Method for digitally watermarking an image, method for extracting a digital watermark and method for detecting falsification of a watermarked image

Abstract

The invention relates to a computer-implemented method for digitally watermarking an image, the method comprising the steps of: - receiving (201) a digital image; - decomposing (202) the digital image into a plurality of sub-blocks having the same predefined dimensions; - generating (203) a digital binary signature; - for each sub-block; i transforming the sub-block into an intermediate space via a predetermined transform; ii. combining the signature with the sub-block in the intermediate space to obtain a signed sub-block in the intermediate space; iii. applying an inverse transform, which is the inverse of the predetermined transform, to the signed sub-block; - concatenating (206) the set of signed sub-blocks to form a signed image.

Description

DESCRIPTION Title of the invention: Method for digitally watermarking an image, method for extracting a digital watermark, and method for detecting falsification of a watermarked image

[0001] The invention relates to the field of digital image integrity monitoring, in particular images acquired by an Earth observation satellite.

[0002] The invention relates more specifically to a method of digitally watermarking an image, a method of extracting a watermark and a method of detecting falsification of a watermarked image.

[0003] In the context of increasing amounts of data, particularly images, transmitted via digital transmission systems, there is a general need to detect intentional or unintentional damage to the integrity of received images.

[0004] In particular, in the field of transmitting images acquired by an Earth observation satellite, this need is particularly present because the acquired images can pass through different equipment, communication networks and transmission platforms.

[0005] In this context, the possibilities of compromising image integrity increase, and it is imperative for a user to be able to verify this integrity before the final use of the images.

[0006] A more specific problem concerns the localization, within the image, of an image falsification, that is, the ability to identify which area of ​​an image has been falsified.

[0007] Another specific problem concerns the robustness of a digital watermarking method to certain so-called classic transformations such as image cropping or reframing transformations.

[0008] Various state-of-the-art solutions exist to address the issue of image integrity.

[0009] One known solution involves calculating, at the time of image production or delivery, a unique key of fixed length representing Specifically, an original object and its characteristics. This key can be recalculated during image integrity checking and compared to the original key to determine if the image has been modified. This first method has the advantage of low complexity; however, it has several drawbacks. First, it is not robust to any transformations applied to the image. The slightest modification, such as cropping or mirroring, results in a different key and thus leads to the detection of image falsification when the image has only undergone an acceptable transformation.

[0010] Another drawback is that this method involves a calculation on the entire image, which does not allow for the localization of areas of the image affected by falsification.

[0011] A second known solution involves applying a digital watermark to an entire image. If the image is modified, the watermark is visibly altered, allowing for the detection of tampering. The main drawback of this method is that it visually and functionally impairs the image's usability. Another disadvantage is that this method does not allow for the localization of tampering to a specific area of ​​the image.

[0012] A third known solution involves analyzing the content of the received image and comparing it to a statistical expectation to assess the probability of authenticity of the information contained within the image. This solution is not applicable to all technical fields because it assumes strong preconceptions about the content of the processed images and any potential falsifications applied to them.

[0013] The invention consists of a new digital watermarking method which involves integrating a watermark on each sub-block of a grid applied to the image in such a way as to allow localization of falsification in the image.

[0014] The invention also relates to a method for extracting the digital tattoo and a method for locating forgeries that does not require a reference tattoo.

[0015] Advantageously, the invention relies on an artificial intelligence engine to perform the learning of tattooing and tattoo extraction functions.

[0016] The proposed method has the particular advantage of being robust to image cropping or reframing operations.

[0017] The invention is advantageously applicable to the verification of the integrity of satellite images of Earth observations.

[0018] The invention relates to a computer-implemented method for digitally watermarking an image, comprising the following steps: - To receive a digital image, - Decompose the digital image into a plurality of sub-blocks of the same predefined dimensions, - Generate a digital binary signature, - For each subblock, i. Transform the subblock into an intermediate space using a predetermined transform, ii. Combine the signature with the subblock in the intermediate space to obtain a signed subblock in the intermediate space, iii. Apply to the signed subblock a transform inverse to the predetermined transform. - Concatenate all the signed sub-blocks to form a signed image.

[0019] In one embodiment, the digital tattooing method according to the invention further comprises the steps of: - Add a binary pattern to the digital binary signature, - Transform the binary signature into a matrix with dimensions smaller than the dimensions of a subblock, - The binary pattern is inserted into the binary signature so that it is positioned at identical predefined positions in the matrix for each sub-block

[0020] According to a particular aspect of the invention, the binary pattern is positioned on at least three areas of the matrix corresponding to three corners of a square, for example three corners of the matrix.

[0021] According to a particular aspect of the invention, the step of combining the signature with a sub-block is carried out by means of a first artificial intelligence engine trained to minimize the visual impact of a signature on a signed image.

[0022] The invention also relates to a computer-implemented method for extracting a digital watermark, comprising the steps of: - To receive a digital image signed using the digital watermarking method according to the invention, - Decompose the digital image into a plurality of sub-blocks of the same dimensions used to generate the signed digital image, - For each sub-block, i. Extract the binary signature tattooed in the sub-block

[0023] According to one particular aspect of the invention, the step of extracting the binary signature of a subblock is carried out by means of a second artificial intelligence engine trained to minimize a difference between the binary signature originally generated and the binary signature extracted from a subblock.

[0024] According to a particular aspect of the invention, the first artificial intelligence engine and the second artificial intelligence engine are trained jointly.

[0025] In one embodiment, the method for extracting a digital tattoo according to the invention further comprises the steps of: - Detect, for at least one sub-block, a binary pattern inserted into the extracted signature, - Compare the positions of the detected binary pattern to a set of expected positions, and if they differ from the expected positions, deduce that the received digital image has undergone a clipping operation. - Reconstruct the binary signature associated with the image before the clipping operation, using the positions of the binary pattern, by applying to the digital signature an operation inverse to the clipping operation.

[0026] The invention also relates to a computer-implemented method for detecting the falsification of a digitally watermarked image using the digital watermarking method according to the invention, the method comprising the steps of: - Execute the method for extracting a digital watermark from an image received according to the invention, - Search, for all sub-blocks of the image, for the predominant extracted signature within that set, - For each sub-block, i. compare the extracted signature for the sub-block to the majority signature and, ii. deduce whether the sub-block of the image has been falsified or not

[0027] According to one particular aspect of the invention, the digital image is an image acquired by a satellite.

[0028] The invention also relates to a computer program comprising instructions for executing the method according to the invention, when the program is executed by a processor.

[0029] The invention also relates to a processor-readable recording medium on which is recorded a program containing instructions for executing the method according to the invention, when the program is executed by a processor.

[0030] Other features and advantages of the present invention will become more apparent from the following description in relation to the following attached drawings.

[0031] [Fig. 1] represents a general schematic diagram of a digital tattooing method according to prior art,

[0032] [Fig. 2] represents a diagram of a digital tattooing method according to one embodiment of the invention,

[0033] [Fig. 3] represents a diagram of a method for extracting a digital tattoo according to an embodiment of the invention,

[0034] [Fig. 4] represents a diagram of an artificial intelligence engine trained to perform a first digital watermarking function and a second digital watermark extraction function according to an embodiment of the invention,

[0035] [Fig. 5] represents an example of an artificial intelligence model trained to perform the first digital watermarking function,

[0036] [Fig. 6] represents an example of an artificial intelligence model trained to perform the second function of digital tattoo extraction,

[0037] [Fig. 7] represents a flowchart detailing the steps for implementing a method for detecting falsification according to an embodiment of the invention,

[0038] [Fig. 8a] represents, for the same example, the extracted digital watermark if no image cropping has taken place,

[0039] [Fig. 8b] represents an example of a digital watermark extracted following an image cropping operation,

[0040] [Fig. 9a] represents a first example of a pattern combined with a digital tattoo according to an embodiment of the invention,

[0041] [Fig. 9b] represents a second example of a pattern combined with a digital tattoo according to an embodiment of the invention,

[0042] Figure 1 illustrates a schematic diagram of a method for digitally watermarking an image to verify image integrity.

[0043] The principle of digital watermarking is to integrate into an image I a watermark or binary signature that is not visible in the sense that it does not alter the content of the image but which is modified when a falsification of the image is carried out during the transmission chain.

[0044] Figure 1 shows an image I which is projected into a particular space by means of a given transformation, for example a Fourier transform, a DCT transform or other to obtain a transformed image I'.

[0045] A binary tattoo or signature W is then combined with the transformed image I' to obtain a tattooed image l' t The tattooed image t in the transformation space is then converted back into the initial space via an inverse transformation to obtain the signed image l t .

[0046] To extract the W signature from the tattooed image l t, inverse operations are applied, that is to say we transform the tattooed image in the transformation space and then we apply the inverse operation to the combination operation to recover the signature.

[0047] It is then possible to compare the extracted signature to the original signature to detect image falsification.

[0048] The combination operation can take different forms but must be robust to certain image processing operations, for example, cropping operations. In other words, a cropping operation of the watermarked image... t must not alter the signature embedded in the image.

[0049] The invention relates to a particular method of digital watermarking which notably allows the localization of a falsification within the image.

[0050] Figure 2 represents a flowchart detailing the steps of a digital tattooing method according to one embodiment of the invention.

[0051] The method begins at step 201 with the reception of a digital image, such as an Earth observation image acquired by a satellite. The watermarking method can be applied, for example, in a ground station belonging to a ground segment of a satellite communications system or in a remote server.

[0052] The image can be a single-channel image or a multi-spectral image comprising several channels at different wavelengths.

[0053] In step 202, the image is decomposed into sub-blocks of identical dimensions according to a predetermined image segmentation grid. The dimensions of the sub-blocks are chosen a priori based on constraints related to data storage and the complexity of the processing, but also depending on the desired level of granularity to locate the impact of image falsification.

[0054] In step 203, a binary signature S, or watermark, is received, which contains, for example, useful information. This signature is then integrated into each sub-block in step 205 according to a digital watermarking process as described in Figure 1. A notable difference is that the watermarking operation is not applied to the entire image I, but to each sub-block into which the same signature is integrated.

[0055] Then in step 206, we reconstruct the signed image by aggregating all the tattooed sub-blocks according to the same decomposition grid used in step 202.

[0056] Figure 3 represents a flowchart detailing the steps of a method for extracting signatures from a digitally watermarked image, according to one embodiment of the invention.

[0057] The method in Figure 3 begins with the reception of an image 301 t which was watermarked using the watermarking method shown in Figure 2. In step 302, the image is decomposed into sub-blocks according to the same slicing grid used to watermark the image. Then, in step 303, a signature is extracted from each sub-block using a signature extraction procedure that performs operations that are the inverse of those used to watermark the image.

[0058] We ultimately obtain a number N of signatures, where N is the number of blocks in the image.

[0059] A 304 tamper detection step can then be applied by comparing these signatures with each other, since these signatures are supposed to be all identical if the image has not been tampered with.

[0060] Figure 4 reveals a particular example of the implementation of the steps of integrating a signature into a subblock (step 205) and extracting signatures from a subblock (step 303).

[0061] In a particular embodiment of the invention, the digital watermarking technique is based on the use of two artificial intelligence engines. The first artificial intelligence engine is trained to learn how to perform an integration function of a binary signature in an image of in order to minimize the visual impact of the signature on the image. The second artificial intelligence engine is trained to learn how to extract a signature from a tattooed image in order to reconstruct the original signature as closely as possible.

[0062] An example of a set of two jointly trained artificial intelligence engines is described in Figure 4.

[0063] An image I is received as input and then decomposed into sub-blocks Bi. Each sub-block is produced as input to a first encoder-type network R1 responsible for transforming each sub-block into a latent space.

[0064] A binary signature S is received as input and transformed into a matrix W of a size at most equal to that of a subblock. The matrix signature W is then concatenated, via a concatenation operation C, to the output vector of the first encoder network R1. This set is then fed to a second decoder network R2, which decodes the information from the latent space to obtain a BSi-signed subblock. A cost function F1 is used to train the autoencoder network consisting of R1 and R2. This cost function is, for example, defined as minimizing the error between a BSi-signed subblock and an original Bi subblock.

[0065] A third network R3 is trained to extract a signature from a signed BSi block to recover an extracted signature Se.

[0066] In one embodiment, the third network R3 is trained by applying certain perturbations to the input blocks to encourage robust learning. The perturbations correspond to alterations to the image that could be made by a third party. For example, this could involve adding different types of noise, changing the intensity or contrast of the image, compressing the image, or performing a geometric transformation such as rotation or scaling.

[0067] A second cost function F2 is used for training the third network R3, for example defined by an error between the extracted signature Se and the original signature S.

[0068] Advantageously, the training of the three networks R1, R2, R3 is carried out jointly from an image database, for example images satellites, by producing at the input of the first network R1 sub-blocks chosen randomly from each image.

[0069] One objective of learning these networks is in particular robustness to certain image transformations, such as for example a mirror transformation, invisibility of the signature buried in the image, absence of functional impact on the usability of the image but also a minimum capacity of the quantities of data that can be included in the signature.

[0070] Different artificial intelligence algorithms can be used to implement the R1, R2, and R3 networks. An example of an AI-compatible model of the system described in Figure 4 is shown in Figures 5 and 6. It is based on the model described in reference [1].

[0071] The diagram in Figure 5 represents an example of an artificial self-encoding neural network to implement the R1 and R2 networks in Figure 4.

[0072] It consists of a first fixed-parameter convolution layer 501 to transform a block of an image into a latent space, a concatenation operation C to concatenate a signature S to the transformed image, an artificial neural network 502 and a second fixed-parameter convolution layer 503 to transform the result obtained into image space.

[0073] The 502 network is trained to learn how to integrate the signature into the image without visual impact. For example, it is composed of several layers of successive convolutions.

[0074] Figure 6 describes an example of an artificial neural network to implement the R3 network in Figure 4.

[0075] It consists of a first 601 convolution layer with fixed parameters and then a set of 602 convolution layers whose parameters are learned to extract a Se signature from a Bs-signed block.

[0076] Without departing from the scope of the invention, any other trainable artificial intelligence model to learn to perform the two aforementioned functions can be used as a replacement for the neural networks described in Figures 5 and 6.

[0077] Figure 7 represents, in a flowchart, the steps for implementing a method for detecting and locating falsification in an image according to an embodiment of the invention. This method corresponds to an example of an embodiment of step 304 of the method in Figure 3.

[0078] The method in Figure 7 begins at step 701 with the extraction of N signatures corresponding to N subblocks of an image as described in Figure 3.

[0079] In step 702, the dominant signature is sought among the N extracted signatures. The dominant or majority signature is the most frequent signature among the N extracted signatures. To perform step 702, the N signatures are compared and grouped into sets of identical signatures. The set containing the largest number of identical signatures is the majority set.

[0080] The following steps, 703 to 707, are executed for each sub-block of the image.

[0081] In step 703, the signature extracted from the subblock with index i is compared to the dominant signature. If these two signatures are identical, it is deduced in step 704 that the subblock with index i has not been tampered with. Otherwise, in step 705, the difference between the two signatures is calculated, and then in step 706, this difference is compared to a predefined detection threshold. If the difference is less than or equal to the threshold, the process returns to step 704 to conclude that the subblock with index i has not been tampered with. Otherwise, it proceeds to step 707 to flag the subblock with index i as a tampered subblock.

[0082] Steps 703 to 707 are iterated for all sub-blocks of the image.

[0083] An optional step 708 allows generating a tamper map indicating which sub-blocks of the image have been tampered with.

[0084] Thus, the method in Figure 7 makes it possible to precisely locate the areas of the image that have been falsified, unlike global methods that integrate a single signature into an image.

[0085] We now describe a particular embodiment of the invention which makes the proposed technique robust to cutting or "cropping" operations on an image.

[0086] Figure 8a illustrates an example of a signature or tattoo represented in the form of a matrix. According to the invention, this tattoo is integrated into each block of an image decomposed into N blocks.

[0087] When an image undergoes cropping or resizing during transmission, the extraction of signatures for each block is distorted, as illustrated in Figure 8b, which shows a signature extracted for a block of an image after it has been resizing. The extracted signature no longer corresponds to the signature embedded within a block because the cropping is no longer consistent with the initial block definition grid.

[0088] To overcome this drawback, it is proposed to add an identifiable pattern to a signature before integrating it into a block.

[0089] An example of such a pattern is given in Figure 9a; it involves placing bits set to 1 in at least three areas 901, 902, 903, and 904, corresponding to corners of the matrix. The patterns thus created in the corners of the matrix must be identical.

[0090] Referring to Figure 2, before the integration of the signature into an image block (step 205), the pattern described in Figure 9a is added to the signature received in step 203, the whole being represented in matrix form. The size of the matrix is ​​less than or equal to that of an image block and is such that the ratio between the size of a block and the size of the matrix is ​​fixed.

[0091] During signature extraction, an additional step 305 (with reference to Figure 3) is executed to recombine the extracted signatures using pattern M to compensate for the image clipping operation.

[0092] To do this, the elements of pattern M are detected in the extracted signature (for example, that of Figure 8b), and then the offset between the initial position of pattern M when it was integrated into a block and the position of this pattern in the extracted signature is deduced. A vertical and a horizontal offset can be estimated. These offsets are then applied to the pixels of the matrix signature to obtain the decoded signature.

[0093] Figure 9b shows another example of a possible M pattern consisting of four elements of the same geometric shape (e.g., crosses) located at the four corners of a square.

[0094] In general, the pattern M consists of at least three elements arranged in three fixed positions, for example, three corners of a square inscribed in the signature matrix. The pattern elements consist of values ​​of 1. Advantageously, they are placed on the edges or corners of the matrix so as to avoid areas where the useful bits of the signature are located.

[0095] The steps of the invention can be implemented as a computer program comprising instructions for its execution. The computer program can be stored on a storage medium readable by a processor.

[0096] The reference to a computer program that, when executed, performs any of the functions described above, is not limited to an application program running on a single host computer. Rather, the terms computer program and software are used here in a general sense to refer to any type of computer code (e.g., application software, firmware, microcode, or any other form of computer instruction) that can be used to program one or more processors to implement aspects of the techniques described herein. Computing means or resources may, in particular, be distributed ("cloud computing"), possibly using peer-to-peer technologies.The software code can be executed on any suitable processor (e.g., a microprocessor) or processor core, or a set of processors, whether located in a single computing device or distributed across multiple computing devices (e.g., as potentially accessible within the device's environment). The executable code for each program, enabling the programmable device to implement the processes according to the invention, can be stored, for example, on the hard drive or in read-only memory. Generally, the program(s) can be loaded into one of the device's storage means before execution. The central processing unit can command and direct the execution of the instructions or portions of software code of the program(s) according to the invention, instructions which are stored in the hard drive. hard drive or in read-only memory or in the other storage elements mentioned above.

[0097] Références

[0098] [1] « ReDMark : Framework for Residual Diffusion Watermarking based on Deep Networks”, M Ahmadi et al

Claims

CLAIMS 1. Method, implemented by computer, of digital tattooing of an image comprising the steps of: - Receive (201) a digital image, - Decompose (202) the digital image into a plurality of sub-blocks of the same predefined dimensions, - Generate (203) a digital binary signature, - For each sub-block, i. Transform the sub-block into an intermediate space via a predetermined transform, ii. Combine the signature with the sub-block in the intermediate space to obtain a signed sub-block in the intermediate space, iii. Apply to the signed sub-block a transform inverse to the predetermined transform, - Concatenate (206) all signed sub-blocks to form a signed image.

2. A digital watermarking method according to claim 1 further comprising the steps of: - Add (204) to the digital binary signature a binary pattern, - Transform the binary signature into a matrix of dimensions smaller than the dimensions of a sub-block, - The binary pattern being inserted into the binary signature so as to be positioned on predefined positions of the matrix identical for each sub-block 3. Digital watermarking method according to claim 2 wherein the binary pattern is positioned on at least three areas of the matrix corresponding to three corners of a square, for example three corners of the matrix.

4. A digital watermarking method according to any preceding claim wherein the step of combining the signature with a sub-block is performed by means of a first artificial intelligence engine trained to minimize the visual impact of a signature on a signed image.

5. Method, implemented by computer, for extracting a digital tattoo comprising the steps of: - Receiving (301) a digital image signed using the digital watermarking method according to claim 1, - Decomposing (302) the digital image into a plurality of sub-blocks of the same dimensions used to generate the signed digital image, - For each sub-block, i. Extract (303) the binary signature watermarked in the sub-block 6. A method of extracting a digital tattoo according to claim 5 wherein the step of extracting the binary signature of a sub-block is performed by means of a second artificial intelligence engine trained to minimize a gap between the originally generated binary signature and the binary signature extracted from a sub-block.

7. Method for extracting a digital tattoo according to claim 6 in which: - the step of combining the signature with a sub-block of the digital watermarking method is performed by means of a first artificial intelligence engine trained to minimize the visual impact of a signature on a signed image, the first artificial intelligence engine and the second artificial intelligence engine are jointly trained.

8. A method of extracting a digital tattoo according to any one of claims 5 to 7 further comprising the steps of: - Detect (305), for at least one sub-block, a binary pattern inserted in the extracted signature, - Compare (305) the positions of the detected binary pattern to a set of expected positions and if they differ from the expected positions, deduce that the received digital image has undergone a cutting operation, - Reconstruct (305) the binary signature associated with the image before the cutting operation from the positions of the binary pattern by applying to the digital signature an operation inverse to the cutting operation 9. A computer-implemented method for detecting falsification of a digitally watermarked image, the method comprising the steps of: - Executing (701) the method of extracting a digital tattoo for a received image according to any one of claims 5 to 8, - Search (702), for all the sub-blocks of the image, the extracted signature which is the majority among this set, - for each sub-block, i. compare (703) the signature extracted for the sub-block to the majority signature and, ii. deduce whether the sub-block of the image has been falsified or not 10. A method according to any preceding claim wherein the digital image is an image acquired by a satellite.

11. Computer program comprising instructions for executing the method according to any one of claims 1 to 10, when the program is executed by a processor.

12. A processor-readable recording medium on which is recorded a program comprising instructions for executing the method according to any one of claims 1 to 10, when the program is executed by a processor.

Images