Intelligent connected vehicle privacy image desensitization system

Abstract

The application discloses a smart connected vehicle privacy image desensitization system and belongs to the technical field of smart connected vehicle data security, comprising an image preprocessing unit, an encryption unit, a transmission and storage unit and a decryption unit; the image preprocessing unit is configured to identify the privacy area in the image through a lightweight target detection model and generate a binary mask, then separate the color image on the red, green and blue color channels, and perform adaptive binary processing on the privacy area and the non-privacy area of each channel to generate three groups of binary matrices; the encryption unit is configured to generate two shares for each pixel in the binary matrix of each color channel based on a visual password scheme, according to whether the pixel belongs to the privacy area or the non-privacy area, and apply different rules. The application can solve the problems of color image privacy protection difficulty and traditional desensitization irreversibility in the prior art, and realize safe, efficient and reversible color image privacy protection.

Description

Technical Field

[0001] This invention relates to the field of data security technology for intelligent connected vehicles, specifically to a privacy image desensitization system for intelligent connected vehicles. Background Technology

[0002] With the widespread adoption of intelligent connected vehicles, in-vehicle cameras are capturing vast amounts of image data containing private information about pedestrians, vehicles, and other road users. The use of this data in vehicle-to-everything (V2X) communication, accident analysis, and cloud storage poses a serious risk of privacy breaches. Therefore, how to protect personal privacy while utilizing this image data has become a critical issue that urgently needs to be addressed.

[0003] Traditional image anonymization methods, such as mosaic and blurring, are irreversible, resulting in the permanent loss of original information and failing to meet legitimate requirements such as post-incident traceability and liability determination. Some reversible anonymization methods suffer from insufficient security or pixel expansion (leading to larger image sizes), increasing the burden on storage and transmission.

[0004] Visual cryptography, as an important security technology, can decompose a secret image into multiple shared copies. Each shared copy does not reveal any information, and the secret can be deciphered by human vision simply by superimposing multiple shared copies. However, traditional visual cryptography schemes are mainly designed for binary images, making them difficult to apply directly to color images generated by intelligent connected vehicles. Furthermore, they often suffer from pixel expansion issues, limiting their application in resource-constrained in-vehicle environments.

[0005] Therefore, those skilled in the art have provided a privacy image desensitization system for intelligent connected vehicles to address the problems mentioned in the background section. Summary of the Invention

[0006] The purpose of this invention is to provide a privacy image desensitization system for intelligent connected vehicles, so as to solve the problems of difficulty in protecting the privacy of color images and the irreversibility of traditional desensitization in the prior art, and to achieve safe, efficient and reversible privacy protection of color images.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] The intelligent connected vehicle privacy image desensitization system includes an image preprocessing unit, an encryption unit, a transmission and storage unit, and a decryption unit;

[0009] The image preprocessing unit is configured to identify privacy regions in the image and generate binary masks using a lightweight object detection model. Subsequently, the color image is separated in the red, green, and blue color channels, and adaptive binarization processing is performed on the privacy and non-privacy regions of each channel to generate three sets of binary matrices.

[0010] The encryption unit is configured to use a visual cryptography scheme, whereby for each pixel in the binary matrix of each color channel, two shared copies are generated based on whether it belongs to a private or non-private region, and the shared copies of the three channels are merged into two complete color noise images.

[0011] The transmission and storage unit is configured to store the two color noise images in the vehicle unit and the remote server respectively, achieving physical isolation;

[0012] The decryption unit is configured to synthesize two color noise images by stacking operations or computational reconstruction operations to recover the original color image.

[0013] As a further aspect of the present invention: the lightweight target detection model in the image preprocessing unit adopts a multi-scale feature enhancement module based on an attention mechanism in the feature fusion stage. This module adaptively adjusts the fused features by calculating the importance weights of feature maps at different scales.

[0014] The specific process is as follows: For multiple scale feature maps to be fused, channel descriptors are first obtained through global average pooling. Then, channel attention weights are generated through a fully connected layer and an activation function. Finally, the enhanced fused feature map is obtained by weighted summation, as expressed by the formula:

[0015] ;

[0016] Among them, F i Let F represent the feature map at the i-th scale, GAP represent the global average pooling operation, MLP represent the multilayer perceptron, sigma represent the sigmoid activation function, and F... enhanced It is the enhanced fused feature map.

[0017] As a further aspect of the present invention: after generating a binary mask, the image preprocessing unit uses a mask post-processing algorithm based on connectivity analysis to optimize the mask region;

[0018] The mask post-processing algorithm identifies and filters out abnormal regions caused by false detections by the detection model by calculating the area, perimeter, and aspect ratio of the bounding rectangle of each connected region. The specific judgment rule is: if the area of ​​a connected region is less than a certain proportion of the total area of ​​the image, or its aspect ratio exceeds a preset range, then the region is removed from the privacy mask.

[0019] As a further aspect of the present invention: the adaptive binarization process employs a hybrid thresholding algorithm that combines global and local statistical characteristics;

[0020] The specific process is as follows: For each color channel, the global threshold is first calculated using the Otsu algorithm. Subsequently, within the privacy region, the mean and standard deviation of grayscale values ​​are calculated within a local window centered on each pixel, thereby obtaining the local adaptive threshold. The final pixel binarization threshold Determined by a weighted average of the global threshold and the local adaptive threshold:

[0021] ;

[0022] in, and These are the mean and standard deviation of the grayscale values ​​within the local window, respectively, with k and α being adjustable parameters.

[0023] As a further aspect of the present invention: when generating the shared copy, the encryption unit uses a pseudo-random number generator based on a linear feedback shift register to drive the selection logic for the randomization process of pixels in the privacy region.

[0024] The pseudo-random number generator produces a uniformly distributed random sequence, ensuring that for the original binary pixel value, the probability of selecting either of the two sharing pairs is 50%, thereby guaranteeing that a single sharing copy does not leak any original information.

[0025] As a further aspect of the present invention: before transmitting the shared copy to the remote server, the transmission and storage unit performs a lossless compression operation on the color noise image;

[0026] The compression operation employs context-adaptive arithmetic coding, which first performs prediction residual calculation on the image pixel values, and then performs arithmetic coding on the residual sequence to reduce storage and transmission bandwidth requirements.

[0027] As a further aspect of the present invention: before performing the XOR operation, the computational reconstruction operation in the decryption unit introduces a shared-part alignment verification mechanism based on image gradient.

[0028] The specific process is as follows: calculate the gradient magnitude maps of the two shared images respectively, and determine whether the shared images have been misaligned during transmission and storage by comparing the similarity of the gradient magnitude maps in the non-privacy background area. If the similarity is lower than the threshold, the automatic realignment process is triggered.

[0029] As a further aspect of the present invention: after the computational reconstruction operation recovers the binary images of each channel by performing the XOR operation, the recovered privacy region is processed by morphological closing operation;

[0030] The morphological closing operation first performs a dilation operation and then an erosion operation to eliminate small holes and burrs in the restored image caused by slight misalignment of shared components or noise, thereby improving the visual quality of the restored image.

[0031] As a further aspect of the present invention, it also includes a metadata management module, which is used to generate and encapsulate auxiliary information related to the privacy area as metadata during the encryption process;

[0032] The metadata includes at least binary mask information defining the location of the privacy region, color space and channel separation parameters, and threshold parameters for binarization processing; this metadata is stored together with the shared copy and read during the decryption process to guide the accurate restoration of non-privacy background regions and the correct reconstruction of color images.

[0033] As a further aspect of the present invention: both the image preprocessing unit and the decryption unit integrate a dynamic performance regulator;

[0034] The dynamic performance regulator monitors the frame rate and latency of the system processing pipeline in real time. When the performance index is lower than the preset threshold, it automatically reduces the processing resolution of the target detection model or skips the detection of some frames to ensure the real-time performance of the system in complex scenarios.

[0035] Compared with the prior art, the beneficial effects of the present invention are:

[0036] The intelligent connected vehicle privacy image desensitization system provided by this invention constructs a complete technical solution integrating intelligent preprocessing, secure encryption, distributed storage, and high-quality recovery, resulting in multiple significant benefits: In terms of security, it utilizes visual cryptography principles to decompose privacy information into two meaningless noise-sharing copies and stores them in physical isolation, achieving "data not leaving the vehicle and plaintext not being stored on disk" from a mechanism perspective, effectively resisting external attacks and the risk of internal data leakage; In terms of practicality, through pass-through processing of non-privacy backgrounds, efficient compression algorithms, and dynamic resource adjustment, it greatly reduces the occupation of onboard computing resources, storage space, and network bandwidth, ensuring the efficient and stable operation of the system in a real-time onboard environment; In terms of recovery quality, by introducing a series of optimization measures such as alignment verification, morphological post-processing, and metadata management, it effectively overcomes the inherent defects of traditional visual cryptography in image recovery, such as decreased contrast and blurred contours, and can recover high-definition, high-fidelity original images under legitimate needs, perfectly meeting the subsequent application requirements in scenarios such as accident liability determination. Attached Figure Description

[0037] Figure 1 This is a structural block diagram of a privacy image desensitization system for intelligent connected vehicles. Detailed Implementation

[0038] 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 some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] As mentioned in the background section of this application, research has found that existing visual cryptography schemes are mainly designed for binary images, which are difficult to apply directly to color images generated by intelligent connected vehicles. They also often suffer from pixel expansion problems, which limit their application in resource-constrained in-vehicle environments and have certain shortcomings.

[0040] To address the aforementioned shortcomings, this application discloses a privacy image desensitization system for intelligent connected vehicles, which solves the problems of difficulty in protecting the privacy of color images and the irreversibility of traditional desensitization in the prior art, and achieves safe, efficient and reversible privacy protection of color images.

[0041] The following will describe in detail, with reference to the accompanying drawings, how the solution of this application solves the above-mentioned technical problems.

[0042] Please see Figure 1 In this embodiment of the invention, the intelligent connected vehicle privacy image desensitization system includes an image preprocessing unit, an encryption unit, a transmission and storage unit, and a decryption unit. The image preprocessing unit is configured to identify privacy regions in an image using a lightweight object detection model and generate a binary mask. Subsequently, it separates the color image in the red, green, and blue color channels and performs adaptive binarization processing on the privacy and non-privacy regions of each channel to generate three sets of binary matrices. The encryption unit is configured to generate two shared copies for each pixel in the binary matrix of each color channel based on whether it belongs to a privacy region or a non-privacy region, according to different rules. The shared copies of the three channels are then merged into two complete color noise images. The transmission and storage unit is configured to store the two color noise images in an on-board unit and a remote server, respectively, to achieve physical isolation. The decryption unit is configured to synthesize the two color noise images through stacking operations or computational reconstruction operations to recover the original color image. This application constructs a complete privacy image processing pipeline based on visual cryptography. Through the collaborative work of four core units—preprocessing, encryption, transmission and storage, and decryption—it achieves automated, secure desensitization and on-demand restoration of privacy regions in images of intelligent connected vehicles, laying the foundation for the system-level framework of the entire solution.

[0043] In this embodiment, the lightweight object detection model in the image preprocessing unit employs a multi-scale feature enhancement module based on an attention mechanism during the feature fusion stage. This module adaptively adjusts the fused features by calculating the importance weights of feature maps at different scales. Specifically, for multiple scale feature maps to be fused, channel descriptors are first obtained through global average pooling, then channel attention weights are generated through a fully connected layer and an activation function, and finally, a weighted sum is calculated to obtain the enhanced fused feature map. The formula is as follows:

[0044] ;

[0045] Among them, F i Let F represent the feature map at the i-th scale, GAP represent the global average pooling operation, MLP represent the multilayer perceptron, sigma represent the sigmoid activation function, and F... enhanced This is the enhanced fused feature map. This setting improves the accuracy of privacy region detection, especially for small-scale targets (such as distant license plates). By introducing an attention mechanism to adaptively weight and fuse multi-scale features, the object detection model can focus more on feature channels containing rich semantic information, thereby generating more accurate initial privacy region proposals.

[0046] In this embodiment, after generating a binary mask, the image preprocessing unit employs a post-processing algorithm based on connectivity analysis to optimize the mask region. The post-processing algorithm identifies and filters out abnormal regions caused by false detections from the detection model by calculating the area, perimeter, and aspect ratio of the bounding rectangle of each connected region. Specifically, if the area of ​​a connected region is less than a certain proportion of the total image area, or its aspect ratio exceeds a preset range, that region is removed from the privacy mask. This setting "cleans" and optimizes the initially generated privacy mask. By analyzing the geometric properties of connected regions, it effectively filters out isolated, abnormal, small-area false detection regions caused by detection model noise, thereby improving the accuracy and reliability of the mask and reducing subsequent false encryption of non-privacy regions.

[0047] In this embodiment, the adaptive binarization process employs a hybrid thresholding algorithm that combines global and local statistical characteristics; specifically, for each color channel, the global threshold is first calculated using the Otsu algorithm. Subsequently, within the privacy region, the mean and standard deviation of grayscale values ​​are calculated within a local window centered on each pixel, thereby obtaining the local adaptive threshold. The final pixel binarization threshold Determined by a weighted average of the global threshold and the local adaptive threshold: ;in, and These are the mean and standard deviation of grayscale values ​​within the local window, respectively, while k and α are adjustable parameters. This setting addresses the issue of poor binarization performance under complex lighting conditions. By combining the globally optimal threshold with a locally adaptive threshold, this scheme can make the binarized contours of privacy regions clearer and more continuous, especially improving the processing effect of uneven brightness areas, thus providing high-quality binary input for subsequent visual cryptography encryption.

[0048] In this embodiment, when generating the shared copy, the encryption unit uses a pseudo-random number generator based on a linear feedback shift register to drive the selection logic for the randomization process of pixels in the privacy region. The pseudo-random number generator produces a uniformly distributed random sequence, ensuring that for the original binary pixel value, the selection probability of each of the two shared pair combinations is 50%, thereby guaranteeing that a single shared copy does not leak any original information. This setting enhances the security and unpredictability of the encryption process. By using a hardware-friendly linear feedback shift register instead of a simple software random function, it can generate pseudo-random sequences with better statistical characteristics and longer periods, ensuring that attackers cannot deduce the original information by analyzing the noise patterns of a single shared copy, thus strengthening the cryptographic security of the system.

[0049] In this embodiment, the transmission and storage unit performs lossless compression on the color noise image before transmitting the shared copy to the remote server. The compression operation employs context-adaptive arithmetic coding, first calculating the prediction residuals of the image pixel values, and then performing arithmetic coding on the residual sequence to reduce storage and transmission bandwidth requirements. This setup reduces the system's resource requirements for storage and communication bandwidth. By efficiently compressing the generated, inherently random, noise image, it saves valuable network resources during transmission between different units (vehicle-edge-cloud) and reduces the storage costs of the remote server, thereby improving the system's economy and practicality.

[0050] In this embodiment, the reconstruction calculation in the decryption unit introduces a shared-part alignment verification mechanism based on image gradients before performing the XOR operation. Specifically, the gradient magnitude maps of the two shared-part images are calculated separately. By comparing the similarity of the gradient magnitude maps in non-privacy background regions, it is determined whether the shared parts have been misaligned during transmission and storage. If the similarity is below a threshold, an automatic realignment process is triggered. This setting ensures the input quality of the decrypted image recovery. By adding an automated alignment verification before the critical decryption XOR operation, it can detect and correct minor deformations or misalignments of the shared-part images that may be caused by storage compression or transmission, thus avoiding blurry or ghosting problems in the recovered image due to input deviations at the source.

[0051] In this embodiment, after the XOR operation recovers the binary images of each channel, the reconstruction operation uses morphological closing to process the recovered privacy regions. The morphological closing operation first performs dilation followed by erosion to eliminate small holes and burrs in the recovered image caused by slight misalignment of shared components or noise, thus improving the visual quality of the recovered image. This post-processing of the decrypted and recovered image enhances visual quality. Through morphological closing, it can repair minor imperfections (such as holes and jagged edges) in the recovered binary image caused by slight misalignment or binarization threshold deviation, resulting in a final image with smoother edges and more complete regions, thus improving the user experience.

[0052] In this embodiment, a metadata management module is also included, used to generate and encapsulate auxiliary information related to the privacy region as metadata during the encryption process. The metadata includes at least binary mask information defining the location of the privacy region, color space and channel separation parameters, and threshold parameters for binarization processing. This metadata is stored together with the shared copy and read during decryption to guide the accurate recovery of non-privacy background areas and the correct reconstruction of the color image. By systematically encapsulating all parameters required for encryption and decryption (mask, color parameters, threshold), it is ensured that in a physically isolated distributed storage environment, the decryption end can reconstruct the color losslessly and accurately distinguish between privacy and background areas, guaranteeing the reliable closed-loop operation of the entire process.

[0053] In this embodiment, both the image preprocessing unit and the decryption unit integrate a dynamic performance regulator. The dynamic performance regulator monitors the frame rate and latency of the system processing pipeline in real time. When performance indicators fall below a preset threshold, it automatically reduces the processing resolution of the target detection model or skips the detection of some frames to ensure the real-time performance of the system in complex scenarios. This setting ensures the real-time performance and robustness of the system in resource-constrained in-vehicle environments. By dynamically monitoring the system load and adaptively adjusting processing strategies (such as reducing resolution), it prioritizes system responsiveness when computing resources are limited, avoiding stuttering or data loss due to processing latency, and ensuring stable service under various operating conditions.

[0054] The intelligent connected vehicle privacy image desensitization system provided by this invention constructs a complete technical solution integrating intelligent preprocessing, secure encryption, distributed storage, and high-quality recovery, resulting in multiple significant benefits: In terms of security, it utilizes visual cryptography principles to decompose privacy information into two meaningless noise-sharing copies and stores them in physical isolation, achieving "data not leaving the vehicle and plaintext not being stored on disk" from a mechanism perspective, effectively resisting external attacks and the risk of internal data leakage; In terms of practicality, through pass-through processing of non-privacy backgrounds, efficient compression algorithms, and dynamic resource adjustment, it greatly reduces the occupation of onboard computing resources, storage space, and network bandwidth, ensuring the efficient and stable operation of the system in a real-time onboard environment; In terms of recovery quality, by introducing a series of optimization measures such as alignment verification, morphological post-processing, and metadata management, it effectively overcomes the inherent defects of traditional visual cryptography in image recovery, such as decreased contrast and blurred contours, and can recover high-definition, high-fidelity original images under legitimate needs, perfectly meeting the subsequent application requirements in scenarios such as accident liability determination.

[0055] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

[0056] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A privacy-preserving image desensitization system for intelligent connected vehicles, characterized in that, It includes an image preprocessing unit, an encryption unit, a transmission and storage unit, and a decryption unit; The image preprocessing unit is configured to identify privacy regions in the image and generate binary masks using a lightweight object detection model. Subsequently, the color image is separated in the red, green, and blue color channels, and adaptive binarization processing is performed on the privacy and non-privacy regions of each channel to generate three sets of binary matrices. The encryption unit is configured based on a visual cryptography scheme. For each pixel in the binary matrix of each color channel, it generates two shared copies according to whether the pixel belongs to a privacy region or a non-privacy region, applying different rules. The shared copies of the three channels are then merged into two complete color noise images. When generating the shared copies, the encryption unit uses a pseudo-random number generator based on a linear feedback shift register to drive the selection logic for the randomization process of pixels in the privacy region. The pseudo-random number generator produces a uniformly distributed random sequence, ensuring that for the original binary pixel value, the selection probability of the two shared pair combinations is 50%. The transmission and storage unit is configured to store the two color noise images in the vehicle unit and the remote server respectively, achieving physical isolation; The decryption unit is configured to synthesize two color noise images by performing a reconstruction operation to recover the original color image. Before performing the XOR operation, the reconstruction operation in the decryption unit introduces a shared copy alignment verification mechanism based on image gradients. The specific process is as follows: the gradient magnitude maps of the two shared copy images are calculated respectively. By comparing the similarity of the gradient magnitude maps in the non-privacy background area, it is determined whether the shared copy has been misaligned during transmission and storage. If the similarity is lower than the threshold, the automatic realignment process is triggered. In the feature fusion stage, the lightweight object detection model in the image preprocessing unit adopts a multi-scale feature enhancement module based on an attention mechanism. This module adaptively adjusts the fused features by calculating the importance weights of feature maps at different scales. The specific process is as follows: For multiple scale feature maps to be fused, channel descriptors are first obtained through global average pooling, then channel attention weights are generated through a fully connected layer and activation function, and finally the enhanced fused feature map is obtained by weighted summation. After generating a binary mask, the image preprocessing unit uses a mask post-processing algorithm based on connectivity analysis to optimize the mask region. The mask post-processing algorithm identifies and filters out abnormal regions caused by false detections by the detection model by calculating the area, perimeter, and aspect ratio of the bounding rectangle of each connected region. The specific judgment rule is: if the area of ​​a connected region is less than a certain proportion of the total image area, or its aspect ratio exceeds a preset range, then the region is removed from the privacy mask. The adaptive binarization process employs a hybrid thresholding algorithm that combines global and local statistical characteristics. Specifically, for each color channel, the global threshold is first calculated using the Otsu algorithm. Then, within the privacy region, the grayscale mean and standard deviation are calculated within a local window centered on the pixel to obtain the local adaptive threshold. Finally, the binarization threshold of the pixel is determined by a weighted average of the global threshold and the local adaptive threshold.

2. The intelligent connected vehicle privacy image desensitization system according to claim 1, characterized in that, Before transmitting the shared copy to the remote server, the transmission and storage unit performs lossless compression on the color noise image. The compression operation employs context-adaptive arithmetic coding, which first calculates the prediction residuals of the image pixel values ​​and then performs arithmetic coding on the residual sequence to reduce storage and transmission bandwidth requirements.

3. The intelligent connected vehicle privacy image desensitization system according to claim 2, characterized in that, After the computational reconstruction operation recovers the binary images of each channel through the XOR operation, the recovered privacy region is processed using a morphological closing operation. The morphological closing operation first performs a dilation operation and then an erosion operation to eliminate small holes and burrs in the restored image caused by slight misalignment of shared components or noise.

4. The intelligent connected vehicle privacy image desensitization system according to claim 3, characterized in that, It also includes a metadata management module, which is used to generate and encapsulate auxiliary information related to privacy areas as metadata during the encryption process; The metadata includes at least binary mask information defining the location of the privacy region, color space and channel separation parameters, and threshold parameters for binarization processing; this metadata is stored together with the shared copy and read during the decryption process to guide the accurate restoration of non-privacy background regions and the correct reconstruction of color images.

5. The intelligent connected vehicle privacy image desensitization system according to claim 4, characterized in that, Both the image preprocessing unit and the decryption unit integrate a dynamic performance regulator. The dynamic performance regulator monitors the frame rate and latency of the system processing pipeline in real time. When the performance index is lower than the preset threshold, it automatically reduces the processing resolution of the target detection model or skips the detection of some frames.

Images