2 "mima" commands from the ALU 2 with 2 waiting cycles). The calculation of a value γ _Xj takes just as long, since the calculation of value values can be done en passant in the ALU 1 and does not take up any additional computing time (the second ISB is used to read in the data). Then in a further 4 cycles (with an ALU-2 command) ai _j must be calculated as the maximum of ß _Xj and γi _j . For rounding and removing ai _j and ß _j , there are 4 cycles each. In total, there are 40 cycles for calculation per tuple (ai _j , ßi _j ). In the first step of the algorithm (1 = 1) only the values aij are calculated, that is 18 cycles each.

Hence wall of 3-- ¹ -18n2 clocks expected an up for the stage-1 = 1, the significance of the pyramid, for the stage-1 = 2 a 16 Effort of 3-— 1 • - 1--40n 2 bars, for 1 = 3 an effort of 4 16

3- - n ² ■ -— ι -40n cycles etc. In total this is less than 6n V47 16

Clocks, with a PE clocked at 200 MHz around 32 ms for n = 1024.

Overall, in the example there is a time requirement of around 150 ms for the creation of the significance pyramid including the calculation of the wavelet coefficients on the VIP, with 512 x 512 images it is correspondingly less than 50 ms.

However, no overhead is taken into account here, such as calculation times for the parameters of the controller programs, initialization times when starting up the VIP controller or waiting times for synchronization of the controllers.

Table 1: Computing times for the significance pyramid

Table 2: Computing times for coding loop, 333 MHz processor

Table 1 compares the values with those that can be expected when calculating a 6-level significance pyramid (including wavelet coefficients) for a Pentium processor with 333 MHz with optimal implementation. The coding loop has to be processed sequentially and cannot be implemented quickly on the VIP. The non-optimal implementation values measured on the Pentium processor gave the values listed in Table 2. These are the values for a coding loop without entropy coding the transmitted symbols.

In order to document that the coding method is expedient, the PSNR error measures for the reconstructed images at different bit rates for the standard image "Lenna" in the format 512 x 512 are finally given in table 3. Again, the values without Entropy coding With entropy coding an improvement of about 0.5 dB is to be expected in each case.

Table 3: Errors when compressing Lenna 512 x 512

Claims

claims

1. Method for compressing data, whereby a distinction is made between important and less important data, so that the significance of data is calculated in a separate method step.

2. The method of claim 1, d a d u r c h g e k e n n z e i c h n e t that the importance of the data is calculated via a coding in wavelet coefficients.

3. The method as claimed in claim 2, so that the significance parameters of the wavelet coefficients are calculated in a separate method step.

4. The method according to claim 3, so that the separate method step takes place in a separate processor element.