Steganographic embedding and image restoration method and device based on remote sensing image sensitive target

Abstract

The present application provides a steganographic embedding and image recovery method and device based on remote sensing image sensitive targets, which relates to the technical field of remote sensing image processing. The present application obtains an original remote sensing image and extracts a sensitive target bounding box, initializes a single-channel all-zero matrix, generates a global binary union mask and a carrier image; then performs non-scaling mapping and packaging on the sensitive target to generate a secret map and a metadata file; a fixed decoder is initialized using a random seed key; texture block selection based on local binary pattern is performed, a secret-containing image is generated through joint loss optimization, and a segmented linear warming schedule is introduced to balance performance; residual disturbance is extracted at the decryption end and analyzed by the decoder, and the target is seamlessly pasted back in place in combination with the metadata file. The present application can avoid scaling distortion, maintain the visual quality of the carrier, realize high-fidelity hiding of multi-scale sensitive targets, pixel-level accurate recovery, and scene-level in-place restoration, and has high security and strong robustness.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing image processing and information security technology, and more specifically, to a method and apparatus for steganography embedding and image restoration based on sensitive targets in remote sensing images. Background Technology

[0002] In the process of publicly sharing and distributing remote sensing data, in order to meet compliance requirements and protect national geographic information security, it is essential to effectively conceal sensitive targets such as military facilities and airports in the imagery. Simultaneously, to ensure that the authorizing party can perform refined decoding and high-precision geographic analysis for specific tasks, these concealed targets must be accurately restored to the original scene when necessary. Currently, existing sensitive target protection technologies mainly employ a scheme combining image inpainting and transform domain digital watermarking. This involves first using background inpainting algorithms to fill sensitive areas to generate pseudo-carriers, and then embedding the sensitive targets as watermark information using traditional algorithms such as discrete wavelet transform.

[0003] However, existing technical solutions have significant shortcomings in practical applications. Firstly, there are issues with multi-target management and capacity bottlenecks. Since remote sensing scenes typically contain multiple sensitive targets with vastly different scales, traditional watermarking schemes involve extremely cumbersome processes for embedding each target individually, easily leading to capacity overflow. Furthermore, they lack a unified spatial location management mechanism to ensure target consistency. Secondly, traditional watermarking technologies heavily rely on fixed frequency domain rules, making them poorly adaptable to the complex and varied textures of remote sensing images, easily producing obvious visual artifacts and resulting in insufficient concealment. In addition, the key and access control mechanisms of existing steganography systems are relatively crude, often relying solely on the non-public nature of the algorithm itself or static traditional cryptographic methods. They lack dynamic key distribution strategies that are deeply integrated with neural network extraction mechanisms and can be flexibly scheduled, making it difficult to cope with complex security access requirements.

[0004] In view of the above, this application is hereby submitted. Summary of the Invention

[0005] This invention aims to provide a method and apparatus for steganography embedding and image restoration based on sensitive targets in remote sensing images, in order to solve the defects in existing remote sensing image sensitive target hiding technologies, such as complex multi-target management logic, limited embedding capacity, insufficient concealment of frequency domain watermarking algorithms, irreversible distortion of micro-texture caused by scaling operations, and low security of key management mechanisms.

[0006] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution:

[0007] A method for steganalysis and image restoration based on sensitive targets in remote sensing images includes: S1, acquire the original remote sensing image and initialize a single-channel all-zero matrix with the same resolution as the original remote sensing image; S2, extract the target bounding box set of the sensitive targets to be protected in the original remote sensing image, map the pixel region covered by each target bounding box to the activation value on a single-channel all-zero matrix, generate a global binary union mask, and output a carrier image with the same size as the original remote sensing image and without sensitive targets based on the global binary union mask; S3, the original remote sensing image is cropped 1:1 without distortion according to the target bounding box to obtain a set of sensitive target pixel blocks, which are packaged into a secret map of a fixed size without scaling, and a metadata file for recording the location information of all sensitive targets and an effective region mask for identifying the location of sensitive target pixels and the boundary of the zero-filled area are generated. S4. Configure a random seed key, initialize a preset neural network with steganographic feature mapping function using the random seed key, and freeze all weights after initialization to build a fixed decoder. S5, the carrier image is divided into several non-overlapping image blocks, the texture complexity of each image block is calculated and sorted, a set of high texture blocks is selected and spliced ​​to obtain a compact domain carrier and the corresponding compact domain perturbation, and at the same time, the image blocks of the compact domain perturbation are written back to the zero-base map with the same resolution as the carrier image using the scattering operator to generate a global sparse perturbation. S6, the global sparse perturbation is superimposed on the carrier image, threshold truncation is performed to obtain a floating-point dense image, and the dense image is output through loss optimization by combining the secret map and the effective region mask; S7. Calculate the pixel difference between the secret image and the carrier image, separate the residual perturbation signal, process it through the extraction and rearrangement operator, input it into the fixed decoder to obtain the recovered secret map, and combine it with the metadata file to restore the full-image scene-level recovered image.

[0008] Preferably, when generating the global binary union mask, the coordinates of each target bounding box in the target bounding box set are traversed, and the pixel coordinates covered by the target bounding box are assigned pixel values ​​on the all-zero matrix. The pixel values ​​of the corresponding regions are set from 0 to 1, forming an independent local binary mask, expressed as: ; in, For the first Local binary mask of the bounding box of each target; These are the row coordinates of the image pixels; These are the column coordinates of the image pixels; , For the first The top-left pixel coordinates of the bounding box of each target; , For the first The bottom right pixel coordinates of the target bounding box; Traverse the local binary masks of all target bounding boxes, perform logical OR merge operations pixel by pixel, perform spatial fusion, and remove overlapping and redundant regions to generate a global binary union mask with the same resolution and single channel as the original remote sensing image.

[0009] Preferably, when the global binary union mask outputs a carrier image with the same size as the original remote sensing image and without sensitive targets, a background inpainting algorithm combined with the global binary union mask is used to fill and inpaint the sensitive target regions in the original remote sensing image, generating an inpainted carrier image without sensitive targets; wherein, the background inpainting algorithm adopts the PatchMatch inpainting algorithm based on image patch matching or the Telea inpainting algorithm based on partial differential equation propagation.

[0010] Preferably, the secret map is obtained by placing each sensitive target pixel block of the sensitive target pixel block set sequentially onto a blank canvas of a fixed size according to a preset grid layout rule without scaling; for grid cells not completely filled by sensitive target pixel blocks, a zero-value filling operation is performed, thereby obtaining the secret map, the expression of which is: ; ; ; in, For secret maps; The global pixel coordinates of the secret graph canvas; For image channels; For the first One sensitive target pixel block; , for Local pixel coordinates; For the first A preset grid area in the secret map for each sensitive target pixel block; The total number of sensitive target pixel blocks; This represents the union of all sensitive target pixel blocks within the grid region of the map; Represents the union; , for The starting coordinates are at the top left corner of the graph.

[0011] Preferably, when configuring the random seed key, two configuration modes are included: Mode 1: Generate a unique random seed key for each image, which is used to encrypt the metadata file before sending; during decryption, receive the encrypted metadata file, extract the random seed key, and initialize the local decoder; Mode 2: When configuring a random seed key, the encryption and decryption ends are pre-configured with the same static shared random seed key to initialize the decoder, thereby improving transmission efficiency when transmitting the metadata file.

[0012] Preferably, S5 specifically includes: The carrier image is divided into several non-overlapping image blocks; Calculate the texture complexity based on the local binary pattern histogram entropy for each image patch; Sort by texture complexity from high to low and select a set of high texture blocks that meet a preset threshold. The high-texture blocks in the high-texture block set are sequentially spliced ​​into a compact mesh domain to obtain a compact domain carrier and the corresponding compact domain perturbation. Using a scattering operator, each image patch of the compact domain perturbation is written back to the zero-base map with the same resolution as the carrier map according to its coordinate position in the set of high-texture patches, thereby generating a global sparse perturbation.

[0013] Preferably, the formula for performing threshold truncation is: ; in, For floating-point dense images; For carrier diagram; For global sparse perturbation; This is a threshold truncation function that restricts the input values ​​to a range. Inside; , These are the upper and lower limits of the set cutoff threshold, respectively.

[0014] Preferably, when combining the secret map and the effective region mask to output a dense image through loss optimization, a joint loss function is used for iterative optimization; The joint loss function includes hidden distortion loss, net load recovery loss, robust recovery loss, and total variational smoothing regularization term; The concealment distortion loss is calculated by measuring the mean square error between the floating-point dense image and the carrier image. The net load recovery loss is combined with the effective region mask to calculate the recovered secret map and the mean square error of the secret map within the effective region; The robust recovery loss is calculated by applying a simulated channel attack operator to the secret image and then calculating the mean square error between the recovered map and the secret map within the effective region under attack conditions. The total variational smoothing regularization term calculates the total variational loss by applying global sparse perturbations; The expression for the joint loss function is: ; ; ;

[0015] ; in, For the joint loss function; This is due to concealed distortion loss; This is the net load recovery loss; Weighted for robust recovery losses; For total variational smoothing regularization; The weights for dynamic concealment loss vary with each iteration. change; , Fixed weighting coefficients; The weighting factor for net load recovery loss; For floating-point dense images; For carrier diagram; Calculation of mean square error; For secret maps; To recover the secret map; For effective area masking; For fixed decoders; For extracting rearrangement operators; To simulate channel attack operators; This is the total variation loss, used to calculate the mean of the absolute differences in the perturbation amplitudes between adjacent pixels; For global sparse perturbation; The dynamic concealment loss weight is dynamically adjusted through a piecewise linear temperature ramping scheduling mechanism, and its expression is as follows: ; in, This refers to the current iteration round; These are the initial baseline weights; For preheating cycles; The maximum upper limit weight; This represents the total number of iterations.

[0016] Preferably, S7 specifically includes: The pixel difference between the secret image and the carrier image is calculated to separate the residual perturbation signal. This residual perturbation signal is then processed by a decimation-rearrangement operator and input into a fixed decoder to obtain the recovered secret map, expressed as: ; in, To recover the secret map; For fixed decoders; For extracting rearrangement operators; For dense images; For carrier diagram; A collection of high-texture blocks; Based on the recovered secret map and the metadata file, zero-filled pixels are stripped, and the sensitive target pixel blocks of the original pixels are restored at a 1:1 ratio without any reverse scaling. The restored sensitive target pixel blocks are pasted back to their corresponding positions in the carrier image according to the coordinate information in the metadata file to obtain a full-scene-level restored image.

[0017] The present invention also provides a steganography embedding and image restoration device based on sensitive targets in remote sensing images, comprising: The image acquisition and initialization unit is used to acquire the original remote sensing image and initialize a single-channel all-zero matrix with the same resolution as the original remote sensing image. The binary mask unit is used to extract the set of target bounding boxes of the sensitive targets to be protected in the original remote sensing image, map the pixel region covered by each target bounding box to the activation value on a single-channel all-zero matrix, generate a global binary union mask, and output a carrier image with the same size as the original remote sensing image and without sensitive targets based on the global binary union mask. The atlas unit is used to perform 1:1 distortion-free cropping of the original remote sensing image according to the target bounding box to obtain a set of sensitive target pixel blocks, and to package them into a secret atlas of a fixed size without scaling, thereby generating a metadata file for recording the location information of all sensitive targets and an effective region mask for identifying the location of sensitive target pixels and the boundary of zero-filled regions. A key configuration unit is used to configure a random seed key, initialize a preset neural network with steganalysis feature mapping function using the random seed key, and freeze all weights after initialization to build a fixed decoder. The global perturbation unit is used to divide the carrier image into several non-overlapping image blocks, calculate the texture complexity of each image block and sort them, select the set of high texture blocks and stitch them together to obtain the compact domain carrier and the corresponding compact domain perturbation. At the same time, the scattering operator is used to write each image block of the compact domain perturbation back to the zero-base map with the same resolution as the carrier image to generate global sparse perturbation. The steganography embedding unit is used to superimpose the global sparse perturbation onto the carrier map, perform threshold truncation to obtain a floating-point dense image, and combine the secret map with the effective region mask to output the dense image through loss optimization. The restoration unit is used to calculate the pixel difference between the secret image and the carrier image, separate the residual perturbation signal, process it through the extraction and rearrangement operator, input it into the fixed decoder to obtain the restored secret map, and combine it with the metadata file to restore the full-image scene-level restored image.

[0018] The present invention also provides a steganography embedding and image restoration device based on sensitive targets in remote sensing images, including a processor and a memory. The memory stores a computer program that can be executed by the processor to implement the steganography embedding and image restoration method based on sensitive targets in remote sensing images as described above.

[0019] The present invention also provides a computer-readable storage medium storing computer-readable instructions, which, when executed by a processor of the device on which the computer-readable storage medium is located, implement the steganalysis embedding and image restoration method based on sensitive targets in remote sensing images as described above.

[0020] In summary, compared with the prior art, the present invention has the following beneficial effects: First, this invention achieves extremely high fidelity in image restoration. By employing a 1:1 scale-free, zero-fill atlasing mechanism, it completely avoids the irreversible distortion of microscopic textures in remote sensing images caused by scaling operations in traditional steganography techniques. This mechanism ensures that the recovered sensitive targets retain their original pixel-level features, thereby meeting the stringent requirements of subsequent high-precision interpretation, target identification, and measurement analysis.

[0021] Secondly, this invention possesses excellent visual concealment. By introducing a texture selection mechanism based on local binary patterns, the system forcibly constrains steganalytic perturbations to visually insensitive high-texture regions of the carrier image, effectively avoiding artificial artifacts in smooth areas such as water surfaces and grass. Simultaneously, combined with a piecewise linear heating scheduling strategy, the concealment constraint is strengthened in the later stages of optimization, resulting in a final dense image that achieves extremely high visual consistency with the original carrier.

[0022] Third, this invention provides efficient multi-target management capabilities. By uniformly encoding multiple sensitive targets with vastly different scales into a single secret map, it achieves the function of "one-time embedding, simultaneous recovery of multiple targets." This map-based management method simplifies the steganography process, reduces the engineering complexity of embedding different targets separately, and utilizes metadata files to achieve precise control over geospatial locations.

[0023] Fourth, this invention constructs a lightweight yet highly secure neural cryptography system. By equating the security of a massive neural network parameter to an extremely simple random seed key, it achieves "one..." Figure 1 The design employs a highly robust security management system. The fixed decoder design allows the receiving end to reconstruct the extraction environment simply by possessing the correct seed key, ensuring the one-way extraction security of the neural steganography model while significantly reducing the system overhead of key distribution and storage.

[0024] Fifth, this invention significantly enhances channel robustness. The optimization process incorporates a simulated channel attack operator and a mask-guided robust recovery loss, enabling the encrypted image to still be parsed into a clear secret map using a fixed decoder even after experiencing real-world transmission channel interference such as JPEG compression, Gaussian noise, or salt-and-pepper noise. The mask-guided mechanism further encourages the model to focus its limited embedding capacity on the true target pixels, improving the accuracy of payload extraction.

[0025] Sixth, this invention achieves scene-level in-situ accurate restoration. Utilizing the geographic association information recorded in the metadata file, the system can accurately re-attach the restored target tiles to their original geographic coordinates. Combined with feathering or Poisson blending techniques, it eliminates stitching marks between the target edges and the background, achieving a visually seamless transition and providing the licensor with complete scene-level restored images suitable for practical analysis. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0027] Figure 1 This is a flowchart illustrating a method for steganography embedding and image restoration based on sensitive targets in remote sensing images, as provided in Example 1.

[0028] Figure 2 This is an overall architecture diagram of a method for steganography embedding and image restoration based on sensitive targets in remote sensing images, provided in Example 1.

[0029] Figure 3 This is a schematic diagram of a steganography embedding and image restoration device based on sensitive targets in remote sensing images, provided in Embodiment 2.

[0030] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0032] Example 1 Embodiment 1 of the present invention provides a method for steganalysis and image restoration based on sensitive targets in remote sensing images, which can be implemented by a steganalysis and image restoration device based on sensitive targets in remote sensing images (hereinafter, the embedding and restoration device), specifically, executed by one or more processors within the embedding and restoration device.

[0033] In this embodiment, the stegume embedding and image restoration device can be an electronic device equipped with a processor. The processor carries a computer program for the stegume embedding and image restoration method based on remote sensing image sensitive targets, and the computer program can be executed. For example, a computer, smartphone, smart tablet, workstation, etc. are not limited here.

[0034] like Figures 1-2 As shown, a method for steganography embedding and image restoration based on sensitive targets in remote sensing images includes steps S1 to S7.

[0035] S1, acquire the original remote sensing image and initialize a single-channel all-zero matrix with the same resolution as the original remote sensing image.

[0036] In this embodiment, a raw remote sensing image to be processed is acquired through a remote sensing imaging device, a satellite data interface, or a storage medium. This image contains sensitive targets to be protected (such as military facilities, key buildings, classified areas, etc.).

[0037] In this embodiment, the single-channel all-zero matrix is ​​a matrix with the same spatial resolution and size as the original remote sensing image. As a mask base, it is used to ensure that there are no blind spots in the hiding process and to avoid size deviations in subsequent image restoration and perturbation generation.

[0038] S2, extract the target bounding box set of the sensitive targets to be protected in the original remote sensing image, map the pixel region covered by each target bounding box to the activation value on a single-channel all-zero matrix, generate a global binary union mask, and output a carrier image with the same size as the original remote sensing image and without sensitive targets based on the global binary union mask.

[0039] When extracting the target bounding box set of sensitive targets to be protected in the original remote sensing image, the bounding boxes of all sensitive targets to be protected in the original remote sensing image can be obtained by target detection algorithms (such as YOLOv8 remote sensing version, Faster R-CNN, etc.) or manual annotation, forming a target bounding box set.

[0040] When generating the global binary union mask, the coordinates of each target bounding box in the target bounding box set are traversed. The pixel coordinates covered by the target bounding box are assigned pixel values ​​on the all-zero matrix, and the pixel values ​​of the corresponding regions are set from 0 to 1, forming an independent local binary mask. The expression is: ; in, For the first Local binary mask of the bounding box of each target; These are the row coordinates of the image pixels; These are the column coordinates of the image pixels; , For the first The top-left pixel coordinates of the bounding box of each target; , For the first The bottom right pixel coordinates of the target bounding box; Traverse the local binary masks of all target bounding boxes, perform logical OR merging operations pixel by pixel to perform spatial fusion, and remove overlapping and redundant regions to generate a global binary union mask with the same resolution and single channel as the original remote sensing image.

[0041] Then, the background restoration algorithm, combined with the global binary union mask, is used to fill and restore the sensitive target areas in the original remote sensing image. By extracting the non-sensitive background texture features around the sensitive target areas, the algorithm performs pixel filling and edge transition processing on the holes covered by the global binary union mask to generate a restored carrier image without sensitive targets. The background restoration algorithm adopts either the PatchMatch restoration algorithm based on image patch matching or the Telea restoration algorithm based on partial differential equation propagation.

[0042] As a public background map, the visual naturalness of this carrier image directly determines the concealment of the subsequent steganographic carrier.

[0043] This step uses bounding box extraction and mask generation to accurately mark the spatial location of all sensitive targets, providing clear regional guidance for hiding sensitive targets and restoring the background. The carrier image serves as the carrier for steganographic embedding (which can be publicly transmitted), achieving the separation of sensitive targets from the background and providing a base image without sensitive information for subsequent steganographic operations.

[0044] S3, the original remote sensing image is cropped 1:1 without distortion according to the target bounding box to obtain a set of sensitive target pixel blocks, which are then packaged into a secret map of a fixed size without scaling, generating a metadata file for recording the location information of all sensitive targets and an effective region mask for identifying the location of sensitive target pixels and the boundary of the zero-filled area.

[0045] In this step, each sensitive target in the original remote sensing image is cropped at its original pixel ratio to obtain multiple sensitive target pixel blocks. This process strictly adheres to 1:1 distortion-free sampling, eliminating any form of interpolation or scaling to preserve the most essential microscopic textures and geometric features of the remote sensing image. Subsequently, a scaling-free atlasing and packaging process is initiated.

[0046] The secret map is obtained by placing each sensitive target pixel block of the sensitive target pixel block set sequentially onto a blank canvas of a fixed size (e.g., 128×128 pixels) according to a preset grid layout rule without scaling; for grid cells not completely filled by sensitive target pixel blocks, a zero-value filling operation is performed, thereby obtaining the secret map, the expression of which is: ; ; ; in, For secret maps; The global pixel coordinates of the secret graph canvas; For image channels; For the first One sensitive target pixel block; , for Local pixel coordinates; For the first A preset grid area in the secret map for each sensitive target pixel block; The total number of sensitive target pixel blocks; This represents the union of all sensitive target pixel blocks within the grid region of the map; , for The starting coordinates are at the top left corner of the graph.

[0047] During this process, a structured metadata file is generated synchronously. This file records in detail, in binary or JSON format, the original geographic coordinates of each sub-target in the secret map, its grid index in the secret map, and its specific zero-padding offset. In addition, by comparing the filling relationship between sensitive target pixel blocks and the canvas, an effective region mask is generated. This mask is used to accurately identify the position of real target pixels in the secret map in subsequent optimizations, filtering out meaningless zero-padding regions.

[0048] S4. Configure a random seed key, initialize a preset neural network with steganography feature mapping function using the random seed key, and freeze all weights after initialization to build a fixed decoder.

[0049] Next, the decoder is built. The system configures a random seed key according to the authorization level, supporting dynamic key mode (i.e., mode one) and static shared key mode (i.e., mode two).

[0050] Mode 1: Generate a unique random seed key for each image, which is used to encrypt the metadata file and send it along with the encrypted image; during decryption, receive the encrypted metadata file, extract the random seed key, and initialize the local decoder. Mode 2: When configuring a random seed key, the encryption and decryption ends are pre-configured with the same static shared random seed key to initialize the decoder, thereby improving transmission efficiency when transmitting the metadata file.

[0051] The random seed key is used to initialize a pre-defined neural network architecture with steganalytic feature mapping capabilities. This neural network architecture may contain multi-layer convolutional neural networks and residual connection modules to achieve a non-linear mapping from image residuals to secret payloads. For example, the pre-defined neural network architecture decoder consists of 5 groups of residual blocks, each containing two 3x3 convolutional layers with a fixed number of channels of 64, using the ReLU activation function, and finally outputting a 3-channel map through a 1x1 convolutional layer.

[0052] After initialization, all network weights of the fixed decoder are completely frozen, so that the parameters remain constant in subsequent gradient iterations, thereby transforming the complex neural network model into a one-way extraction key controlled by a random seed.

[0053] S5, the carrier image is divided into several non-overlapping image blocks, the texture complexity of each image block is calculated and sorted, a set of high texture blocks is selected and spliced ​​to obtain a compact domain carrier and the corresponding compact domain perturbation, and at the same time, the image blocks of the compact domain perturbation are written back to the zero-base map with the same resolution as the carrier image using the scattering operator to generate a global sparse perturbation.

[0054] This step aims to implant the secret map into the carrier map in the form of imperceptible, subtle perturbations.

[0055] Specifically, the carrier image is divided into several non-overlapping image blocks (e.g., into several 8×8 image blocks). For each image patch, calculate the texture complexity based on Local Binary Pattern Histogram Entropy (LBP); Sort by texture complexity from high to low and select a set of high texture blocks that meet a preset threshold. Subsequently, the extraction and rearrangement operator can be invoked to rearrange the high-texture blocks scattered throughout the entire image according to the index order and stitch them together into a continuous, compact tensor structure, thus obtaining the compact domain carrier and the corresponding compact domain perturbation.

[0056] This mechanism forces steganographic perturbations to act only on visually insensitive complex texture areas, effectively avoiding the risk of visual artifacts in smooth background areas.

[0057] Finally, the scattering operator is used to write each image patch of the compact domain perturbation back to the zero-base map with the same resolution as the carrier map according to the coordinate position in the set of high-texture patches, thus generating a global sparse perturbation.

[0058] In this embodiment, the scattering operator is the inverse mapping of the extraction rearrangement operator, which backfills the compact domain perturbation block to the corresponding index position and fills the remaining positions with zeros.

[0059] The extraction and rearrangement design allows the fixed decoder to process only the high-density information grid after dimensionality reduction, completely avoiding the waste of computing power and memory overflow caused by ineffective background smoothing in remote sensing images.

[0060] S6, the global sparse perturbation is superimposed on the carrier image, threshold truncation is performed to obtain a floating-point dense image, and the dense image is output through loss optimization by combining the secret map and the effective region mask.

[0061] This step overlays a global sparse perturbation onto the carrier image and forces the pixel values ​​to be limited to a legal floating-point range using a hyperbolic tangent function or threshold truncation operation, thus obtaining a floating-point dense image.

[0062] The formula for performing threshold truncation can be set as follows: ; in, For floating-point dense images; For carrier diagram; For global sparse perturbation; This is a threshold truncation function that restricts the input values ​​to a range. Inside; , These are the upper and lower limits of the set cutoff threshold, respectively. For example, set to... , .

[0063] During the optimization iteration process, the update of the perturbation variables is driven by constructing a multi-level joint loss function.

[0064] The joint loss function includes hidden distortion loss, net load recovery loss, robust recovery loss, and total variational smoothing regularization term; The concealment distortion loss is calculated by measuring the mean square error between the floating-point dense image and the carrier image. The net load recovery loss is combined with the effective region mask to calculate the recovered secret map and the mean square error of the secret map within the effective region, so as to force the model to focus the embedding energy on the real target pixel; The robust recovery loss is calculated by applying a simulated channel attack operator (such as JPEG compression, Gaussian noise, and salt-and-pepper noise) to the dense image, and then calculating the mean square error between the recovered map and the secret map within the effective region under the attack condition. The total variational smoothing regularization term improves the spatial smoothness of the perturbation by calculating the total variational loss on the global sparse perturbation.

[0065] The expression for the joint loss function is: ; ; ;

[0066] ; in, For the joint loss function; This is due to concealed distortion loss; This is the net load recovery loss; Weighted for robust recovery losses; For total variational smoothing regularization; The weights for dynamic concealment loss vary with each iteration. change; , Fixed weighting coefficients; The weighting factor for net load recovery loss; For floating-point dense images; For carrier diagram; This is for calculating the mean squared error, used to measure the average squared error between two sets of data. For secret maps; To recover the secret map; For effective area masking; For fixed decoders; For extracting rearrangement operators; To simulate channel attack operators; This is the total variation loss, used to calculate the mean of the absolute differences in the perturbation amplitudes between adjacent pixels; This represents a global sparse perturbation.

[0067] To resolve the conflict between recovery accuracy and concealment constraints in the early stages of optimization, the system sets the concealment loss weight to change dynamically with each iteration.

[0068] The dynamic concealment loss weight is dynamically adjusted through a piecewise linear temperature ramping scheduling mechanism, and its expression is as follows: ; in, This refers to the current iteration round; These are the initial baseline weights; For preheating cycles; The maximum upper limit weight; This represents the total number of iterations.

[0069] In a preferred embodiment The value is 0.18. The value is 0.03. The value is 0.5. Set it to 20% of the total rounds to ensure that the recovery logic is learned first in the early stages.

[0070] During the pre-set warm-up phase, the concealment weights are kept at a low baseline value, allowing the perturbation to have a large search space for rapid convergence and accuracy recovery. After the warm-up phase, the weights increase linearly with each round to a maximum upper limit, forcing the perturbation to evolve in an imperceptible direction. After completing the pre-set number of iterations, the floating-point dense image is subjected to 8-bit quantization processing, outputting the final dense image for public release.

[0071] S7. Calculate the pixel difference between the secret image and the carrier image, separate the residual perturbation signal, process it through the extraction and rearrangement operator, input it into the fixed decoder to obtain the recovered secret map, and combine it with the metadata file to restore the full-image scene-level recovered image.

[0072] After acquiring the encrypted image, carrier image, and metadata file, the receiving end first reconstructs the fixed decoder using the matched random seed key.

[0073] The pixel difference between the secret image and the carrier image is calculated to separate the residual perturbation signal from the complex background. This residual perturbation signal is then processed by a decimation-rearrangement operator and input into a fixed decoder to obtain the recovered secret map, expressed as: ; in, To recover the secret map; For fixed decoders; For extracting rearrangement operators; For dense images; For carrier diagram; A collection of high-texture blocks.

[0074] The fixed decoder recovers the secret map from the weak residual signal through a single forward propagation calculation.

[0075] Based on the recovered secret map and the metadata file, zero-filled pixels are stripped, and the sensitive target pixel blocks of the original pixels are restored at a 1:1 ratio without any reverse scaling. The restored sensitive target pixel blocks are pasted back to their corresponding positions in the carrier image according to the coordinate information in the metadata file to obtain a full-scene-level restored image.

[0076] After the coordinate information is pasted back, in order to eliminate the stitching gap between the target edge and the background carrier, feathering blending technology or Poisson blending technology based on gradient field reconstruction can be used for edge blending. Feathering blending creates a smooth visual transition between the target and the background by setting an adjustable transparency gradient band at the target boundary; Poisson blending, on the other hand, solves the Poisson equation to deeply blend the detailed texture of the target with the lighting tone of the background.

[0077] Ultimately, the system outputs a full-scene restored image with pixel-level restoration accuracy, natural edges, and accurate geographic coordinates, providing reliable data support for high-precision remote sensing interpretation.

[0078] In a specific engineering application of this embodiment, the carrier image size is set to 512×512 pixels, and the secret map packaging size is set to 128×128 pixels. During the steganography embedding process, the perturbation truncation amplitude parameter is set to 0.20 to maximize the recovery quality while ensuring concealment. The total number of iterations for optimization is configured to be 1500, with the first 750 iterations being a weight warm-up phase. The concealment loss weight increases linearly from the initial 0.25 to 1.80. In the simulated channel attack configuration, the JPEG compression quality factor is randomly distributed between 60 and 90, and the standard deviation of additive Gaussian noise is set to 0.0002. With this parameter configuration, when processing complex urban remote sensing images containing multiple targets, the peak signal-to-noise ratio (PSNR) of the generated concealed image and the original carrier remains above 45dB, and the average structural similarity (SSIM) between the recovered sensitive target and the original pixels reaches above 0.98, fully verifying the technical advantage of this invention in achieving high-fidelity restoration while maintaining high concealment.

[0079] In summary, compared with the prior art, the present invention has the following beneficial effects: This invention solves the challenge of multi-scale management of remote sensing targets by utilizing a scaling-free map-based packaging mechanism and eliminates interpolation distortion. The combination of texture block selection based on Local Binary Pattern Histogram Entropy (LBP) and extraction / rearrangement operators significantly reduces the computational overhead of the neural network while precisely guiding steganographic energy to complex regions of the image, greatly improving visual concealment. The mask-guided joint loss function and piecewise linear temperature ramping scheduling strategy enable the system to achieve extremely high information extraction accuracy within a limited embedding capacity. A fixed decoder environment driven by random seeds constructs a lightweight and highly secure cryptographic barrier, ensuring the controlled use of sensitive information. The in-situ seamless fusion technology provided during scene restoration transforms the hidden information back into a complete scene image with practical value.

[0080] This invention can not only be applied to the protection of sensitive targets in the field of remote sensing, but can also be extended to fields with strict requirements for image fidelity, such as medical image privacy desensitization and secure transmission of industrial visual data.

[0081] Example 2 like Figure 3 As shown, the second embodiment of the present invention also provides a steganography embedding and image restoration device based on remote sensing image sensitive targets, comprising: The image acquisition and initialization unit is used to acquire the original remote sensing image and initialize a single-channel all-zero matrix with the same resolution as the original remote sensing image. The binary mask unit is used to extract the set of target bounding boxes of the sensitive targets to be protected in the original remote sensing image, map the pixel region covered by each target bounding box to the activation value on a single-channel all-zero matrix, generate a global binary union mask, and output a carrier image with the same size as the original remote sensing image and without sensitive targets based on the global binary union mask. The atlas unit is used to perform 1:1 distortion-free cropping of the original remote sensing image according to the target bounding box to obtain a set of sensitive target pixel blocks, and to package them into a secret atlas of a fixed size without scaling, thereby generating a metadata file for recording the location information of all sensitive targets and an effective region mask for identifying the location of sensitive target pixels and the boundary of zero-filled regions. A key configuration unit is used to configure a random seed key, initialize a preset neural network with steganalysis feature mapping function using the random seed key, and freeze all weights after initialization to build a fixed decoder. The global perturbation unit is used to divide the carrier image into several non-overlapping image blocks, calculate the texture complexity of each image block and sort them, select the set of high texture blocks and stitch them together to obtain the compact domain carrier and the corresponding compact domain perturbation. At the same time, the scattering operator is used to write each image block of the compact domain perturbation back to the zero-base map with the same resolution as the carrier image to generate global sparse perturbation. The steganography embedding unit is used to superimpose the global sparse perturbation onto the carrier map, perform threshold truncation to obtain a floating-point dense image, and combine the secret map with the effective region mask to output the dense image through loss optimization. The restoration unit is used to calculate the pixel difference between the secret image and the carrier image, separate the residual perturbation signal, process it through the extraction and rearrangement operator, input it into the fixed decoder to obtain the restored secret map, and combine it with the metadata file to restore the full-image scene-level restored image.

[0082] Example 3 The third embodiment of the present invention also provides a steganography embedding and image restoration device based on sensitive targets in remote sensing images, which includes a memory and a processor. The memory stores a computer program, which can be executed by the processor to realize the steganography embedding and image restoration method based on sensitive targets in remote sensing images as described above.

[0083] Example 4 The fourth embodiment of the present invention also provides a computer-readable storage medium storing computer-readable instructions. When the computer-readable instructions are executed by the processor of the device where the computer-readable storage medium is located, they implement the above-described method for steganography embedding and image restoration based on sensitive targets in remote sensing images.

[0084] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for steganalysis and image restoration based on sensitive targets in remote sensing images, characterized in that, include: S1, acquire the original remote sensing image and initialize a single-channel all-zero matrix with the same resolution as the original remote sensing image; S2, extract the target bounding box set of the sensitive targets to be protected in the original remote sensing image, map the pixel region covered by each target bounding box to the activation value on a single-channel all-zero matrix, generate a global binary union mask, and output a carrier image with the same size as the original remote sensing image and without sensitive targets based on the global binary union mask; S3, the original remote sensing image is cropped 1:1 without distortion according to the target bounding box to obtain a set of sensitive target pixel blocks, which are packaged into a secret map of a fixed size without scaling, and a metadata file for recording the location information of all sensitive targets and an effective region mask for identifying the location of sensitive target pixels and the boundary of the zero-filled area are generated. S4. Configure a random seed key, initialize a preset neural network with steganographic feature mapping function using the random seed key, and freeze all weights after initialization to build a fixed decoder. S5, the carrier image is divided into several non-overlapping image blocks, the texture complexity of each image block is calculated and sorted, a set of high texture blocks is selected and spliced ​​to obtain a compact domain carrier and the corresponding compact domain perturbation, and at the same time, the image blocks of the compact domain perturbation are written back to the zero-base map with the same resolution as the carrier image using the scattering operator to generate a global sparse perturbation. S6, the global sparse perturbation is superimposed on the carrier image, threshold truncation is performed to obtain a floating-point dense image, and the dense image is output through loss optimization by combining the secret map and the effective region mask; S7. Calculate the pixel difference between the secret image and the carrier image, separate the residual perturbation signal, process it through the extraction and rearrangement operator, input it into the fixed decoder to obtain the recovered secret map, and combine it with the metadata file to restore the full-image scene-level recovered image.

2. The method for steganography embedding and image restoration based on sensitive targets in remote sensing images according to claim 1, characterized in that... When generating the global binary union mask, the coordinates of each target bounding box in the target bounding box set are traversed. The pixel coordinates covered by the target bounding box are assigned pixel values ​​on the all-zero matrix, and the pixel values ​​of the corresponding regions are set from 0 to 1, forming an independent local binary mask. The expression is: ； in, For the first Local binary mask of the bounding box of each target; These are the row coordinates of the image pixels; These are the column coordinates of the image pixels; , For the first The top-left pixel coordinates of the bounding box of each target; , For the first The bottom right pixel coordinates of the target bounding box; Traverse the local binary masks of all target bounding boxes, perform logical OR merge operations pixel by pixel, perform spatial fusion, and remove overlapping and redundant regions to generate a global binary union mask with the same resolution and single channel as the original remote sensing image.

3. The method for steganography embedding and image restoration based on sensitive targets in remote sensing images according to claim 1, characterized in that... When the global binary union mask outputs a carrier image that is the same size as the original remote sensing image and does not contain sensitive targets, a background inpainting algorithm combined with the global binary union mask is used to fill and inpaint the sensitive target regions in the original remote sensing image, generating an inpainted carrier image that does not contain sensitive targets; wherein, the background inpainting algorithm adopts the PatchMatch inpainting algorithm based on image patch matching or the Telea inpainting algorithm based on partial differential equation propagation.

4. The method for steganography embedding and image restoration based on sensitive targets in remote sensing images according to claim 1, characterized in that... The secret map is obtained by placing each sensitive target pixel block of the sensitive target pixel block set sequentially onto a fixed-size blank canvas according to a preset grid layout rule without scaling; for grid cells not completely filled by sensitive target pixel blocks, a zero-value filling operation is performed, thereby obtaining the secret map, the expression of which is: ； ； ； in, For secret maps; The global pixel coordinates of the secret graph canvas; For image channels; For the first One sensitive target pixel block; , for Local pixel coordinates; For the first A preset grid area in the secret map for each sensitive target pixel block; The total number of sensitive target pixel blocks; This represents the union of all sensitive target pixel blocks within the grid region of the map; Represents the union; , for The starting coordinates are at the top left corner of the graph.

5. The method for steganalysis and image restoration based on sensitive targets in remote sensing images according to claim 1, characterized in that... When configuring a random seed key, there are two configuration modes: Mode 1: Generate a unique random seed key for each image, which is used to encrypt the metadata file before sending; during decryption, receive the encrypted metadata file, extract the random seed key, and initialize the local decoder; Mode 2: When configuring a random seed key, the encryption and decryption ends are pre-configured with the same static shared random seed key to initialize the decoder, thereby improving transmission efficiency when transmitting the metadata file.

6. The method for steganography embedding and image restoration based on sensitive targets in remote sensing images according to claim 1, characterized in that... Specifically, S5 includes: The carrier image is divided into several non-overlapping image blocks; Calculate the texture complexity based on the local binary pattern histogram entropy for each image patch; Sort by texture complexity from high to low and select a set of high texture blocks that meet a preset threshold. The high-texture blocks in the high-texture block set are sequentially spliced ​​into a compact mesh domain to obtain a compact domain carrier and the corresponding compact domain perturbation. Using a scattering operator, each image patch of the compact domain perturbation is written back to the zero-base map with the same resolution as the carrier map according to its coordinate position in the set of high-texture patches, thereby generating a global sparse perturbation.

7. The method for steganography embedding and image restoration based on sensitive targets in remote sensing images according to claim 1, characterized in that... The formula for applying threshold truncation is: ； in, For floating-point dense images; For carrier diagram; For global sparse perturbation; This is a threshold truncation function that restricts the input values ​​to a range. Inside; , These are the upper and lower limits of the set cutoff threshold, respectively.

8. The method for steganalysis and image restoration based on sensitive targets in remote sensing images according to claim 1, characterized in that... When combining the secret map and the effective region mask to optimize the output of the dense image through loss, a joint loss function is used for iterative optimization. The joint loss function includes hidden distortion loss, net load recovery loss, robust recovery loss, and total variational smoothing regularization term; The concealment distortion loss is calculated by measuring the mean square error between the floating-point dense image and the carrier image. The net load recovery loss is combined with the effective region mask to calculate the recovered secret map and the mean square error of the secret map within the effective region; The robust recovery loss is calculated by applying a simulated channel attack operator to the secret image and then calculating the mean square error between the recovered map and the secret map within the effective region under attack conditions. The total variational smoothing regularization term calculates the total variational loss by applying global sparse perturbations; The expression for the joint loss function is: ； ； ； ； in, For the joint loss function; This is due to concealed distortion loss; This is the net load recovery loss; Weighted for robust recovery losses; For total variational smoothing regularization; The weights for dynamic concealment loss vary with each iteration. change; , Fixed weighting coefficients; The weighting factor for net load recovery loss; For floating-point dense images; For carrier diagram; Calculation of mean square error; For secret maps; To recover the secret map; For effective area masking; For fixed decoders; For extracting rearrangement operators; To simulate channel attack operators; This is the total variation loss, used to calculate the mean of the absolute differences in the perturbation amplitudes between adjacent pixels; For global sparse perturbation; The dynamic concealment loss weight is dynamically adjusted through a piecewise linear temperature ramping scheduling mechanism, and its expression is as follows: ； in, This refers to the current iteration round; These are the initial baseline weights; For preheating cycles; The maximum upper limit weight; This represents the total number of iterations.

9. A method for steganography embedding and image restoration based on sensitive targets in remote sensing images according to claim 1, characterized in that... Specifically, S7 is: The pixel difference between the secret image and the carrier image is calculated to separate the residual perturbation signal. This residual perturbation signal is then processed by a decimation-rearrangement operator and input into a fixed decoder to obtain the recovered secret map, expressed as: ； in, To recover the secret map; For fixed decoders; For extracting rearrangement operators; For dense images; For carrier diagram; A collection of high-texture blocks; Based on the recovered secret map and the metadata file, zero-filled pixels are stripped, and the sensitive target pixel blocks of the original pixels are restored at a 1:1 ratio without any reverse scaling. The restored sensitive target pixel blocks are pasted back to their corresponding positions in the carrier image according to the coordinate information in the metadata file to obtain a full-scene-level restored image.

10. A steganalysis and image restoration apparatus based on sensitive targets in remote sensing images, used to implement the steganalysis and image restoration method based on sensitive targets in remote sensing images as described in any one of claims 1-9, characterized in that, include: The image acquisition and initialization unit is used to acquire the original remote sensing image and initialize a single-channel all-zero matrix with the same resolution as the original remote sensing image. The binary mask unit is used to extract the set of target bounding boxes of the sensitive targets to be protected in the original remote sensing image, map the pixel region covered by each target bounding box to the activation value on a single-channel all-zero matrix, generate a global binary union mask, and output a carrier image with the same size as the original remote sensing image and without sensitive targets based on the global binary union mask. The atlas unit is used to perform 1:1 distortion-free cropping of the original remote sensing image according to the target bounding box to obtain a set of sensitive target pixel blocks, and to package them into a secret atlas of a fixed size without scaling, thereby generating a metadata file for recording the location information of all sensitive targets and an effective region mask for identifying the location of sensitive target pixels and the boundary of zero-filled regions. A key configuration unit is used to configure a random seed key, initialize a preset neural network with steganalysis feature mapping function using the random seed key, and freeze all weights after initialization to build a fixed decoder. The global perturbation unit is used to divide the carrier image into several non-overlapping image blocks, calculate the texture complexity of each image block and sort them, select the set of high texture blocks and stitch them together to obtain the compact domain carrier and the corresponding compact domain perturbation. At the same time, the scattering operator is used to write each image block of the compact domain perturbation back to the zero-base map with the same resolution as the carrier image to generate global sparse perturbation. The steganography embedding unit is used to superimpose the global sparse perturbation onto the carrier map, perform threshold truncation to obtain a floating-point dense image, and combine the secret map with the effective region mask to output the dense image through loss optimization. The restoration unit is used to calculate the pixel difference between the secret image and the carrier image, separate the residual perturbation signal, process it through the extraction and rearrangement operator, input it into the fixed decoder to obtain the restored secret map, and combine it with the metadata file to restore the full-image scene-level restored image.

Images