WO2015075185A2

WO2015075185A2 - Method and device for digital image compression

Info

Publication number: WO2015075185A2
Application number: PCT/EP2014/075293
Authority: WO
Inventors: Carole THIEBAUT; Roberto CAMARERO; Christophe Latry
Original assignee: Centre National D'etudes Spatiales
Priority date: 2013-11-21
Filing date: 2014-11-21
Publication date: 2015-05-28
Also published as: WO2015075185A3; FR3013490B1; FR3013490A1

Abstract

The invention relates to a method and device for digital image compression, comprising the application (6) of a transformation in order to obtain a representation of the image comprising a set of coefficient blocks, each coefficient block comprising a plurality of coefficients, termed frequency coefficients, followed by the application of quantification and coding (10) of the coefficient blocks, in order to obtain a set of digital data representative of the compressed digital image. The method according to the invention comprises, after obtaining the coefficient blocks, for at least one processed block selected from among the coefficient blocks, calculation (8) of a first value representative of said block processed using the frequency coefficients of said process block, and the obtaining (8) of a second value representative of said process block, calculated as a function of a model of noise associated with said processed block. The quantification and coding (10) of the processed block takes account of said first and second values associated with the processed block.

Description

Method and device for digital image compression

The present invention relates to a digital image compression method represented by at least one two-dimensional array of image samples. It also relates to an associated compression device.

The invention lies in the field of digital image compression, and more particularly in the field of image compression with a quality objective to achieve.

Digital images obtained by various sensors are generally represented by arrays of image samples or pixels, each image sample having a given dynamic intensity.

Nowadays, digital images have very large sizes, whether they are images in the field of photography using high quality devices, and in the field of remote sensing imaging, or in the field of medical imaging.

In order to store and transmit such images, it is known to apply digital image compression.

Various compression techniques exist.

In general, the compression methods include the application of three treatments to a source digital image, which are: decorrelation, quantization, coding. In several digital image compression standards, the decorrelation is obtained by applying a transformation, applied to the pixels to obtain transformed coefficients.

The JPEG ("Joint Photography Expert Group") coding standard uses a discrete cosine transform (DCT) applied to 8x8-sized image blocks, followed by block quantization and entropy coding. Quantization consists in applying a quantization matrix to the coefficients of a block, and it is possible to define a quality level of compression, a multiplicative factor then being applied to the quantization of all the blocks of the source digital image. However, such a level of quality results in a more or less strong compression according to the blocks of the source digital image, and does not allow to finely control the quality of the image to be encoded.

The JPEG2000 encoding standard was developed after JPEG. It uses a discrete wavelet transform (DWT) of the source digital image. Quantization and coding are performed in a single step, thanks to quality layer coding, obtained by bitmap coding of the transformed coefficients. In order to limit the amount of information data relating to coding parameters, the JPEG2000 standard uses coding units which are called blocks ", typically of size at least equal to 32x32 pixels. Thus, JPEG2000 does not allow a fine control of the coding quality.

In the field of image processing obtained by satellite, the Image Data Compression (IDC) coding standard defined by the Consultative Committee for Space Data Systems (CCSDS) also uses a discrete wavelet transformation, followed by a grouping. spatial transformed coefficients obtained in blocks of predetermined size, typically 8x8 size, and a segment coding grouping a predetermined number of consecutive blocks. Each block of transformed coefficients comprises a so-called low frequency coefficient, representative of the average of the intensity of the pixels of the block, and a plurality of so-called frequency coefficients, corresponding to three levels of wavelet decomposition for a block of size 8 × 8. The segments are encoded independently of each other, the parameters necessary for the compression being encoded in the header of each segment.

The frequency coefficients of the blocks of a segment are coded by bit planes. As in the case of JPEG2000, the set of binary coding data associated with a segment, also referred to as the bit stream associated with the segment, is organized hierarchically and can be truncated at any point, to obtain a smaller but quality data volume. degraded.

The CCSDS IDC coding standard provides two modes, allowing quality and / or data volume control for each segment. The quality is controlled in particular by the number of coded bit plans for the segment. As a result, the actual quality of the reconstructed image depends on the characteristics of the source digital image. Therefore, in practice, it is particularly difficult to find parameters to ensure a desired quality for a set of images. Indeed, with constant parameters, one can obtain strong differences in terms of signal-to-noise ratio between two different digital source images.

There is therefore a need to improve the existing compression methods by allowing fine control of the local quality level, while guaranteeing a minimum flow rate for a given target local quality level.

For this purpose, the invention proposes, according to a first aspect, a digital image compression method represented by at least a two-dimensional matrix of image samples, each sample having an associated intensity value, comprising steps of :

applying a transformation to obtain a representation of the image comprising a set of coefficients blocks, each coefficient block comprising a coefficient called low frequency coefficient corresponding to an average value of the intensity of the image samples of said block and a plurality of coefficients called frequency coefficients;

quantization and coding of the coefficient blocks, in order to obtain a set of digital data representative of the compressed digital image.

This compression method comprises, after obtaining blocks of coefficients, the steps of:

for at least one treated block selected from the blocks of coefficients:

Calculating a first representative value of said processed block from the frequency coefficients of said processed block,

Obtaining a second value representative of said processed block calculated according to an image acquisition noise model associated with said processed block,

Determining a quality target according to said first and second values associated with said processed block,

Quantifying and encoding said processed block according to said determined quality target.

Advantageously, the compression method of the invention makes it possible to apply a differentiated coding for treated blocks as a function of a first value calculated from the frequency coefficients of a processed block and a second value representative of a model. noise associated with the processed block.

Advantageously, a fine control of the quality is obtained, thanks to the selection of a quality target per block.

The compression method according to the invention may have one or more of the features below, taken alone or in combination, in any technically possible combination.

According to one characteristic, for a processed block, said quality target is a value representative of the mean squared error between the coefficients of the block processed before quantization and the coefficients of the block processed after quantization.

The method further comprises, for each block processed, a step of determining a quantization level to meet the quality target.

According to one embodiment, the method comprises, for each block treated, a comparison of said first value with a first threshold level calculated from said second value, and when said first value is lower than said first threshold level, a step of setting zero of the frequency coefficients of said processed block.

When said first value is greater than said first threshold level and less than a second threshold level, said processed block is considered to be a block exception for which said quality target is selected to minimize the mean squared error between the coefficients of the processed block before quantization and the coefficients of the processed block after quantization.

According to one characteristic, the first value representative of a processed block is representative of the variance of the frequency coefficients of the processed block, and the second value representative of a processed block is a function of a variance of a noise model depending on the value of the low frequency coefficient of the treated block.

The processed blocks are selected by applying a bit mask.

The method comprises, for each block not selected as a block processed by applying said mask, a step of setting the frequency coefficients of said block to zero.

The method comprises a step of encoding an indication relating to a quantization level to be applied for each block processed, a said coding indication being associated with the set of digital data representative of the compressed digital image.

According to a variant, the quantization and coding step is implemented by a coding by bit planes representative of the coefficients, and said indication is an indication relating to the number of bit planes to be coded.

According to one variant, the quantization step is implemented by applying a block quantization matrix, and said indication is an indication relating to a weighting coefficient to be applied to a predetermined quantization matrix.

The coefficient blocks are obtained by applying a discrete wavelet transformation and spatial grouping of the transformation coefficients obtained.

The coefficient blocks are obtained by applying a discrete cosine transform applied on blocks of predetermined size.

According to a second aspect, the invention relates to a digital image compression device represented by at least a two-dimensional array of image samples, each sample having an associated intensity value, comprising modules capable of producing:

an application of a transformation in order to obtain a representation of the image comprising a set of coefficients blocks, each coefficient block comprising a coefficient called a low frequency coefficient corresponding to an average value of the intensity of the image samples said block and a plurality of coefficients called frequency coefficients;

a quantization and a coding of the blocks of coefficients, in order to obtain a set of digital data representative of the compressed digital image. This device comprises, modules, able to achieve, after obtaining blocks of coefficients, for at least one treated block selected from the blocks of coefficients:

A calculation of a first representative value of said processed block from the frequency coefficients of said processed block,

Obtaining a second value representative of said processed block calculated according to a noise model associated with said processed block,

A determination of a quality target according to said first and second values associated with said processed block,

A quantization and a coding of said processed block according to said target of determined quality.

The compression device comprises modules able to implement the steps of the digital image compression method briefly described above.

This compression device provides similar advantages to the advantages of the compression method briefly described above.

Other features and advantages of the invention will emerge from the description given below, by way of indication and in no way limiting, with reference to the appended figures, among which:

FIG 1 is a block diagram of an image compression system implementing a compression method according to the invention;

FIG. 2 diagrammatically represents a set of sub-bands obtained by applying a discrete wavelet transformation to a digital image;

FIG. 3 schematically represents a block of transformed coefficients obtained by a wavelet transformation;

FIG. 4 is a diagram representing the functional blocks of a programmable device capable of implementing the invention;

FIG. 5 is a block diagram of a step of determining a block quality target according to one embodiment of the invention;

FIG. 6 is a table illustrating the association of quality targets and a value representative of the block frequency coefficients to be encoded;

FIG. 7 schematically illustrates a set of bit planes of a block;

FIG. 8 is a block diagram of a quantification and coding step using a quality target according to a first embodiment of the invention;

FIG. 9 is a block diagram of a compression method implementing the determination of a quality target according to a second embodiment of the invention. Figure 1 illustrates a system 1 of compression / decompression of images. The system is composed of a first part 2 performing digital image compression, a second storage / transmission part 3 and a third decompression part 4.

Parts 2 and 4 are respectively implemented by devices as briefly described below with reference to Figure 4, comprising one or more processing units, for example processors, and able to perform calculations.

The second storage and transmission part 3 is performed by storing data in a memory of a programmable device or an electronic device and / or by a communications network capable of transmitting digital data.

In a variant, the methods of the invention are implemented by electronic devices of the programmable logic circuit type, such as electronic cards based on FPGAs or ASICs.

A digital image to compress I is presented at the input of the first compression part 2. The digital image I is composed of one or more arrays of image samples or pixels, each image sample having an associated intensity value.

Subsequently, it is considered to simplify the explanation that a digital image to be compressed is composed of a matrix of dimensions M × N of pixels, each pixel having an intensity initially coded on P bits. For example, P = 8.

The digital image I is supplied to a first module 6 for transforming the digital image to be compressed into blocks of transformed coefficients of predetermined size. The coefficient blocks are arranged spatially.

According to one embodiment, a discrete wavelet transformation (DWT) is applied over the entire image matrix, on 3 levels of decomposition, as schematically illustrated in FIG. 2.

In a known manner, the application of a DWT is first carried out on the complete image matrix, making it possible to obtain four sub-bands, each subband including transformed coefficients corresponding to the application of a low-pass filtering L and a high-pass filtering H, firstly in a first direction of the matrix, then in a second direction, perpendicular to the first direction.

The application of a first level of decomposition to the matrix I makes it possible to obtain the sub-bands of coefficients LL _1; LH _1; HLi and HHi. The subband comprises a low-pass version of the initial image I at a lower resolution level than the resolution level of the initial image I. If the initial image is composed of M rows and N columns, each LL ₁ sub-band _; LH _1; HL _1; HHi's first level of decomposition has M / 2 rows and N / 2 columns.

The application of the respective filters L and H is iterated on the sub-band LL _1; to obtain the sub-bands LL ₂ , LH ₂ , HL ₂ and HH ₂ , second decomposition level and size M / 4 rows and N / 4 columns.

The LL ₂ sub-band is filtered again to obtain the third decomposition sub-bands LL ₃ , LH ₃ , HL ₃ and HH ₃ .

Thus, after applying a wavelet transformation on three decomposition levels, as illustrated in FIG. 2, the digital image is represented by a set of transformed coefficient subbands. At a decomposition level p, the size of the subbands is M 12 ^P x N 12 ^P.

The transformed coefficients of the subbands are grouped in blocks of spatially located coefficients of size 2 ^P x 2 ^p for p decomposition levels, thus blocks of size 8x8 for a decomposition with 3 levels of decomposition.

Thus, as illustrated in FIGS. 2 and 3, a block B _k of transformed coefficients is obtained by spatially grouping coefficients of the DWT transformation corresponding to the same spatial position of the initial image I.

Each block of coefficients B _k comprises a low-frequency coefficient DC, coming from the sub-band LL ₃ in the example, and several AC frequency coefficients derived from the sub-bands LH ₃ , HL ₃ , HH ₃ , LH ₂ , HL ₂ , HH ₂ , LH ^ HL ^ HHi as illustrated.

Alternatively, according to an alternative embodiment, the image I is divided into blocks of pixels of size 8 × 8 and a discrete cosine transformation (DCT) is applied to each block of pixels, making it possible to obtain blocks of 8 × 8 localized size. spatially transformed coefficients.

In general, any decorrelation transformation, making it possible to obtain blocks of coefficients of size mxn, is applicable.

Returning to FIG. 1, following the transformation, a representation of the block image B _k of transformed coefficients C _t _i including a low frequency coefficient C _0.0 , also called DC coefficient, and frequency coefficients C _{t i} is obtained.

The coefficient blocks are then processed by a block quality determination module 8 per block of the image.

This block quality target determination module 8 is specific to the invention. It takes as input Pi parameters relating to a noise model associated with the processed image and P ₂ parameters relating to the quality target.

Optionally, one or more MK bit masks to distinguish between blocks of interest and non-interest blocks is also provided. For example, if the image digital I to compress is a remote sensing image, acquired by a sensor on board a satellite, a mask distinguishing blocks belonging to cloud areas is provided, the cloud blocks being blocks of no interest to the mission assigned to the satellite because without information content.

In a more general manner, for a given target application, in a given use context, it is possible to provide such MK masks for defining the interest of each block so as to be able to adapt the associated target image quality.

Subsequently, only the blocks of interest are processed to establish an associated quality target.

The non-interest blocks are simply coded by their low frequency coefficient.

The determination of a quality target per block of interest, also called treated block, will be detailed hereinafter with reference to FIG.

In one embodiment of the invention, the module 8 for determining a quality target per processed block provides, for each processed block B _k , a mean squared error target value MSE _k (Mean Square Error) of coding. for block B _k .

Each coefficient Q _j of the block B _k is coded by a quantized coefficient d ^* .

The mean squared error of coding is given by:

MSE _k = - Σ (C ,, - C ^* ,) ² (Eq l)

(U) ≠ (0,0)

for a block B _k of size mxn.

These MSE values _{k as} well as the transformed coefficient blocks are provided at the input of a quantization / coding module.

For each processed block B _k , the quantization / coding module 10 determines statistically or by measurement of the mean squared error (MSE), an associated quantization level and performs the entropy coding by applying the determined quantization level.

According to a first embodiment, the module 10 performs quantization and coding by coding the coefficients of a block processed by bit planes.

According to a second embodiment, the quantization is performed by applying a quantization matrix per block, and the coding is an entropic coding of the quantized coefficients.

At the output of the module 10, a set of binary data or bitstream representative of the quantized coded coefficients 12 as well as a set of indications 14 relating to the quantization / coding parameters are obtained. In one embodiment, groups of blocks or picture segments are independently encoded. An information set 14 and a set of binary data 12 are associated with each segment to form a bit stream of compressed data associated with the segment. The set of these bit streams per segment forms a set of compressed digital data F representative of the image I.

As a variant, only one set of information 14 is associated with all the blocks of the image I.

The set of compressed digital data F is either stored, for example in the form of a file, or transmitted via a communications network, by a storage / transmission part 3 of the system 1.

The set of compressed digital data is received by the third decompression part 4.

A decoding and dequantization module 16 receives, for each segment of the image, the compressed digital data F comprising the binary data 12 and the associated information sets 14, and decodes transformed coefficients associated with the blocks B _k of the image I.

Quantized coefficients C for each of the blocks B _k of the image represented by the compressed digital data F are obtained at the output of the module 16.

Then an inverse transformation by block of coefficients is applied by the module 18.

If the transformation applied by the transformation module 6 into coefficient blocks is a DWT as explained above, the coefficients of the blocks B _k are redistributed by subbands, and an inverse DWT transformation is applied by the application module 18 inverse transformation, depending on the number of decomposition levels.

If the transformation applied by the transformation module 6 is a block DCT, a block inverse DCT is applied by the module 18 for applying an inverse transformation.

At the output of the module 18, a reconstructed image I ^* is obtained, the intensity values of the samples of the image I ^* being obtained by applying the inverse transformation as explained above.

The methods of the invention are implemented by an electronic device, for example a computer-type, workstation-type programmable device whose main functional blocks are illustrated in FIG. 4.

An electronic device 20 capable of implementing the methods of the invention comprises a central processing unit 22, for example a CPU, capable of executing preprogrammed operations or computer program instructions when the device 20 is powered on. In one embodiment, a multiprocessor CPU is used to perform parallel computations. The device 20 also comprises information storage means 24, for example registers, capable of storing executable code instructions allowing the implementation of programs comprising code instructions able to implement the methods according to the invention. .

The device 20 comprises control means 26 for updating parameters and receiving commands from an operator. When the programmable device 20 is an onboard device, the control means 26 comprise a telecommunication device for receiving remote commands and parameter values.

Alternatively and optionally, the control means 26 are means for inputting commands from an operator, for example a keyboard.

Optionally, the programmable device 20 comprises a screen 28 and additional pointing means 30, such as a mouse.

The various functional blocks of the device 20 described above are connected via a communication bus 32.

FIG. 5 illustrates an embodiment of a step of determination of quality target by transformed coefficient blocks of a digital image I, implemented by modules of the processor 22 of a device 20 capable of implementing the invention or by an electronic card developed for the implementation of the invention.

The determination of a quality target is implemented for each block of transformed coefficients B _k resulting from the image I.

A block B _k is provided at the input of an optional stage 34 for selecting blocks of interest. If a binary MK mask defining blocks of interest and blocks of no interest is provided, step 34 consists in applying the mask MK to determine if the block B _k is a block of interest. In the case of a negative response, step 34 is followed by a step 35 of determining at least one coding parameter for block B _k .

In step 35, the block B _k is modified into a block B ' _k to be coded in which all the AC frequency coefficients are set to zero. Advantageously, the coding rate associated with the non-interest blocks is thus minimized, a minimum information relating to these blocks nevertheless being coded by coding the low-frequency coefficient of the block. As a variant, a mean squared error MSE _k target is associated with the block B _k in step 35.

If the block B _k is a block of interest, step 34 is followed by a step 36 of calculating a first value representative of the frequency coefficients of the block B _k .

In the preferred embodiment, the variance of the frequency coefficients of the block is calculated:

VAB _k) = Σ ¹ _(C, -C) ²

(xn - l) _{i eBi}

Calculating the first value takes into account all the coefficients _C, of the B block _k except the low frequency coefficient C _0,0, C is the mean value of the frequency coefficients of the block B _k.

Alternatively, the standard deviation of the frequency coefficients of the block B _k or another value representative of the statistics of the coefficients, such as a higher order moment, is calculated.

In a step 38, a second value V ₂ , representative of an image acquisition noise model associated with the block B _k , is calculated.

In the preferred embodiment, a parameterized noise pattern is used.

The second value V ₂ is a function of this noise model and the value of the low frequency coefficient C _0.0 , representative of the average of the intensity of the pixels of the block B _k , according to the following formula:

V ₂ {B _k ) = a ² + - C _0fi

The second value V ₂ is the variance of the local noise, and the parameters a and β are provided in the set of input parameters of the process. These parameters a and β are predetermined constants.

For example, the constants a and β are representative of the acquisition sensor of the image I to be compressed and are determined by calibration of the sensor and stored. For the PLEIADES-1 A satellite, the values a and β are equal to 2.267 and 0.0393, respectively. Alternatively, for a non-white detection noise, the noise calculation may take into account different models for each sub-band or transform coefficient. The noise is therefore obtained with the following formula:

V ₂ (B _k ) = Σ 0.0

> J≡B _k

(U) ≠ (0,0)

It should be noted that the calculation steps 36, 38 can be performed in parallel or their order can be reversed. According to the invention, the quantization and the coding of the processed block are carried out as a function of the first value and the second value V ₂ calculated for the block.

In a step of determining the type of block 40, the first value \ (Βκ) is compared with the second calculated value, weighted by a coefficient K.

The coefficient K is a constant. This coefficient is a parameter provided by the user, and makes it possible to determine a threshold level for separating the blocks containing useful information from the noise blocks. In one embodiment, K is preferably chosen between 0.5 and 4, for example K = 1.

If the first value \ (Β _Κ ) is lower than the first threshold level KV ₂ (B _k ), it is concluded that the processed block B _k is a block containing noise. In this case, the step 40 is followed by a step 42, similar to the step 35 described above, in which the block B _k is modified into a block B ' _k to be coded in which all the AC frequency coefficients are set to zero. Indeed, it is useless to allocate an important coding rate to a block which comprises only noise.

If the first value \ (Βκ) is greater than the first threshold level KV ₂ (B _k ), it is concluded that the block B _k contains useful information.

Step 40 is followed by a step 44 of comparing the first value V ^ B _k ) with a second predetermined threshold level S _E.

If the value \ (Β _Κ ) is lower than this second threshold level, it is concluded that the signal level of the block B _k , estimated by the value \ (Β _Κ ) is low, slightly higher than a level noise type signal. For example, S _E is between 16 and 32.

These blocks are so-called exception blocks, having information while having a low signal level. This type of block is heavily compressed by conventional compression methods, even for low compression rates.

For some applications, especially in the field of remote sensing, it has been noted that it is important to preserve such blocks.

As a result, a particular treatment is applied to them.

In the preferred embodiment, the quality target for these blocks is set to a high level, which is achieved for example by setting to 0, in step 46, the mean square error MSE _k for the block B _k .

The noise model presented above, can be used, during the optional step 47, to perform a denoising by thresholding or weighting coefficients for this type of block, to ensure a more efficient coding of the useful information. In this case MSE _k = 0 must be guaranteed between the denoised block obtained at the output of denoising step 47 and the block after coding.

If the value \ (Β) is greater than the threshold S _E , the block B _k is considered as a regular block, for which a quality target is set in step 48.

In this step 48, a mean square error of MSE target coding _k is set for the block B _k .

Preferably, the value MSE _k is set according to the first value ((Β) equal to the variance of the frequency coefficients in this embodiment.

Target mean MSE target error values as a function of the variance or standard deviation values are tabulated and stored in order to be provided among the P ₂ input parameters of the process.

For example, these values are stored in tables, as for example the table shown in FIG. 6 linking block frequency coefficient variance values to target target MSE target values.

Thus, for example, using the table T of FIG. 6, if the block B _k has a first associated value strictly less than 64, the target mean squared error is MSE _k = 2.

As a variant, several tables relating the variance of the frequency coefficients to a quality target value are stored and used, for example, as a function of the membership of the block treated with a type of region. For that, a preliminary classification in regions must be carried out.

Once the quality target determination performed for a block B _k , it is stored in order to be supplied to the quantization and coding module 10, for a coding of the block respecting this quality target.

As a variant, the noise model presented above can also be used, during an optional step 49, to perform a denoising by thresholding or weighting of the coefficients of the block B _k , in order to guarantee a more efficient coding of the useful information. In this case, the target mean squared error value MSE _k associated with the block B _k must be guaranteed between the denoised block obtained after the denoising step 49 and the block after coding.

In the particular case where the AC frequency coefficients are set to 0, the modified block B ' _k is directly provided for quantization and coding, without indication of the particular quality target associated.

FIGS. 7 to 8 illustrate the implementation of the quality targets and coding parameters determined in a first embodiment, in which the blocks of coefficients are encoded by bit planes. This embodiment is in particular compatible with the CCSDS IDC compression standard.

Figure 7 schematically illustrates the principle of bitmap coding.

For a given block B _k , a 4x4 block being illustrated in FIG. 7, each coefficient C has an associated value represented on a number B of bits in binary representation, arranged in the order between the C bit, ^{6 "1} of most significant bit and the least significant bit C ^{0 0} , or LSB for "least significant bit".

Thus, it is possible to represent a block by the bit planes of its coefficients, a bit plane grouping together all the bits of the same weight of all the coefficients of a block.

The maximum number of bit planes needed to represent all the frequency coefficients of a block, that is, the number of bits needed to encode the largest frequency coefficient of a block of coefficients, is called depth or " bit-depth ".

In the example of FIG. 7, the block B _k is represented by B = 5 bit planes, denoted BP ₀ to BP ₄ , so its depth is equal to B = 5.

The bit plane coding consists of successively coding the representative bit planes of the coefficients of one or more consecutive blocks of an image, starting from the MSB plane, as indicated by the direction of the arrow in FIG. 7. This makes it possible to obtain a binary set of data or a hierarchically nested bit stream, the most important information being at the head of the bit stream.

Quantization is performed by stopping the coding at a number of bit planes for a block B _k .

For example, the coding of the BP ₄ and BP ₃ bit planes may be sufficient to obtain a target quadratic error value.

According to the CCSDS IDC compression standard, multi-block segments are encoded independently. For each segment, the coding indications are encoded in the segment header.

The low-frequency coefficient or DC coefficient of each block is coded by predictive coding, separately from the frequency coefficients or AC coefficients of the blocks of the segment. Preferably, a Rice encoding is applied to the differences between successive quantized DC coefficients of the segment coefficient blocks.

Then, the depths BitDepthAC_Block _m associated with the coding of the frequency coefficients of the successive blocks B _m of coefficients of the segment are coded by entropy coding, for example the coding of Rice. Then, for each block, the bit planes of all the successive AC frequency coefficients, starting from the most significant bit plan, are coded, by coding steps or "stages" in English, the coding steps allowing a hierarchy of the data encoded by bit plane. A maximum of 4 coding steps for each bit plan is provided by the CCSDS IDC standard.

Thus, a quality layer bit stream is obtained for a coefficient block segment.

According to the invention, the coding for each block of coefficients of a segment takes into account the quality target or the pre-quantization by setting to 0 the frequency coefficients implemented in the block quality target determination step. of coefficients described above.

FIG. 8 illustrates the main steps of a method of quantization and coding of the frequency coefficients taking into account the quality targets determined and stored in blocks, according to a first embodiment of the invention. For each block processed, a quantization level per block processed is determined according to the quality target previously calculated for the block.

For a block B _k of a given image segment to be compressed, a first step 50 of initialization of the coding parameters is implemented. During this initialization step 50, the depth B of the block B _k to be coded is calculated, an index b of the current bit plane BP _b to be coded is set to B-1, and an index s representative of the step of encoding of the bit plane b is set to 0.

Then, during a quantization step 52, the frequency coefficients of the block B _k , quantized according to the current parameters b and s, are calculated by the function denoted as (B _k , b, s).

The current quantization mean square error, denoted mse, is also calculated during the quantization step 52.

During a comparison step 54, the mean squared error mse of the calculated block is compared with the mean squared error MSE _k for the block B _k , previously determined and stored.

If the quantization mean quantization error mse, with the current encoding parameters, is smaller than the mean squared error MSE _k , step 54 is followed by a step of storing a coding indication of index b of bit plane to be coded and associated coding step s for the current block B _k .

The block coding indication is associated with the binary data encoded and transmitted with the compressed data for later use during decoding, as explained above. If the current quantization mean quantization error, with the current encoding parameters, is greater than the target mean squared error, step 54 is followed by a step 58 of comparing the index of the encoding step If the value of the index sa reaches the value 3, then step 58 is followed by a step 60 of updating the parameters. The bit plane index to be coded is decremented by 1 and the subscript s is again set to 0. Otherwise, the subscript s is incremented by 1 during step 62 of incrementing the encoding step.

Thus, four bit plan coding steps are allowed, these coding steps being applied according to the CCSDS IDC compression standard.

The respective steps 60 and 62 are followed by the quantization step 52 previously described.

Steps 50 to 62 are applied for each block B _k to be coded for each segment. For blocks for which the AC frequency coefficients have all been set to 0, only the low-frequency coefficient is quantized and coded, the quantization level being determined conventionally for the whole segment.

In one embodiment, a file containing the set of coding indications for the set of blocks of a segment is created for each segment encoded and stored / transmitted with the coded data of the segment, for example in the same manner. that the sequence BitDepthAC_Block _m .

As a variant, two simplifications can be introduced in the quantization / coding step intended to limit computational complexity.

According to a first variant, the calculation of the mean squared error mse is done once per bit plane, for the same step "s" of coding. In this way steps 58 and 62 can be deleted.

According to a second variant, requiring even less computational resources, the values of "b" and "s" are associated with target image quality levels depending on the variance of the block. These values are determined statistically for a given application or sensor.

Thus, the "b" and "s" necessary to achieve (statistically) the target image quality for the block can be applied directly from step 50. Steps 52, 54, 58, 60 and 62 then become unnecessary.

FIG. 9 illustrates a second embodiment of the invention, compatible with the JPEG compression standard.

The digital image to be compressed I is provided as input. A first step 80 for obtaining 8x8 blocks of the digital image I and application of an 8x8 block DCT is implemented. Blocks of transformed coefficients are obtained. Next, a quality target determining step 82, according to the invention, for all the coefficient blocks is applied. Block quality targets are stored.

Then, for each block B _k of image coefficients, a quantization of the coefficients by applying a quantization matrix TQ (U, V) is applied to the quantization step 86. The quantization matrix Q (u, v) is a stored predetermined matrix, and γ is a weighting coefficient applied to all the coefficients of the matrix Q (u, v). In one embodiment, all the coefficients of the matrix Q (u, v) are multiplied by γ. During initialization, γ = 1.

A block of quantized coefficients, denoted qB _k , is obtained.

The mean squared error mse between the quantized coefficient block qB _k and the block B _k is computed at the quantization error calculation step 88.

The squared mean error mse is compared to the mean squared error target in step 90.

If the mean squared error mse is less than the MSE quality target _k , the value of the current weighting coefficient γ is stored as an encoding indication during a step 92.

If the mean squared error mse is greater than the quality target MSE _k for the current block B _k , the weighting coefficient y is decreased during a step 94 of updating the weighting coefficient. For example, γ is divided by 2. Alternatively, a new value of γ is selected from a set of predetermined values stored.

Steps 86 to 94 are implemented for each block, and make it possible to determine a quantization level according to the target of MSE quality _k to be respected. The coding indication for each block is stored and indicated in association with the compressed data, for example in the header of a compressed data file.

The steps described above are followed by a conventional range-length event encoding step applied to the frequency coefficients of each quantized block.

Alternatively, several quantization matrices Q ^* (u, v) are previously stored, and one of the stored matrices is selected for quantization. An indication relating to the matrix chosen from the set of stored matrices is then coded. According to another variant, making it possible to simplify the calculations, the values of γ or the matrices to be applied can also be determined statistically according to the variance of the block and the target image quality. They can therefore be applied from step 82. Steps 88, 90 and 94 then become unnecessary

The method of the invention has been described above in particular in its adaptation with the CCSDS IDC compression standard and with the JPEG compression standard.

However, it is clear that the method according to the invention generally applies with any other compression method, in particular with the JPEG2000 compression standard.

In particular, any size of the treated blocks can be implemented with the method of the invention.

Claims

1. A digital image compression method represented by at least a two-dimensional array of image samples, each sample having an associated intensity value, comprising steps of:

quantization and coding of the blocks of coefficients, in order to obtain a set of digital data representative of the compressed digital image,

characterized by comprising, after obtaining blocks of coefficients, the steps of:

for at least one treated block selected from the blocks of coefficients:

Calculating (36) a first value representative of said processed block from the frequency coefficients of said processed block,

Obtaining (38) a second value representative of said processed block calculated according to an image acquisition noise model associated with said processed block,

Determining (46, 48) a quality target according to said first and second values associated with said processed block,

Quantization and coding of said processed block according to said target of determined quality.

2. - Compression method according to claim 1, characterized in that, for a processed block, said quality target is a value representative of the mean squared error between the coefficients of the block processed before quantization and the coefficients of the block processed after quantification.

3. - Compression method according to any one of claims 1 or 2, characterized in that it further comprises, for each processed block, a step of determining (56) a quantization level to meet the target quality.

4. - Compression method according to any one of claims 1 to 3, characterized in that it comprises, for each treated block, a comparison (40) of said first value to a first threshold level calculated from said second value, and when said first value is lower than said first threshold level, a step of zeroing (46) the frequency coefficients of said processed block.

5. - Compression method according to claim 4, characterized in that, when said first value is greater than said first threshold level and less than a second threshold level, said processed block is considered to be an exception block for which said target quality is selected to minimize the mean squared error between the coefficients of the block processed before quantization and the coefficients of the block processed after quantization.

6. - A compression method according to any one of claims 1 to 5, characterized in that said first value representative of a processed block is representative of the variance of the frequency coefficients of the treated block, and in that said second representative value of a processed block is a function of a variance of a noise model depending on the value of the low frequency coefficient of the processed block.

7. A compression method according to any one of claims 1 to 6, characterized in that said treated blocks are selected by application (34) of a bit mask.

8. - Compression method according to claim 7, characterized in that it comprises, for each block not selected as a block processed by applying said mask, a step of zeroing (35) the frequency coefficients of said block.

9. - A compression method according to any one of the preceding claims, characterized in that it comprises a step of coding (56, 92) an indication relating to a quantization level to be applied for each treated block, a so-called coding indication being associated with the set of digital data representative of the compressed digital image.

The compression method according to claim 9, wherein the quantization and coding step is implemented by bit plane coding. representative of the coefficients, characterized in that said indication is an indication relating to the number of bit planes to be encoded.

1 1. - A compression method according to any one of the preceding claims, characterized in that the blocks of coefficients are obtained by applying a discrete wavelet transformation and spatial grouping of the conversion coefficients obtained.

12. The compression method as claimed in claim 9, in which the quantization step is implemented by applying a block quantization matrix, characterized in that said indication is an indication relating to a weighting coefficient to be applied. to a predetermined quantization matrix.

13. - A compression method according to claim 12, characterized in that the coefficient blocks are obtained by applying a discrete cosine transformation applied to blocks of predetermined size.

14. - Digital image compression device represented by at least one two-dimensional array of image samples, each sample having an associated intensity value, comprising:

a module (6) for applying a transformation in order to obtain a representation of the image comprising a set of coefficients blocks, each coefficient block comprising a coefficient called a low frequency coefficient corresponding to an average value of the intensity of the image samples of said block and a plurality of coefficients called frequency coefficients;

a module (10) for quantizing and coding the coefficient blocks, in order to obtain a set of digital data representative of the compressed digital image,

characterized in that it comprises, modules, able to achieve, after obtaining blocks of coefficients, for at least one treated block selected from the blocks of coefficients:

A calculation (8) of a first value representative of said processed block from the frequency coefficients of said processed block,

Obtaining (8) a second value representative of said processed block calculated according to a noise model associated with said processed block, A determination of a quality target according to said first and second values associated with said processed block,

A quantization and a coding (10) of said processed block according to said target of determined quality.