Methods, devices and computer programs for compressing and decompressing remote sensing images

Abstract

Methods, devices and computer programs for compressing and decompressing remote sensing images. This remote sensing image compression method comprises the steps of: separating (80) an initial digital image into a first and a second digital image, the initial digital image (Im) being equal to the maximum, pixel by pixel, of these digital images; applying a first compression method (64) to the first digital image to obtain first compressed data (D1); wavelet decomposition (66) of the second digital image into an approximation subband (LLK) and a plurality of detail subbands; applying a second compression method (70) to the approximation subband, to obtain second compressed data (D2);for each retail sub-band, estimation (72) of at least two statistical parameters representative of a statistical distribution of the values ​​of said retail sub-band, and coding of said statistical parameters in third compressed data (D3); grouping (74) of the first, second and third compressed data. Figure for the abstract: Figure 2;

Description

Title of the invention: Methods, devices and computer programs for compressing and decompressing remote sensing images

[0001] The present invention relates to a remote sensing image compression method, a remote sensing image compression device and an associated computer program, as well as an associated remote sensing image decompression method, device and computer program.

[0002] The invention lies in the field of image processing obtained by a remote sensing sensor, for example a radar or a sonar, and more particularly in the field of compression / decompression of digital images obtained by radar or sonar.

[0003] Radar or sonar remote sensing sensors enable the acquisition of digital terrain images, which are composed of relatively large pixel matrices. In many applications, remote sensing sensors are mounted on mobile platforms, whether terrestrial, aerial, marine, or underwater, etc. In order to use such digital images, for example, for target recognition, they must be stored and / or transmitted via a given transmission channel to data processing centers, generally located on the ground and possessing sufficient memory and computing resources.

[0004] However, it is not possible to transmit images acquired by remote sensing sensors in raw format over communication channels with limited bandwidth, for example in the case of underwater acoustic communication. The temporary storage of such digital images before processing would also require very large memory resources.

[0005] It is therefore necessary to compress the digital images acquired by remote sensing sensors, for the purpose of their storage and transmission over communication channels.

[0006] Various methods of digital image compression are known, which fall into two main categories: on the one hand, the category of non-destructive (or lossless) compression methods, and on the other hand, the category of destructive or lossy compression methods.

[0007] Non-destructive compression methods make it possible to obtain, after compressing an initial digital image and then decompressing it, a decompressed digital image that is strictly identical to the initial digital image. Known non-destructive compression methods include, for example, the methods LZW (for "Lempel-Ziv-Welsh"), Shannon-Fano, RLE (for "run-length encoding"), or Huffman. There is a theoretical limit to the capacity of non-destructive compression, which can be calculated using Shannon's information entropy. Since the entropy of remote sensing images is relatively high, the compression factor achievable with non-destructive compression methods is therefore limited and insufficient for the intended applications.

[0008] Lossy compression methods induce a difference between the original digital image and the decompressed digital image. Examples of this category include the JPEG and JPEG2000 image compression standards, as well as the SPIHT (Set Partitioning in Hierarchical Trees) algorithm. The higher the compression factor, the greater the information loss, and the more significant the visual difference between the original and decompressed digital images. Furthermore, such methods introduce visual artifacts. Remote sensing images have high entropy and exhibit multiplicative noise, also known as speckle noise, which is specific to this type of image.In the field of remote sensing, technical experts have observed that beyond a certain compression factor using known lossy compression algorithms, the resulting digital images become unusable, for example, for target recognition applications, and contain artifacts that obscure important information. This compression ratio has been estimated at approximately 16, which proves insufficient for transmission over certain types of transmission channels.

[0009] There is therefore a need to develop compression and decompression processes, adapted for remote sensing images and allowing to obtain a higher compression factor while allowing application exploitation of the decompressed images.

[0010] To this end, the invention relates to a remote sensing image compression method comprising, for an initial digital image, formed of a pixel matrix of a given size, each pixel having a numerical value, acquired by a remote sensing sensor, steps, implemented by a computing processor, of: - separating said initial digital image into a first digital image and a second digital image, each of the first and second digital images having the same size as said initial digital image, the first digital image being obtained by keeping the pixels of the initial digital image with a value greater than a predetermined threshold value, the other pixels being set to zero, the initial digital image being equal to the maximum, pixel by pixel, of said first and second digital images, - application of a first compression method to said first digital image to obtain first compressed data, - wavelet decomposition of the second digital image, on a number K of decomposition levels, K being a positive integer, allowing us to obtain an approximation sub-band and a plurality of detail sub-bands; • application of a second compression method on the approximation sub-band, allowing obtaining second compressed data; • for each retail sub-band, estimation of at least two statistical parameters representative of a statistical distribution of the values ​​of said retail sub-band, and coding of said statistical parameters in compressed third data; - grouping of said first, second and third compressed data into a data structure forming a compressed representation of said initial digital image.

[0011] Advantageously, the proposed compression method is particularly well-suited to remote sensing images. It allows for the separation of strong echoes, typically corresponding to structures, in the first digital image from weak echoes, typically corresponding to noise, in the second digital image. Furthermore, advantageously, the method includes statistical parameter modeling of the detail sub-bands of the wavelet decomposition of the second digital image, thus providing an adequate representation of the shimmering noise. Indeed, advantageously, the detail sub-bands consist mainly of noise.

[0012] Thus, advantageously, the proposed compression process makes it possible to obtain a better compression factor than the known destructive compression methods for coherent remote sensing digital images, while making it possible to obtain, thanks to the separation between a first and a second digital image, and the specific processing of these first and second digital images, a decompressed digital image usable for the applications intended.

[0013] According to other advantageous aspects of the invention, the compression process comprises one or more of the following features, taken individually or in all technically possible combinations.

[0014] The second digital image is obtained by preserving the pixels of the initial digital image with a value lower than said threshold value, the other pixels being set to said threshold value.

[0015] The first compression method is a non-destructive compression method.

[0016] The process further comprises, before the application of the first compression method, a first quantification step of the values ​​of said first digital image.

[0017] The second compression method is a non-destructive compression method.

[0018] The method further comprises a second quantification step of the values ​​of said approximation subband before the application of said second compression method.

[0019] The statistical parameters include, for each sub-band of detail, a mean and a standard deviation of the values ​​of said sub-band of detail.

[0020] The statistical parameters further include, for each sub-band of detail, a skewness coefficient and a kurtosis coefficient for the values ​​of said sub-band of detail.

[0021] The invention also relates to a remote sensing image compression device, comprising a computing processor configured, for an initial digital image, formed of a pixel matrix of a given size, each pixel having a numerical value, acquired by a remote sensing sensor, to implement: - a module for separating said initial digital image into a first digital image and a second digital image, each of the first and second digital images having the same size as said initial digital image, the first digital image being obtained by retaining the pixels of the initial digital image with a value greater than a predetermined threshold value, the other pixels being set to zero, the initial digital image being equal to the maximum, pixel by pixel, of said first and second digital images, - a module for applying a first compression method to said first digital image to obtain first compressed data, - a wavelet decomposition module of the second digital image, on a number K of decomposition levels, K being a positive integer, allowing to obtain a sub-band of approximation level K and a plurality of sub-bands of detail; • an application module for a second compression method on the approximation sub-band, allowing the obtaining of second compressed data; • for each retail sub-band, an estimation module of at least two statistical parameters representative of a statistical distribution of the values ​​of said retail sub-band, and coding of said statistical parameters in compressed third data; - a module for grouping said first, second, and third compressed data into a data structure forming a compressed representation of said initial digital image. The invention also relates to a computer program comprising software instructions which, when executed by a computer, implement a remote sensing image compression method as defined above.

[0022] According to another aspect, the invention relates to a remote sensing image decompression method comprising a receiver and a data structure comprising first, second, and third compressed data obtained by a compression method as briefly described above. The decompression method comprises the following steps, implemented by a computing processor: - first decompression of the first compressed data by applying a first decompression method to obtain a first decompressed digital image; - second decompression of the second compressed data by applying a second decompression method to obtain a decompressed approximation sub-band; - wavelet recomposition in K numbers of decomposition levels, using the decompressed approximation sub-band, and detail sub-bands of chosen values, allowing to obtain a second decompressed digital image; - pixel-by-pixel combination of the said first and second decompressed digital images to obtain a decompressed digital image.

[0023] Advantageously, the decompression process, for sub-bands of details with chosen values ​​equal to zero (null values), makes it possible to reconstruct a decompressed digital image which is denoised, while the first digital image, representative of strong echoes, and the approximation sub-band are preserved.

[0024] According to one feature, the decompression process comprises, for each detail sub-band, receiving a scalar parameter between zero and one, pseudo-random noise generation as a function of at least one statistical decompression parameter, said statistical decompression parameter being obtained by multiplying by said scalar parameter at least one of the statistical parameters extracted from the third compressed data and representative of a statistical distribution of the values ​​of said detail sub-band, said pseudo-random noise generation enabling the synthesis of a noise sub-band, the recomposition step in wavelets using the decompressed approximation sub-band and the generated noise sub-bands to obtain said second decompressed digital image.

[0025] Advantageously, the pseudo-random noise generation respects the statistical parameters, and therefore the generated noise, in each sub-band of detail, follows a statistical distribution analogous to that of the noise in the corresponding sub-band of detail in the initial digital image. Thus, a shimmering noise having substantially the same statistical distribution is introduced into the decompressed digital image when the scalar parameter is equal to 1.

[0026] According to one characteristic, for each sub-band of detail, said at least one statistical parameter extracted from the third compressed data is the standard deviation.

[0027] According to another aspect, the invention also relates to a remote sensing digital image decompression device, comprising a receiver and a data structure comprising first, second, and third compressed data obtained by a compression device as briefly described above, the decompression device comprising a computing processor configured to implement: - a decompression module for the first compressed data by applying a first decompression method to obtain a first decompressed digital image; - a module for decompressing the second compressed data by applying a second decompression method to obtain a decompressed approximation sub-band; - a wavelet recomposition module with K levels of decomposition, using the decompressed approximation sub-band and detail sub-bands of chosen values, allowing a second decompressed digital image to be obtained, - a pixel-by-pixel combining module of the said first and second decompressed digital images to obtain a decompressed digital image.

[0028] According to one feature, the decompression device is configured to receive a scalar parameter between 0 and 1, the decompression device further comprising a pseudo-random noise generation module based on at least one statistical decompression parameter, said statistical decompression parameter being obtained by multiplying by said scalar parameter at least one of the statistical parameters extracted from the third compressed data and representative of a statistical distribution of the values ​​of said detail sub-band, said pseudo-random noise generation enabling the synthesis of a noise sub-band, the wavelet recomposition module being configured to use the sub- decompressed approximation band and the noise sub-bands generated to obtain said second decompressed digital image

[0029] According to another aspect, the invention also relates to a computer program comprising software instructions which, when executed by a computer, implement a remote sensing image decompression method as defined above.

[0030] The invention will become clearer upon reading the following description, given solely by way of non-limiting example, and made with reference to the drawings in which:

[0031] [Fig-1] [Fig.1] schematically represents an image processing system of remote sensing, including a remote sensing image compression device and a remote sensing image decompression device;

[0032] [Fig.2] [Fig.2] is a synoptic diagram of the main steps of a process of remote sensing image compression in one embodiment of the invention;

[0033] [Fig.3] [Fig.3] schematically represents sub-bands resulting from a wavelet decomposition of a digital image on 2 levels of decomposition;

[0034] [Fig.4] [Fig.4] is an example of a detail sub-band and a histogram partner ;

[0035] [Fig.5] [Fig.5] is a synoptic diagram of the main steps of a remote sensing image decompression process in one embodiment of the invention.

[0036] Fig. 1 schematically illustrates a remote sensing image processing system 2, the remote sensing images being digital images acquired by a remote sensing sensor 4.

[0037] The remote sensing sensor 4, sometimes also called a coherent sensor, is for example a radar or a sonar, which is configured to acquire digital images.

[0038] Each digital image is defined by at least one pixel matrix of a given size, each pixel having a value representative of the captured signal, coded on a predetermined number of bits, for example 8 to 32 bits.

[0039] The remote sensing sensor 4 is, for example, mounted on a mobile platform 6, for example a ship or a submarine. Of course, these examples are given by way of non-limiting implication, as other types of mobile platforms are conceivable.

[0040] The system 2 includes a remote sensing digital image compression device 8 and a remote sensing digital image decompression device 10.

[0041] The compression device 8 is for example mounted on the mobile platform 6, and connected to the remote sensing sensor 4, for example via a wired link.

[0042] Thus, the compression device 8 is configured to receive digital remote sensing images 22 acquired by the remote sensing sensor 4.

[0043] The compression device 8 and the decompression device 10 are programmable electronic devices, for example computers.

[0044] In the embodiment of [Fig. 1], the compression devices 8 and The 10 decompression points are remote and adapted to communicate via a communication channel 12, using respective communication interfaces 14 and 16. For example, communication channel 12 is a wireless communication channel, and communication interfaces 14 and 16 allow communication via communication channel 12 according to a given communication protocol. Information can thus be advantageously transmitted via a radio-frequency, optical, or acoustic communication protocol (for example, the NATO JANUS underwater acoustic communication protocol referenced under STANAG 4748), noting that the underwater acoustic data rate is generally very low, with a typical range-data rate product on the order of 100 km.kb / s currently, thus justifying compression.

[0045] For example, the decompression device 10 is located in a ground-based data processing center, while the compression device 8 is mounted on a mobile platform.

[0046] The compression device 8 is a programmable electronic device comprising, in addition to the communication interface 14, an electronic memory 18, adapted for storing data in the form of files or registers, for example, and a processing unit 20, which is, for example, a processing unit, capable of executing calculations and computer program code instructions when the device 8 is powered on. The elements 14, 18, and 20 of the compression device are adapted to communicate via communication buses 15.

[0047] The processor 20 of the compression device 8 is configured to implement:

[0048] - a module 24 for separating an initial digital image into a first image digital and a second digital image;

[0049] -a module 26 for applying a first compression method to the first digital image to obtain first compressed DI data;

[0050] -a wavelet decomposition module 28 of the second digital image, on a number K of decomposition levels, K being a positive integer, allowing to obtain an approximation sub-band and a plurality of detail sub-bands;

[0051] -a module 30 for applying a second compression method on the approximation sub-band, allowing the obtaining of second compressed data D2;

[0052] - a module 32 for estimating at least two representative statistical parameters of a statistical distribution of the values ​​of each sub-band of detail and coding of the statistical parameters in third compressed data D3;

[0053] -a module 34 for grouping the first compressed data D1, second compressed data D2 and third compressed data D3 into a data structure 36 forming a compressed representation of the initial digital image.

[0054] In some embodiments, the same compression method is used as both the first compression method and the second compression method.

[0055] The data structure 36 is for example a file, comprising the compressed data D1, D2 and D3, as well as parameters used during compression and useful for decompression, for example the number K of decomposition levels, arranged according to a chosen representation structure.

[0056] Modules 24, 26, 28, 30, 32, 34 are adapted to cooperate, as described in more detail below, to implement the remote sensing image compression method according to the invention.

[0057] In one embodiment, modules 24, 26, 28, 30, 32, 34 are implemented in the form of software instructions forming a computer program, which, when executed by a programmable electronic device, implements a remote sensing image compression method according to the invention.

[0058] In an alternative not shown, modules 24, 26, 28, 30, 32, 34 are each implemented as programmable logic components, such as FPGAs (Field Programmable Gate Arrays), microprocessors, GPGPUs (General-purpose processing on graphics processing), or dedicated integrated circuits, such as ASICs (Application-Specific Integrated Circuits).

[0059] The computer program comprising software instructions is further capable of being stored on a computer-readable medium, not shown. A computer-readable medium is, for example, a medium capable of storing electronic instructions and being connected to a bus of a computer system. For example, a readable medium is an optical disc, a magneto-optical disc, a ROM, a RAM, any type of non-volatile memory (e.g., EPROM, EEPROM, FLASH, NVRAM), a magnetic card, or an optical card.

[0060] The data structure 36 forming a compressed representation of the initial digital image is adapted to be transmitted, via the communication channel 12, to the remote sensing image decompression device 10.

[0061] The decompression device 10 is a programmable electronic device which includes, in addition to the communication interface 16, an electronic memory 38, adapted for storing data in the form of files or registers, for example, and a processing unit 40, which is, for example, a processing unit, capable of executing calculations and computer program code instructions when the device 10 is powered on, as well as a human-machine interface 42, allowing in particular the display of decompressed images.

[0062] Elements 16, 38, 40, 42 of the decompression device are adapted to communicate via communication buses 35.

[0063] The processor 40 of the decompression device 10 is configured to implement: - a module 44 for first decompression of the first compressed data by applying a first decompression method to obtain a first decompressed digital image; - a module 46 of second decompression of the second compressed data by application of a second decompression method to obtain a decompressed approximation sub-band; - a 48 wavelet recomposition module with K levels of decomposition, allowing a second decompressed digital image to be obtained; - a module 50 for combining pixel by pixel the said first and second decompressed digital images to obtain a decompressed digital image; - an optional module 52, for pseudo-random noise generation based on corresponding statistical parameters extracted from the third compressed data in order to obtain a noise sub-band.

[0064] The first decompression method is adapted to decompress digital images compressed using the first compression method.

[0065] The second decompression method is adapted to decompress digital images compressed using the second compression method.

[0066] Modules 44, 46, 48, 50, 52 are adapted to cooperate, as described in more detail below, to implement the remote sensing image decompression process according to the invention.

[0067] In one embodiment, modules 44, 46, 48, 50, 52 are implemented in the form of software instructions forming a computer program, which, when executed by a programmable electronic device, implements a remote sensing image decompression method according to the invention.

[0068] In an alternative not shown, modules 44, 46, 48, 50, 52 are each implemented as programmable logic components, such as FPGAs (Field Programmable Gate Arrays), microprocessors, GPGPUs (General-purpose processing on graphics processing), or dedicated integrated circuits, such as ASICs (Application-Specific Integrated Circuits).

[0069] The computer program comprising software instructions is further capable of being stored on a computer-readable medium, not shown. A computer-readable medium is, for example, a medium capable of storing electronic instructions and being connected to a bus of a computer system. For example, a readable medium is an optical disc, a magneto-optical disc, a ROM, a RAM, any type of non-volatile memory (e.g., EPROM, EEPROM, FLASH, NVRAM), a magnetic card, or an optical card.

[0070] Fig. 2 is a synoptic diagram of the main steps of a remote sensing image compression process according to one embodiment, the process being implemented by a processor of a remote sensing image compression device.

[0071] The method comprises, for an initial digital image Im, of a given size, acquired by a remote sensing sensor, a first step 60 of separating the initial digital image Im into two distinct digital images of the same size, a first digital image Iml and a second digital image Im2, both of the same size as the initial digital image Im.

[0072] For example, when the initial digital image is formed by a pixel matrix Im(l,c) of L rows and C columns, for example L=2P and C=2Q, each of the digital images Iml and Im2 is also formed by a pixel matrix of L=2P rows and C=2Q columns.

[0073] The separation is carried out according to a predetermined threshold value S.

[0074] For example, the value of the threshold S is fixed and previously stored.

[0075] According to one variant, the threshold value S varies locally and is calculated by an automatic gain control or AGC method, applied to an initial normalized digital image, the automatic gain control comprising a division of the image by its local average determined by a low-pass filtering.

[0076] According to another variant, the threshold value S varies locally and is calculated by a time variable gain control method or TVG (from the English "Time Variable Gain").

[0077] A gain control commonly used in sonar or radar imaging consists of compensating for the geometric propagation loss over distance. Thus, in the so-called sonar or radar equation giving the power received at the receiver: - some terms are constant and represent the transmitting power, the antenna gain in transmitting and receiving in the direction considered, and various noise or detection threshold terms; - others are variable with the transmitter-target range R (therefore also with the round trip time), classically with the geometric propagation loss term in -40 log R and an absorption term by the medium in " ®log R - the last one is variable and depends on the target, it describes the fraction of power which is re-emitted towards the receiver (this is the so-called index term).

[0078] A possible TVG algorithm is then the opposite of the variable terms in R, namely +40 log R + a log R so that the data recorded at the receiver only represents the fluctuations of the target independently of R, the threshold value S then being defined with respect to the dynamics that the values ​​corrected by the TVG term take.

[0079] According to a third variant, the threshold value S varies locally and is equal to a certain quantile (for example 95%) of the values ​​taken by the image on a sliding window of predefined size.

[0080] The first digital image Iml, also called the strong echo image, comprises the pixels of the initial digital image Im with a value greater than or equal to the threshold S, the values ​​of the other pixels being set to zero:

[0081] [Math.l] Im 1 ( z, j) = Inii, j) if [Math.l] In^i, j) > 5

[0082] Im 1 (i, j) = 0 if Im(i, j)

[0083] The second digital image Im2, also called the image of weak echoes, includes the pixels of the initial digital image Im whose value is less than the threshold value S, the other pixels being set to S: Im2 is therefore the image Im said to be clipped or "clamped" to the threshold value S.

[0084] [Math.2] Im2(i,j) = Im(ij)​ if [Math.2] Im(i,j)

[0085] Im2(i, j) =S if Im(i, j) >S

[0086] The initial digital image is obtained by a pixel-by-pixel combination of the first digital image and the second digital image, the combination consisting of taking the maximum value in each pixel:

[0087] [Math.3] Im(i, j) = max(lmI(i, j), Im2(i, j))

[0088] for all values ​​of (i,j), i.e. 0 < i < L; 0 < j <c

[0089] Separate processing is then applied to the first digital image Iml and to the second digital image Im2.

[0090] For the first digital image Iml, the separation step 60 is optionally followed by a first quantization step 62, in order to reduce the number of bits on which each pixel value of the first digital image is represented. The first quantization thus makes it possible to go from a representation of the value of each pixel on B bits to a representation on B' bits, with B' <b.

[0091] The first quantization step 62, or the separation step 60 when the first quantization is not applied, is then followed, for the first digital image Iml, by an application 64 of a first compression method on the first digital image (or the first quantized digital image), preferably normalized beforehand to the full dynamic range (between 0 and 1, or between 0 and 255 depending on the image representation mode) to obtain first compressed DI data.

[0092] When normalization has been applied, the minimum and maximum values ​​of the image Iml before compression are stored alongside the compressed image Dl, so that an approximation of Iml can be restored from the first compressed data DL

[0093] Preferably, the first compression method is a lossless compression method. For example, the PNG algorithm is applied. Advantageously, the first digital image has a large number of zeros and a much lower entropy than the initial digital image, which makes it possible to obtain a compression factor much higher than that obtained for the initial digital image with a first lossless compression method.

[0094] According to one embodiment, a lossy compression method, parameterized to obtain a decompressed image close, according to a chosen distance measurement, to the original digital image, is applied. For example, JPEG or JPEG2000 compression with a quality factor greater than 90 is applied (the quality factor is arbitrary and determined solely by objective or subjective quality criteria of the reconstructed image).

[0095] Thus are obtained first compressed data Dl representative of the first digital image or image of strong echoes of the initial digital image.

[0096] For the processing of the second digital image Im2, or image of weak echoes, the process includes a wavelet decomposition step 66 on a previously chosen number K of decomposition levels, K being a positive integer, preferably between 1 and 4, and preferably equal to 3.

[0097] Wavelet decomposition is known in the field of image processing by the acronym DWT (for "Digital Wavelet Transform"). Wavelet decomposition is a transformation of a digital image I, represented in the spatial domain, into a hybrid pyramidal representation in spatio-frequency sub-bands of different levels of decomposition, simply called levels hereafter, by application of row and / or column decomposition filters.

[0098] By way of example, with reference to [Fig.3], for a digital image I, on K=2 levels of decomposition, the wavelet decomposition makes it possible to obtain a level 2 sub-band hereafter called the approximation sub-band, referenced LL2 in [Fig.3] and six sub-bands hereafter called detail sub-bands, respectively referenced HLb LHiet HHiau level 1, and HL2, LH2 and HH2au level 2.

[0099] In the illustrated example, the digital image I is a square image of size 2qx2q. Level 0 corresponds to the digital image I itself; level 1 comprises sub-strips, each with a size of 2q'x2q', thus the number of rows and columns is halved compared to the digital image I; level 2 comprises sub-strips, each with a size of 2q2x2q2, thus the number of rows and columns is halved compared to the digital image I. More generally, level K comprises sub-strips, each with a size of 2qKx2qK, thus the number of rows and columns is halved compared to the digital image I.

[0100] Wavelet decomposition is carried out in successive levels by applying analysis filters, respectively a low-pass filter (or approximation filter) and a high-pass filter (or detail filter), associated with decimators by two, applied in the horizontal direction (column scanning) and in the vertical direction (row scanning). Thus, at the first level of decomposition, the digital image I is decomposed into a first approximation sub-band, LL1, corresponding to a low-pass filter in each direction; a sub-band LH1, corresponding to a low-pass filter in the horizontal direction and a high-pass filter in the vertical direction; a sub-band HL1, corresponding to a high-pass filter in the horizontal direction and a low-pass filter in the vertical direction; and a sub-band HH1, corresponding to a high-pass filter in both the horizontal and vertical directions.

[0101] At the next decomposition level (level 2), the first approximation sub-band, LL1, is treated analogously to the treatment applied to the digital image I: the approximation sub-band LL1 is decomposed into an approximation sub-band, LL2, corresponding to a low-pass filtering in each of the directions; a sub-band LH2, corresponding to a low-pass filtering along the horizontal direction and high-pass filtering along the vertical direction; a sub-band HL2, corresponding to high-pass filtering along the horizontal direction and low-pass filtering along the vertical direction; a sub-band HH2, corresponding to high-pass filtering along the horizontal and vertical directions.

[0102] In general, to go from a level d to a level (d+1), the approximation subband LLd is decomposed by low-pass and high-pass filters associated with decimators as described above.

[0103] Each approximation sub-band is a small image (i.e., a smaller image) representative of the digital image I, while the detail sub-bands include components representative of the noise present in the digital image I.

[0104] Wavelet decomposition is preferably carried out according to the Mallat algorithm, described in the article "A Theory for Multiresolution Signal Decomposition: The Wavelet Representation", published in IEEE Transactions on Pattern Analysis and Machine Intelligence, 11, July 1989, which uses so-called separable orthogonal discrete wavelets, allowing a calculation with a linear algorithmic complexity depending on the number of pixels in the image, therefore faster than with arbitrary wavelets.

[0105] In a preferred embodiment, the decomposition implements the 5th order Daubechies wavelet filters or the 5th order Coiflet type filters, described in "Ten Lectures on Wavelets" by I. Daubechies, published in SIAM, 1992.

[0106] Returning to [Fig.2], the process therefore includes a step 66 of wavelet decomposition of the second digital image on K levels of decomposition, by applying a previously chosen wavelet filter.

[0107] The number of decomposition levels and the wavelet filter coefficients are parameters of the compression process.

[0108] According to a first embodiment, these parameters are fixed in advance and known to the compression device and the decompression device.

[0109] According to a second embodiment, the type of wavelets used and the number of decomposition levels used is indicated in the compressed data by ad hoc data fields (one for the type, the other for the number of levels, both fields classically having far fewer than 256 possibilities, which ultimately represents a negligible additional cost of two bytes).

[0110] According to one variant, the number of decomposition levels and the wavelet filter coefficients are selected dynamically at compression, and inserted into the data structure forming a compressed representation of the initial digital image, in order to be used by the decompression process.

[0111] Following the wavelet decomposition, 66 are obtained: a K-level approximation sub-band, and a plurality of detail sub-bands, comprising 3 detail sub-bands per level, so a total of 3K detail sub-bands, as explained above.

[0112] For example, for K=2, the six detail sub-bands HLi, LHiet HHiau level 1, and HL2, LH2et HH2au level 2 are obtained.

[0113] By way of example, for an initial digital image of size 512x512, and K=2 levels of decomposition, the approximation subband LL2 is a matrix of 128x128 values, the detail subbands HLi, LHi and HHi at level 1 each have a size of 256x256 and the detail subbands HL2, LH2 and HH2 at level 2 each have a size of 128x128.

[0114] The process then again includes differentiated treatments of the K-level approximation sub-band and the detail sub-bands.

[0115] The LLK approximation subband is optionally quantized at the second quantization step 68, in order to reduce the number of bits on which each value of the LLK approximation subband is represented. The second quantization thus makes it possible to go from a representation of the value of each pixel on B bits to a representation on B' bits, with B' <b.

[0116] The method comprises, after the second quantization step 68 or directly after the decomposition step, an application 70 of a second compression method on the approximation subband (or the quantized approximation subband) LLK, preferably previously normalized between 0 and 1 (or 0 and 255 depending on the image representation mode), to obtain second compressed data D2. When normalization is applied, the minimum and maximum of the LLK image before compression are stored alongside the compressed image D2, so that an approximation of LLK can be restored from D2.

[0117] Preferably, the second compression method is a lossless compression method. For example, and without limitation, the PNG algorithm is applied.

[0118] According to one variant, the second compression method is a lossy compression method, parameterized to obtain a decompressed image close, according to a chosen distance measure, to the initial image.

[0119] In one embodiment, the second compression method is the same compression method as the first compression method.

[0120] The method further comprises, for each respective sub-band of detail among {HLi, LH^ HHi}, i = 1... K, a step 72 of estimating at least two statistical parameters representative of a statistical distribution of the values ​​of the sub-band of detail.

[0121] The statistical parameters include, in particular, the mean of the values ​​in the sub-band of detail and the variance or standard deviation of the values ​​in the sub-band of detail. These parameters are also known as the first and second moments of a statistical distribution.

[0122] Thus, for example, in step 72 the following are calculated: for the detail sub-band HLi, the mean mHLi and the standard deviation δHli; for the detail sub-band LHi, the mean mL Hi and the standard deviation δLhi; for the detail sub-band HHi, the mean mHHi and the standard deviation δHhii; for the detail sub-band HL2, the mean mHL2 and the standard deviation δHl2; for the detail sub-band LH2, the mean mL H2 and the standard deviation δLH2; for the detail sub-band HH2, the mean mHH2 and the standard deviation δHH2. In this case, for K=2 levels of decomposition, 3 x K x 2 statistical parameters are obtained.

[0123] In another embodiment, which provides an improvement in the approximation accuracy of the statistical distribution of the detail sub-bands, step 72 of estimating the statistical parameters further includes the calculation of the skewness coefficient q (or third-order moment), also called "skewness" in English, of the values ​​of the detail sub-band, and optionally the calculation of the kurtosis coefficient K of the values ​​of the sub-band.

[0124] As is known in the field of statistics, for a random variable X, whose mean is denoted m and standard deviation is denoted 0, the skewness coefficient is obtained by the formula:

[0125] [Math.4] 7 / = e[ ( ) ]

[0126] And kurtosis is obtained by the formula:

[0127] [Math.5] K = e[ (^ ) ]

[0128] Where E[.] denotes the expectation, or in other words the mean over the available values, which are samples of the statistical distribution under consideration.

[0129] In this embodiment, four respective statistical parameters are calculated for each sub-band, and for K=2 levels of decomposition, 3 x K x 4 statistical parameters are obtained.

[0130] The statistical parameters are coded, for example each parameter is coded on O bytes, where O is for example equal to 4, and form a third compressed data D3.

[0131] Advantageously, even using 4 statistical parameters per sub-band of detail, each coded on 4 bytes, the third compressed data has a significantly smaller number of bits compared to the case where these sub-bands of detail are coded in a traditional way.

[0132] The third compressed data is representative of the distributions of the detail subbands of the wavelet decomposition of the second digital image Im2.

[0133] The compression process then includes a step 74 grouping of the first, second and third compressed data D1, D2, D3 into a data structure, Comp(Im), forming a compressed representation of the initial digital image.

[0134] As explained previously, depending on the embodiment, additional parameters, such as the number of decomposition levels and the coefficients of the applied wavelet filter, are also optionally inserted into the data structure.

[0135] Fig. 4 illustrates an example of two-level wavelet decomposition, referenced 73, of a digital image 71. The detail subband HH1, referenced 75, is enlarged, and the histogram 77 of the values ​​of the detail subband HH1 is also shown.

[0136] In a known manner, the histogram of values ​​is representative of the statistical distribution of the values ​​considered.

[0137] Thus, in the example of [Fig.4], we observe a statistical distribution close to a Gaussian distribution, which can then be represented by two statistical parameters, the mean and the standard deviation.

[0138] Fig. 5 is a synoptic diagram of the main steps of a remote sensing image decompression process according to one embodiment, the process being implemented by a processor of a remote sensing image decompression device.

[0139] The decompression process is associated with the compression process, and implements a decompression from a data structure forming a compressed representation of an initial numeric, the data structure being obtained by a compression process as described above.

[0140] The decompression process includes a step 80 of extracting the first compressed data D1, the second compressed data D2 and the third compressed data D3 from a received data structure Comp(Im).

[0141] The process then comprises applying a first decompression method to the first compressed data to obtain a first decompressed digital image hnl', which is then rescaled, if necessary, to recover the dynamic range of Iml using the minimum and maximum parameters that had been saved in Comp(Im). The first decompression method is adapted to decompress digital images compressed using the first compression method.

[0142] For example, the non-destructive PNG method is implemented.

[0143] Any other method of first decompression, non-destructive or destructive, associated with the first compression method, is applicable.

[0144] The process includes a step 84 of applying a second decompression method to the second compressed data D2, accompanied, where applicable, by the saved minimum and maximum values ​​of the approximation subband, thereby obtaining a decompressed approximation subband. The second decompression method is adapted for decompressing digital images compressed using the second compression method.

[0145] For example, the non-destructive PNG method is implemented. Any other second decompression method, non-destructive or destructive, associated with the second compression method, is applicable.

[0146] In one embodiment, the method further comprises, for each set of statistical parameters of the third compressed data D3, corresponding to a sub-band of detail, a step 86 of pseudo-random noise generation as a function of the corresponding statistical parameters, in order to obtain a noise sub-band which will be used as a sub-band of detail in a subsequent wavelet recomposition step 88.

[0147] According to one variant, step 86 is not implemented, and detail sub-bands formed of zero values ​​are used in wavelet recomposition step 88.

[0148] Wavelet recomposition 88 consists of applying the inverse wavelet transform (or inverse DWT) of that applied in the wavelet decomposition step during compression. In particular, the filter coefficients used during compression are known at the decompression stage, either by default or extracted from the received data structure.

[0149] At the end of the wavelet recomposition step, a second decompressed digital image Im2' is obtained.

[0150] In the case where the detail sub-bands are formed of zero values, the second decompressed digital image Im2' is denoised.

[0151] In the case where the detail sub-bands are synthesized by pseudo-random noise generation from the representative parameters of the statistical distribution of each detail sub-band, estimated by the statistical estimation module 32 during compression, the second decompressed digital Im2' has a shimmer noise with a statistical distribution similar to the statistical distribution of shimmer noise in the initial digital image.

[0152] In the case where the detail sub-bands are formed by pseudo-random noise generation with a standard deviation greater than zero and less than that estimated by the statistical estimation module 32, the second decompressed digital Im2' features a statistically distributed, intermediate intensity shimmer noise, thus providing partial image denoising.

[0153] Advantageously, it is possible to use a single scalar parameter P_scal between 0 and 1, adjustable by the operator. A scalar parameter P_scal is optionally provided as input to step 86, for example by a user.

[0154] Alternatively, the scalar parameter P_scal is fixed to a given value according to the intended application.

[0155] The scalar parameter P_scal is used to weight all standard deviations, and more generally statistical parameters, initially estimated by the estimation module 32, during the reconstruction of the noise sub-bands in generation step 86.

[0156] In other words, the process implements a multiplication of one or more of the statistical parameters extracted from the third compressed data to form statistical decompression parameters, the statistical decompression parameters being used for the synthesis by pseudo-random generation of the detail sub-bands.

[0157] A value of zero for the scalar parameter corresponds to the absence of noise, a value of 1 for the scalar parameter corresponds to a reconstruction as close as possible to the original image, and an intermediate value to a partially denoised reconstructed image.

[0158] The decompression process then includes a step 90 of pixel-by-pixel combination of the first decompressed digital image Iml' and the second decompressed digital image Im2' to obtain a decompressed digital image Im'.

[0159] The combination consists of taking the maximum, for each pixel, between the value contained in the first decompressed digital image Iml' and the value contained in the second decompressed digital image Im2'.

[0160] [Math.6] = max( / m 1 '(i, j), Im2'(J, j))

[0161] Advantageously, thanks to the image separation of strong and weak echoes, wavelet decomposition and the introduction of statistical parameters characterizing the sub-bands of detail, the proposed compression method makes it possible to obtain high compression factors for remote sensing images, on the order of 16 to 100 (on images presenting an almost uniform background such as a sandbank) while preserving a quality satisfactory for the interpretation of these images. < / c> ​

Claims

1. Demands A remote sensing image compression method, comprising, for an initial digital image formed from a pixel matrix of a given size, each pixel having a numerical value, acquired by a remote sensing sensor, steps, implemented by a computing processor, of: separation (80) of said initial digital image (Im) into a first digital image (Iml) and a second digital image (Im2), each of the first and second digital images having the same size as said initial digital image, the first digital image being obtained by preserving the pixels of the initial digital image with a value greater than a predetermined threshold value, the other pixels being set to zero, the initial digital image (Im) being equal to the maximum, pixel by pixel, of said first and second digital images, application of a first compression method (64) on said first digital image (Iml) to obtain first compressed data (Dl), wavelet decomposition (66) of the second digital image (Im2), on a number K of decomposition levels, K being a positive integer, allowing to obtain an approximation sub-band (LLK) and a plurality of detail sub-bands; • application of a second compression method (70) on the approximation subband, allowing to obtain second compressed data (D2); • for each retail sub-band, estimation (72) of at least two statistical parameters representative of a statistical distribution of the values ​​of said retail sub-band, and coding of said statistical parameters in compressed third data (D3); grouping (74) of the said first, second and third compressed data (D1, D2, D3) into a data structure forming a compressed representation of said initial digital image.

2. A compression method according to claim 1, wherein said second digital image (Im2) is obtained by preserving the pixels of the initial digital image with a value lower than said threshold value, the other pixels being set to said threshold value.

3. Compression method according to any one of claims 1 or 2, wherein said first compression method is a non-destructive compression method.

4. Compression method according to any one of claims 1 to 3, further comprising, prior to the application of the first compression method, a first quantification step (62) of the values ​​of said first digital image.

5. Compression method according to any one of claims 1 to 4, wherein said second compression method is a non-destructive compression method.

6. Compression method according to any one of claims 1 to 5, further comprising a second quantification step (68) of the values ​​of said approximation subband before the application of said second compression method.

7. Compression method according to any one of claims 1 to 6, wherein said statistical parameters include, for each detail subband, a mean and a standard deviation of the values ​​of said detail subband.

8. Compression method according to claim 7, wherein said statistical parameters further comprise, for each detail sub-band, an asymmetry coefficient and a flattening coefficient of the values ​​of said detail sub-band.

9. A computer program comprising software instructions which, when executed by a programmable electronic device, implement a remote sensing image compression method according to claims 1 to 8.

10. Remote sensing image compression device, comprising a computing processor configured, for an initial digital image, formed of a pixel matrix of a given size, each pixel having a numerical value, acquired by a remote sensing sensor, to implement:

11. - a separation module (24) of said initial digital image (Im) into a first digital image (Iml) and a second digital image (Im2), each of the first and second digital images having the same size as said initial digital image, the first digital image (Iml) being obtained by preserving the pixels of the initial digital image with a value greater than a predetermined threshold value, the other pixels being set to zero, the initial digital image being equal to the maximum, pixel by pixel, of said first and second digital images, - a module (26) for applying a first compression method to said first digital image (Iml) to obtain first compressed data (Dl), - a wavelet decomposition module (28) of the second digital image (Im2), on a number K of decomposition levels, K being a positive integer, allowing to obtain a sub-band of approximation level K and a plurality of sub-bands of detail; • a module (30) for applying a second compression method on the approximation subband, allowing the obtaining of second compressed data (D2); • for each detail sub-band, an estimation module (32) of at least two statistical parameters representative of a statistical distribution of the values ​​of said detail sub-band, and of coding said statistical parameters into compressed third data (D3); - a module (34) for grouping said first, second and third compressed data into a data structure forming a compressed representation of said initial digital image. A method for decompressing digital remote sensing images, comprising a receiver and a data structure comprising first, second, and third compressed data obtained by a compression method according to claims 1 to 8, The process comprises the following steps, implemented by a computing processor: - first decompression (82) of the first compressed data (Dl) by applying a first decompression method to obtain a first decompressed digital image; - second decompression (82) of the second compressed data (D2) by applying a second decompression method to obtain a decompressed approximation subband; - wavelet recomposition (88) in a number K of decomposition levels, using the decompressed approximation sub-band, and detail sub-bands of chosen values, allowing to obtain a second decompressed digital image; - combination (90) pixel by pixel of said first and second decompressed digital images to obtain a decompressed digital image (Im').

12. A decompression method according to claim 11, further comprising, for each detail sub-band, receiving a scalar parameter (P_scal) between zero and one, pseudo-random noise generation (86) as a function of at least one statistical decompression parameter, said statistical decompression parameter being obtained by multiplying by said scalar parameter (P_scal) at least one of the statistical parameters extracted from the third compressed data (D3) and representative of a statistical distribution of the values ​​of said detail sub-band, said pseudo-random noise generation enabling the synthesis of a noise sub-band, the wavelet recomposition step (88) using the decompressed approximation sub-band and the generated noise sub-bands to obtain said second decompressed digital image.

13. Decompression method according to claim 12, wherein, for each detail sub-band, said at least one statistical parameter extracted from the third compressed data is the standard deviation.

14. A computer program comprising software instructions which, when executed by a programmable electronic device, implement a remote sensing digital image decompression method in accordance with claims 11 and 13.

15. A remote sensing digital image decompression device, comprising a receiver and a data structure comprising first, second and third compressed data obtained by a compression device according to claim 10, the decompression device comprising a computing processor configured to implement: - a module (44) for decompressing the first compressed data (D1) by applying a first decompression method to obtain a first decompressed digital image; - a module (46) for decompressing the second compressed data (D2) by applying a second decompression method to obtain a decompressed approximation sub-band;- a wavelet recomposition module (48) with K levels of decomposition, using the decompressed approximation sub-band and detail sub-bands of chosen values, allowing a second decompressed digital image to be obtained, - a pixel-by-pixel combination module of said first and second decompressed digital images to obtain a decompressed digital image.

16. A decompression device according to claim 15, configured to receive a scalar parameter (P_scal) between 0 and 1, the decompression device further comprising a pseudo-random noise generation module (52) based on at least one statistical decompression parameter, said statistical decompression parameter being obtained by multiplying by said scalar parameter (P_scal) at least one of the statistical parameters extracted from the third compressed data (D3) and representative of a statistical distribution of the values ​​of said detail sub-band, said pseudo-random noise generation allowing the synthesis of a noise subband, the wavelet recomposition module (48) being configured to use the decompressed approximation subband and the generated noise subbands to obtain said second decompressed digital image.

Images