A template-embedding-based remote sensing image robust digital watermarking method and system
By designing template embedding and a customized distortion layer for remote sensing images, combined with a deep learning watermark extraction network, the problems of robustness and spectral integrity in the copyright protection of remote sensing images are solved, achieving efficient watermark extraction and security protection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHEJIANG UNIV
- Filing Date
- 2026-03-17
- Publication Date
- 2026-07-10
Smart Images
Figure CN122367705A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of remote sensing image copyright protection technology, and in particular to a robust digital watermarking method and system for remote sensing images based on template embedding. Background Technology
[0002] Remote sensing imagery, as a crucial spatial information resource, holds significant application value in fields such as land resource surveys, environmental monitoring, disaster assessment, and urban planning. With the rapid development of remote sensing technology, the commercial value of high-resolution remote sensing imagery is continuously increasing, making the issue of copyright protection for remote sensing data increasingly prominent. Statistics show that the global remote sensing imagery market exceeds several billion US dollars annually; however, the illegal copying, alteration, and misuse of remote sensing images occur frequently, severely damaging the legitimate rights and interests of remote sensing data providers.
[0003] Traditional methods for protecting the copyright of remote sensing images mainly rely on visible watermarks or metadata annotations, but these methods have significant drawbacks. Visible watermarks can compromise the integrity of remote sensing images, affecting subsequent image interpretation and applications; metadata annotations are easily deleted or tampered with, resulting in low security. Therefore, there is an urgent need for an invisible and robust remote sensing image copyright protection technology.
[0004] Digital watermarking technology offers a potential solution for copyright protection of remote sensing images. By embedding invisible copyright information into remote sensing images, digital watermarking achieves copyright declaration while maintaining the image's visual quality. However, existing image watermarking methods are primarily designed for natural images and do not fully consider the unique characteristics of remote sensing images.
[0005] First, the acquisition and processing of remote sensing images differs significantly from that of natural images. During the process from satellite platform to ground reception, remote sensing images undergo physical distortions such as atmospheric scattering and sensor noise. In the ground processing stage, preprocessing operations such as radiometric calibration, atmospheric correction, and geometric correction are required, which introduce complex distortions into the image. Existing digital watermarking methods are primarily designed to address common digital image processing distortions such as JPEG compression and Gaussian noise, and lack robustness to the unique physical and preprocessing distortions of remote sensing images.
[0006] Secondly, remote sensing images typically contain multiple spectral bands, carrying rich spectral information that is crucial for remote sensing applications. If the watermarking process damages these spectral characteristics, it will severely impact the subsequent applications of the remote sensing image. Therefore, remote sensing image watermarking methods need to protect copyright while preserving the integrity of the image's spectral information as much as possible.
[0007] While existing deep learning-based watermarking methods have achieved good results on natural images, their distortion layer designs are primarily designed for print-to-photograph or screen-to-photograph scenarios, failing to accurately simulate the distortion characteristics of remote sensing images. This results in a significant drop in watermark extraction accuracy in practical remote sensing applications, making it difficult to meet real-world needs. Summary of the Invention
[0008] To address the aforementioned issues, this invention provides a robust digital watermarking method and system for remote sensing images based on template embedding. By employing template-based watermark embedding, customized distortion simulation of remote sensing images, and deep learning-based watermark extraction, highly robust copyright protection for remote sensing images is achieved.
[0009] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention proposes a robust digital watermarking method for remote sensing images based on template embedding, including a watermark embedding stage, a distortion simulation stage, and a watermark extraction stage. The watermark embedding stage includes: receiving the remote sensing image to be protected and copyright watermark information; performing BCH encoding on the copyright watermark information to generate a watermark bit vector; repeatedly embedding and rearranging the watermark bit vector into a matrix, and encrypting it through a key-controlled scrambling matrix to generate a watermark message matrix; generating a corresponding Gaussian circular bit template based on the value of each bit in the watermark message matrix, and combining them into a watermark template with the same resolution as the original remote sensing image; and superimposing the watermark template onto the original remote sensing image with adjustable embedding strength parameters and template strength parameters to generate a watermarked remote sensing image. The distortion simulation stage includes: constructing a customized distortion layer for the remote sensing image, and generating a distorted watermarked remote sensing image by taking the watermarked remote sensing image as input. The customized distortion layer for remote sensing images includes: an atmospheric scattering simulation module for simulating the effects of atmospheric scattering on remote sensing images during atmospheric transmission; a sensor noise simulation module for simulating thermal noise, quantization noise, and stripe noise of remote sensing sensors; a radiometric correction distortion simulation module for simulating brightness and contrast adjustments during radiometric calibration and atmospheric correction processing; and a geometric distortion simulation module for simulating geometric transformations caused by geometric correction, registration, cropping, or rotation of remote sensing images. The watermark extraction stage includes: training a watermark extraction network using a watermarked remote sensing image, a distorted watermarked remote sensing image, and their corresponding watermark message matrix labels; segmenting the remote sensing image to be detected into a sequence of image blocks of the same size as the Gaussian circular bit template; inputting the image block sequence into the watermark extraction network to predict the watermark message matrix; performing inverse scrambling on the predicted watermark message matrix based on the key information, and fusing repeatedly embedded watermark information based on a voting mechanism to obtain a watermark bit vector; and performing BCH decoding on the watermark bit vector to recover the final copyright watermark information.
[0010] Furthermore, in the watermark embedding stage, the process of generating the watermark message matrix includes: The watermark bit vector is repeatedly embedded to generate a robust message bit vector. ; Robust message bit vector Rearranged into a two-dimensional matrix of a preset size a×(a+1) ; The matrix is scrambled using a row scrambling matrix P and a column scrambling matrix Q determined by a key K. Encryption is performed to generate a watermarked message matrix: ; Where P is an a×a row scramble matrix, and Q is an (a+1)×(a+1) column scramble matrix. This means using the a+1 bit key K as the first row, and encrypting it with the a×(a+1) watermark matrix after row scrambling matrix P and column scrambling matrix Q. By vertically concatenating the data, a complete (a+1)×(a+1) watermark message matrix is generated. .
[0011] Furthermore, in the watermark embedding stage, the watermarked remote sensing image is represented as follows: ; in, Parameters for controlling the watermark embedding strength, Parameters for controlling the strength of the template itself, This is a watermarked remote sensing image. This is a watermark template. This is the original remote sensing image.
[0012] Furthermore, in the distortion simulation stage, the atmospheric scattering simulation module uses an atmospheric scattering physical model, adding random atmospheric illumination A and scattering coefficient t to the RGB channels respectively. The simulation formula is as follows: ; in, Let be the atmospheric transport rate randomly sampled from a uniform distribution U[0.7, 0.95]. The atmospheric illumination is randomly sampled from a uniform distribution U[0.5, 0.8]. These are the original pixel values. The pixel value is the result of scattering.
[0013] Furthermore, in the distortion simulation stage, the noise simulated by the sensor noise simulation module includes: a standard deviation of... Randomly sampled Gaussian thermal noise and noise density within the range Salt-and-pepper quantization noise randomly sampled within the range, and strip noise randomly sampled within the range of [1,5].
[0014] Furthermore, in the distortion simulation phase, the simulation process of the radiation correction distortion simulation module includes: Add a random offset to each channel of the image, with the offset ranging from 1 to 2. ; Using linear transformations Adjust the image's contrast and brightness, including the gain. offset , Indicates a uniform distribution; Adaptive adjustment is performed on different brightness areas of the image, enhancing low-brightness areas and suppressing high-brightness areas; The color changes during radiometric correction are simulated by randomly mixing RGB images with their grayscale versions.
[0015] Furthermore, in the distortion simulation stage, the geometric transformations simulated by the geometric distortion simulation module include: random perspective transformation with a four-corner perturbation range of ±25 pixels, random rotation within an angle range of [-15, 15], random scaling within a scaling factor range of [0.85, 1.15], and random cropping within a cropping ratio range of [0.8, 0.95]. The blank areas generated after rotation are filled with boundary pixel values, and after random scaling and cropping, they are adjusted back to their original size.
[0016] Furthermore, the watermark extraction network adopts the ResNet18 architecture, and the loss function is: ; in, For real watermarked message matrix tags, To predict the watermark message matrix from a watermarked remote sensing image, The watermark message matrix is predicted from the distorted watermarked remote sensing image. This is the weighting coefficient, with a value of 2.
[0017] Furthermore, in the watermark extraction stage, the process of inversely scrambling the predicted watermark message matrix based on the key information and fusing the repeatedly embedded watermark information based on a voting mechanism to obtain the watermark bit vector includes: Extract the first row vector from the predicted watermark message matrix as the prediction key; calculate the Hamming distance between the prediction key and the original key used in the watermark embedding stage; if the Hamming distance is less than the preset fault tolerance threshold, the key verification is passed and the watermark is valid; otherwise, the image is determined not to contain a valid watermark. After verifying the validity of the watermark, the corresponding scrambling matrices P and Q are recovered based on the original key. Then, the remaining part of the watermark message matrix except for the first row is decrypted to obtain an intermediate watermark matrix containing multiple repeated embeddings. The intermediate watermark matrix is divided into multiple groups along the column direction, with each group having the same number of columns, and each group corresponds to a repetition of the watermark information. For each bit position in the original watermark information, the corresponding predicted value is extracted from each of the multiple groups. The extracted predicted values are processed by a majority voting function, and the bit value that appears most frequently is determined as the final value of that bit position. All bit positions are traversed to generate the watermark bit vector.
[0018] Secondly, this invention proposes a template-embedded robust digital watermarking system for remote sensing images, which is used to implement the above-mentioned template-embedded robust digital watermarking method for remote sensing images.
[0019] The beneficial effects of this invention are as follows: (1) This invention accurately simulates various distortions in remote sensing images during acquisition, transmission, and processing by carefully designing a customized distortion layer, including atmospheric scattering, sensor noise, radiometric correction, and geometric transformation. This allows the watermark extraction network to fully learn the unique distortion characteristics of remote sensing images during training, thus exhibiting excellent robustness in practical applications. Experiments show that compared to existing image watermarking methods, this invention improves the bit accuracy of watermark extraction from remote sensing images by approximately 10%, reaching over 97%.
[0020] (2) This invention adopts a template-based watermark embedding method, using adjustable intensity parameters. and Balancing the invisibility and robustness of watermarks. Experiments verified that when , At that time, the PSNR of the watermarked remote sensing image reached over 36dB and the SSIM reached over 0.95, with excellent visual quality. At the same time, it maintained the integrity of the spectral information of the remote sensing image and did not affect subsequent remote sensing applications.
[0021] (3) This invention significantly improves the robustness of watermarks to local distortions through a four-times repeated embedding and voting fusion mechanism combined with BCH error correction coding. Even if the remote sensing image has local distortions such as cloud cover and stripe noise, the watermark information can still be reliably extracted. This is of great significance for practical remote sensing applications, because remote sensing images are often affected by factors such as clouds and shadows.
[0022] (4) This invention employs a key-based scrambling matrix to encrypt watermark information, enhancing watermark security. Even if an unauthorized user obtains a watermarked remote sensing image, they cannot extract or tamper with the watermark information without knowing the key. This provides reliable protection for the copyright and integrity verification of remote sensing images.
[0023] (5) The watermarking method of this invention is applicable to various types of remote sensing images, including optical remote sensing images and hyperspectral remote sensing images. By adjusting the distortion layer parameters, it can also be extended to other types of remote sensing data such as SAR images. This provides a unified technical solution for comprehensive copyright protection of remote sensing images.
[0024] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and will become apparent from the description or may be learned by practice of the invention. Attached Figure Description
[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0026] Figure 1 An overall framework diagram of a template-embedded robust digital watermarking method for remote sensing images provided as an example of the present invention; Figure 2 A detailed flowchart of the watermark embedding module provided as an example of the present invention; Detailed Implementation
[0027] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of the invention.
[0028] This embodiment provides a robust digital watermarking method for remote sensing images based on template embedding, such as... Figure 1 As shown, the method includes the following steps: I. Watermark Embedding Stage (1.1) Watermark Encoding like Figure 2 As shown, this embodiment uses a Landsat-8 remote sensing image as an example for explanation. First, a 400×400 pixel Landsat-8 true-color remote sensing image is acquired. The image to be protected contains three bands: RGB. The copyright watermark information M is set to a 28-bit binary sequence, such as "1010110101011010101101010110".
[0029] During watermark encoding, the 28-bit watermark information M is divided into four groups of 7 bits each. Each group is encoded using BCH(15,7) code, thus encoding the 7-bit information into a 15-bit codeword. BCH encoding has the ability to correct up to 2 bit errors, enabling the correction of some erroneous bits during watermark extraction. After encoding, a 60-bit watermark bit vector is obtained. Where n represents the number of groups of watermark information M, and 15 represents the fixed length of each group after BCH encoding. Representing vectors The i-th bit in the n-th encoded block.
[0030] To enhance robustness to local distortions (such as cloud cover and stripe noise), the aforementioned 60-bit watermark bit vector B is repeatedly embedded four times to generate a 240-bit robust message bit vector. .Will Rearranged into a 15×16 matrix .
[0031] Generate a 16-bit key K, for example, "1100101110010110". Based on the key K, generate a 15×15 scrambling matrix P and a 16×16 scrambling matrix Q using a pseudo-random number generation algorithm. The scrambling matrices are permutations of rows and columns, ensuring their linear independence.
[0032] The watermark matrix is encrypted using a scrambling matrix to enhance its security, generating the final watermark message matrix. The calculation formula is expressed as: Here, `concatenate` represents a vertical concatenation operation (row-by-row concatenation). Specifically, it takes the 1×16 key vector K as the first row of the concatenated matrix and concatenates it vertically with the 15×16 watermark matrix PB′′Q, which has been encrypted using row scrambling matrix P and column scrambling matrix Q, to form a final 16×16 watermark message matrix. .
[0033] (1.2) Template generation Considering the potential local distortion in remote sensing images, Gaussian circles were chosen as the basic unit of the bit template. Gaussian circles possess good spatial continuity and can resist local distortion to a certain extent.
[0034] In this embodiment, for the watermark message matrix For each bit in the array (with a value of 1 or 0), a corresponding Gaussian circular bit template is generated. When the bit value is 1, a positive Gaussian circle is generated, and when the bit value is 0, a negative Gaussian circle (or a weaker Gaussian circle) is generated.
[0035] Let the bit template size be = 25 × 25 pixels. For positions with a bit value of 1, generate a positive Gaussian circular template: For positions with a bit value of 0, generate a negative Gaussian circular template: Where (x, y) are the coordinates relative to the center of the template. The standard deviation of the Gaussian function is set in this embodiment. =8.
[0036] The 16×16 bit templates are arranged according to the watermark message matrix. The positions of the images are arranged and stitched together to form a complete 400×400 pixel watermark template T. Template T is then compared with the original remote sensing image. They have the same resolution and number of channels.
[0037] (1.3) Watermark Embedding Set watermark embedding strength parameters = 0.25 and template strength parameters = 0.4. Overlay the watermark template T onto the original remote sensing image. Above, generate watermarked remote sensing images. : Watermarked remote sensing images With the original image Visually, they are almost indistinguishable. PSNR and SSIM are calculated to assess visual quality: In this embodiment, PSNR = 36.8 dB and SSIM = 0.95, indicating that the watermark embedding has little impact on the image's visual quality.
[0038] To verify the impact of watermark embedding on the spectral information of remote sensing images, the spectral angle distance (SAD) between the watermarked image and the original image in each band was calculated. Experimental results show that the average SAD is less than 1.8 degrees, indicating that the watermark embedding basically preserves the spectral characteristics of the remote sensing image.
[0039] II. Customized Distortion Layer for Remote Sensing Images To train a robust watermark extraction network, a distortion layer needs to be designed to simulate various distortions that remote sensing images may suffer in real-world applications. The remote sensing image-customized distortion layer of this invention includes the following modules: (2.1) Atmospheric Scattering Simulation Module This module employs an atmospheric scattering physics model to simulate the effects of atmospheric scattering on remote sensing images during atmospheric transmission. Atmospheric scattering primarily includes Rayleigh scattering (caused by atmospheric molecules) and Mie scattering (caused by aerosols and suspended particles). This module simulates the combined effect of these two types of scattering by randomly sampling atmospheric transmissivity and illumination. The atmospheric scattering physics model used in this invention is as follows: During training, atmospheric transmissivity *t* and atmospheric illumination *A* are randomly sampled for each image. Specifically, *t* is sampled from a uniform distribution U[0.7, 0.95], and *A* is sampled from U[0.5, 0.8]. *t* and *A* are sampled independently for each of the three RGB channels to simulate wavelength-dependent scattering characteristics. For example, for the R channel, = 0.85, = 0.65; For the G channel, =0.88, = 0.62; For channel B, = 0.92, = 0.58. After applying the atmospheric scattering physics model, the contrast of the remote sensing image decreased, exhibiting a slight fogging effect.
[0040] (2.2) Sensor noise simulation module Sensor noise simulation includes three types of noise: (a) Thermal noise: Zero-mean Gaussian noise is added to the image. Random sampling standard deviation. Generate a Gaussian random matrix of the same size as the image. Overlay it onto the image: (b) Quantization noise: Add salt and pepper noise to the image. Randomly sample noise density d ∼ U[0,0.02], randomly select pixels at a ratio of d, and set their values to 0 (salt noise) or 1 (pepper noise).
[0041] (c) Strip noise: Adds a systematic offset to random rows or columns of an image. Randomly select several columns (or rows) of the image and add a fixed offset to the selected columns. Strip width Pixels. For example, randomly select columns 50-53 (width 4) and add an offset of 0.02 to all pixels in these 4 columns to simulate the response differences of the detector unit.
[0042] (2.3) Radiation Correction Distortion Simulation Module Radiation correction distortion simulation includes the following steps: (a) Channel offset: Add random offsets to each of the three RGB channels. For example, add -0.08 to the R channel, 0.05 to the G channel, and 0.12 to the B channel.
[0043] (b) Contrast and brightness adjustment: Apply linear transformation to each channel. Among them, the gain offset .For example, , The transformed pixel value is .
[0044] (c) Adaptive Brightness Adjustment: The image is converted from RGB space to HSV space to obtain the brightness component V. For low-brightness areas where V < 0.5, an enhancement factor is applied. Enhancement is applied; for high-brightness regions with V > 0.7, a suppression factor is applied. Suppress the image. After adjustment, convert the image back to RGB space.
[0045] (d) Saturation Adjustment: The RGB image is randomly blended with its grayscale version to simulate color changes during radiometric correction. Blending Ratio The mixing formula is as follows: in, This is the grayscale version of the image.
[0046] (2.4) Geometric Distortion Simulation Module Geometric distortion simulation includes the following transformations: (a) Perspective Transformation: Add random perturbations to the four corner points of the image. Assume the image size is 400×400, and the initial coordinates of the four corner points are (0,0), (400,0), (400,400), (0,400). Add random perturbation amounts to each corner point. , For example, the perturbed corner coordinates are (8,-15), (415,12), (388,422), (-10,385). The perspective transformation matrix H is calculated based on the perturbed corner coordinates, and bilinear interpolation is used to perform a perspective transformation on the entire image.
[0047] (b) Rotation: Randomly sample the rotation angle Rotate the image using a rotation transformation matrix centered at the image center. After rotation, blank areas may appear at the image boundaries; these are filled using the boundary pixel values.
[0048] (c) Scaling: Random sampling scaling factor After applying the scaling transformation, bilinear interpolation is used to resize the image back to its original size of 400×400.
[0049] (d) Cropping: Cropping ratio based on random sampling Randomly select the top-left corner coordinates of the cropping area and crop. The original size image patch is then resized back to the original size of 400×400 using interpolation.
[0050] (2.5) Distortion layer combination When training the watermark extraction network, the above modules are applied sequentially to the watermarked remote sensing images. In each training iteration, the parameters of each module are randomly sampled to enhance the network's adaptability to different distortion conditions. The distortion process is as follows: Watermarked image → Atmospheric scattering simulation → Sensor noise simulation → Radiometric correction distortion simulation → Geometric distortion simulation → Distorted watermarked remote sensing image.
[0051] To ensure training efficiency, computationally intensive distortion operations (such as perspective transformation and rotation) are implemented using PyTorch's differentiable image transformation functions to ensure gradient backpropagation. Non-differentiable operations (such as salt-and-pepper noise) are treated as independent augmentation steps during training and do not participate in gradient calculation.
[0052] III. Watermark Extraction Stage (3.1) Training the watermark extraction network The watermark extraction network adopts the ResNet18 architecture. The network input is a 25×25×3 image block, which corresponds to a bit template region; the network output is the predicted watermark bit (probability of 0 or 1) for that image block.
[0053] Training data preparation: 40,000 remote sensing images were randomly selected from remote sensing databases such as Landsat and Sentinel, and their resolution was adjusted to 400×400 pixels. A random 28-bit watermark was generated for each image, and watermarked remote sensing images were generated according to the watermark embedding process. .
[0054] During training, each watermarked remote sensing image was processed. Two samples were generated respectively: (1) Clear sample: directly... The image is divided into 16×16=256 blocks, each 25×25 pixels in size, corresponding to the watermark message matrix. One bit position.
[0055] (2) Distorted samples: Input a remote sensing image, customize a distortion layer, and generate a distorted watermarked remote sensing image. Then it was divided into 256 image blocks.
[0056] Loss function design: The first term is the clear sample loss. The first term represents the network's prediction of sharp image patches; the second term represents the loss for distorted samples. For the network to predict distorted image patches; These are the weighting coefficients. Using the Adam optimizer, the learning rate is 5e. -5 Train for 3 epochs.
[0057] (3.2) Watermark Extraction Process For the remote sensing image to be detected (The watermark may have undergone various processing and distortion processes) To extract the watermark, follow these steps: Step 1: Image patch segmentation.
[0058] Will The image is divided into 16×16=256 image block sequences according to a 25×25 grid. .
[0059] Step 2: Watermark bit prediction.
[0060] 256 image patches are input into the trained watermark extraction network to obtain 256 bits of prediction probability. For each bit, if the prediction probability is > 0.5, it is predicted as 1; otherwise, it is predicted as 0. This yields the predicted watermark message matrix. Size: 16×16.
[0061] Step 3: Key verification and reverse scrambling.
[0062] extract The first line is used as the prediction key. .calculate The Hamming distance to the real key K is used to determine the key verification. If the Hamming distance is less than 3, the key verification is considered successful and the watermark is valid; otherwise, the image is considered not to contain a valid watermark.
[0063] After key verification is successful, the scrambling matrices P and Q are restored based on key K. The remaining part is inversely scrambled to obtain the intermediate matrix: We obtain a 15×16 intermediate matrix. .
[0064] Step 4: Voting and Merging.
[0065] Because the watermark was embedded four times... The 16 columns can be divided into 4 groups, with each group of 4 columns corresponding to the same BCH-encoded watermark. For each bit position i (i = 0, 1, ..., 14), the predicted values in the 4 repetitions are counted, and the value that appears most frequently is selected: in, This is the majority voting function; i = 0, 1, ..., 14, corresponding to a 15-bit BCH-encoded watermark. Four sets of BCH-encoded watermarks are obtained through voting. Each group consists of 15 bits, for a total of 60 bits.
[0066] Step 5: BCH decoding.
[0067] Each of the four BCH-encoded watermarks is decoded using BCH(15,7), yielding a 7-bit original watermark from each group. The four original watermarks are then concatenated to obtain the final 28-bit copyright watermark information. .
[0068] (3.3) Experimental verification To verify the effectiveness of the method of the present invention, experiments were conducted under various distortion conditions.
[0069] Experimental setup: 1000 test images were randomly selected from a remote sensing database, and a random 28-bit watermark was embedded in each image. Test distortion included: Atmospheric scattering: , ; Sensor noise: Gaussian noise Salt and pepper noise ; Radiation correction: m ∈ [0.5, 1.5], b ∈ [-0.3, 0.3]; Geometric Transformation: Rotation Angle scaling factor .
[0070] Evaluation indicators: (1) Bit Accuracy: The ratio of the extracted watermark bits to the original watermark bits.
[0071] (2) Message Accuracy: The proportion of images that completely and correctly restore the original watermark information.
[0072] (3) PSNR and SSIM: Visual quality assessment of watermarked images compared to the original images.
[0073] Experimental results: Table 1: Watermark extraction performance under different distortion conditions Table 2: Comparison with existing methods (under conditions of comprehensive distortion) The experimental results show that: (1) The method of the present invention maintains a high watermark extraction accuracy under various distortion conditions, with a bit accuracy of over 95%. Even under combined distortion conditions (multiple distortions applied simultaneously), the bit accuracy still reaches 94.2%, and the message accuracy reaches 87.5%.
[0074] (2) Compared with existing image watermarking methods, the performance of this invention on remote sensing images is significantly better than traditional methods. Compared with the second-ranked LIM method, the bit accuracy is improved by 7.3 percentage points. This is due to the fact that the customized distortion layer of the remote sensing image accurately simulates the distortion characteristics unique to remote sensing images.
[0075] (3) This invention maintains good visual quality while maintaining high robustness. The PSNR reaches 36.8 dB and the SSIM reaches 0.95, which is better than or close to existing methods, indicating that the watermark embedding has little impact on the quality of remote sensing images.
[0076] (3.4) Ablation test To verify the contribution of each module in the customized distortion layer of remote sensing images, ablation experiments were conducted. Multiple watermark extraction networks were trained, and one module of the distortion layer was removed each time. The performance under real remote sensing distortion conditions was evaluated.
[0077] Table 3: Ablation Experiment Results (Bit Accuracy, %) The ablation experiment results show that each module of the distortion layer contributes to the final performance. Among them, the geometric distortion simulation module and the radiometric correction distortion simulation module have the greatest impact; removing these two modules results in the most significant performance degradation. This indicates that the geometric and radiometric correction processes of remote sensing images have the greatest impact on watermarking, and these distortions must be fully simulated during training.
[0078] (3.5) Parameter sensitivity analysis The watermark embedding strength parameters were analyzed. and Impact on performance. Fixed ,change ;fixed ,change .
[0079] Table 4: Results of parameter sensitivity analysis The parameter sensitivity analysis shows that, and Increasing the value improves watermark robustness but reduces visual quality (PSNR decreases). In practical applications, a balance needs to be struck between robustness and visual quality based on specific requirements. Recommended parameters for this invention. , It can maintain good visual quality ( At the same time, it achieves high robustness (bit accuracy > 94%).
[0080] Based on the same inventive concept, this embodiment also provides a template-embedded robust digital watermarking system for remote sensing images, including: The watermark encoding module is used to receive the remote sensing image to be protected and the copyright watermark information, and to perform BCH encoding, repeated embedding and matrix rearrangement and encryption operations on the copyright watermark information to generate a watermark message matrix. The template generation module is used to generate a Gaussian circular bit template based on the watermark message matrix. The watermark embedding module is used to overlay a watermark template onto the original remote sensing image with adjustable embedding strength parameters and template strength parameters to generate a watermarked remote sensing image. The remote sensing distortion simulation module is used to construct a customized distortion layer for remote sensing images. It takes a watermarked remote sensing image as input, simulates the unique distortions of the remote sensing image, and generates a distorted watermarked remote sensing image. The customized distortion layer for remote sensing images includes an atmospheric scattering simulation module, a sensor noise simulation module, a radiometric correction simulation module, and a geometric distortion simulation module. A watermark extraction network is used to extract the watermark message matrix from distorted remote sensing images. The watermark decoding module is used to reverse scramble, merge votes, and decode the extracted watermark message matrix to recover the final copyright watermark information.
[0081] For the system embodiments, since they basically correspond to the method embodiments, relevant details can be found in the descriptions of the method embodiments; the implementation methods of the remaining modules will not be repeated here. The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without any creative effort.
[0082] The system embodiments of the present invention can be applied to any device with data processing capabilities, such as a computer or other similar device. The system embodiments can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution.
[0083] The above-described embodiments are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. Those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.
Claims
1. A robust digital watermarking method for remote sensing images based on template embedding, characterized in that, It includes the watermark embedding stage, the distortion simulation stage, and the watermark extraction stage; The watermark embedding stage includes: receiving the remote sensing image to be protected and copyright watermark information; performing BCH encoding on the copyright watermark information to generate a watermark bit vector; repeatedly embedding and rearranging the watermark bit vector into a matrix, and encrypting it through a key-controlled scrambling matrix to generate a watermark message matrix; generating a corresponding Gaussian circular bit template based on the value of each bit in the watermark message matrix, and combining them into a watermark template with the same resolution as the original remote sensing image; and superimposing the watermark template onto the original remote sensing image with adjustable embedding strength parameters and template strength parameters to generate a watermarked remote sensing image. The distortion simulation stage includes: constructing a customized distortion layer for the remote sensing image, and generating a distorted watermarked remote sensing image by taking the watermarked remote sensing image as input. The customized distortion layer for remote sensing images includes: an atmospheric scattering simulation module for simulating the effects of atmospheric scattering on remote sensing images during atmospheric transmission; a sensor noise simulation module for simulating thermal noise, quantization noise, and stripe noise of remote sensing sensors; a radiometric correction distortion simulation module for simulating brightness and contrast adjustments during radiometric calibration and atmospheric correction processing; and a geometric distortion simulation module for simulating geometric transformations caused by geometric correction, registration, cropping, or rotation of remote sensing images. The watermark extraction stage includes: training a watermark extraction network using a watermarked remote sensing image, a distorted watermarked remote sensing image, and their corresponding watermark message matrix labels; segmenting the remote sensing image to be detected into a sequence of image blocks of the same size as the Gaussian circular bit template; inputting the image block sequence into the watermark extraction network to predict the watermark message matrix; performing inverse scrambling on the predicted watermark message matrix based on the key information, and fusing repeatedly embedded watermark information based on a voting mechanism to obtain a watermark bit vector; and performing BCH decoding on the watermark bit vector to recover the final copyright watermark information.
2. The robust digital watermarking method for remote sensing images based on template embedding according to claim 1, characterized in that, The process of generating the watermark message matrix during the watermark embedding stage includes: The watermark bit vector is repeatedly embedded to generate a robust message bit vector. ; Robust message bit vector Rearranged into a two-dimensional matrix of a preset size a×(a+1) ; The matrix is scrambled using a row scrambling matrix P and a column scrambling matrix Q determined by a key K. Encryption is performed to generate a watermarked message matrix: ； Where P is an a×a row scramble matrix, and Q is an (a+1)×(a+1) column scramble matrix. This means using the a+1 bit key K as the first row, and encrypting it with the a×(a+1) watermark matrix after row scrambling matrix P and column scrambling matrix Q. By vertically concatenating the data, a complete (a+1)×(a+1) watermark message matrix is generated. .
3. The robust digital watermarking method for remote sensing images based on template embedding according to claim 1, characterized in that, During the watermark embedding stage, the watermarked remote sensing image is represented as follows: ； in, Parameters for controlling the watermark embedding strength, Parameters for controlling the strength of the template itself, This is a watermarked remote sensing image. This is a watermark template. This is the original remote sensing image.
4. The robust digital watermarking method for remote sensing images based on template embedding according to claim 1, characterized in that, In the distortion simulation phase, the atmospheric scattering simulation module uses an atmospheric scattering physical model, adding random atmospheric illumination A and scattering coefficient t to the RGB channels respectively. The simulation formula is as follows: ； in, Let be the atmospheric transport rate randomly sampled from a uniform distribution U[0.7, 0.95]. The atmospheric illumination is randomly sampled from a uniform distribution U[0.5, 0.8]. These are the original pixel values. These are the pixel values after scattering.
5. The robust digital watermarking method for remote sensing images based on template embedding according to claim 1, characterized in that, During the distortion simulation phase, the noise simulated by the sensor noise simulation module includes: a standard deviation of... Randomly sampled Gaussian thermal noise and noise density within the range Salt-and-pepper quantization noise randomly sampled within the range, and strip noise randomly sampled within the range of [1,5].
6. The robust digital watermarking method for remote sensing images based on template embedding according to claim 1, characterized in that, In the distortion simulation phase, the simulation process of the radiation correction distortion simulation module includes: Add a random offset to each channel of the image, with the offset ranging from 1 to 2. ; Using linear transformations The contrast and brightness of the entire image, including the gain. offset , Indicates a uniform distribution; Adaptive adjustment is performed on different brightness areas of the image, enhancing low-brightness areas and suppressing high-brightness areas; The color changes during radiometric correction are simulated by randomly mixing RGB images with their grayscale versions.
7. The robust digital watermarking method for remote sensing images based on template embedding according to claim 1, characterized in that, In the distortion simulation stage, the geometric transformations simulated by the geometric distortion simulation module include: random perspective transformation with a four-corner perturbation range of ±25 pixels, random rotation within an angle range of [-15, 15], random scaling within a scaling factor range of [0.85, 1.15], and random cropping within a cropping ratio range of [0.8, 0.95]. The blank areas generated after rotation are filled with boundary pixel values, and after random scaling and cropping, they are adjusted back to their original size.
8. The robust digital watermarking method for remote sensing images based on template embedding according to claim 1, characterized in that, The watermark extraction network adopts the ResNet18 architecture, and the loss function is: ； in, For real watermarked message matrix tags, To predict the watermark message matrix from a watermarked remote sensing image, The watermark message matrix is predicted from the distorted watermarked remote sensing image. This is the weighting coefficient, with a value of 2.
9. The robust digital watermarking method for remote sensing images based on template embedding according to claim 2, characterized in that, In the watermark extraction stage, the predicted watermark message matrix is inversely scrambled based on the key information, and the repeatedly embedded watermark information is fused based on a voting mechanism to obtain the watermark bit vector, including: Extract the first row vector from the predicted watermark message matrix as the prediction key; calculate the Hamming distance between the prediction key and the original key used in the watermark embedding stage; if the Hamming distance is less than the preset fault tolerance threshold, the key verification is passed and the watermark is valid; otherwise, the image is determined not to contain a valid watermark. After verifying the validity of the watermark, the corresponding scrambling matrices P and Q are recovered based on the original key. Then, the remaining part of the watermark message matrix except for the first row is decrypted to obtain an intermediate watermark matrix containing multiple repeated embeddings. The intermediate watermark matrix is divided into multiple groups along the column direction, with each group having the same number of columns, and each group corresponds to a repetition of the watermark information. For each bit position in the original watermark information, the corresponding predicted value is extracted from each of the multiple groups. The extracted predicted values are processed by a majority voting function, and the bit value that appears most frequently is determined as the final value of that bit position. All bit positions are traversed to generate the watermark bit vector.
10. A robust digital watermarking system for remote sensing images based on template embedding, used to implement the watermarking method of claim 1, characterized in that, The system includes: The watermark encoding module is used to receive the remote sensing image to be protected and the copyright watermark information, and to perform BCH encoding, repeated embedding and matrix rearrangement and encryption operations on the copyright watermark information to generate a watermark message matrix. The template generation module is used to generate a Gaussian circular bit template based on the watermark message matrix. The watermark embedding module is used to overlay a watermark template onto the original remote sensing image with adjustable embedding strength parameters and template strength parameters to generate a watermarked remote sensing image. The remote sensing distortion simulation module is used to construct a customized distortion layer for remote sensing images. It takes a watermarked remote sensing image as input, simulates the unique distortions of the remote sensing image, and generates a distorted watermarked remote sensing image. The customized distortion layer for remote sensing images includes an atmospheric scattering simulation module, a sensor noise simulation module, a radiometric correction simulation module, and a geometric distortion simulation module. A watermark extraction network is used to extract the watermark message matrix from distorted remote sensing images. The watermark decoding module is used to reverse scramble, merge votes, and decode the extracted watermark message matrix to recover the final copyright watermark information.