CN-121462702-B - Intelligent network-connected automobile privacy image desensitization system

CN121462702BCN 121462702 BCN121462702 BCN 121462702BCN-121462702-B

Abstract

The invention discloses an intelligent network connection automobile privacy image desensitization system, which belongs to the technical field of intelligent network connection automobile data security, and comprises an image preprocessing unit, an encryption unit, a transmission and storage unit and a decryption unit, wherein the image preprocessing unit is configured to recognize privacy areas in images through a lightweight object detection model and generate binary masks, then separate color images on three color channels of red, green and blue, perform self-adaptive binarization processing on the privacy areas and the non-privacy areas of each channel to generate three groups of binary matrixes, and the encryption unit is configured to generate two sharing parts for each pixel in the binary matrixes of each color channel according to the privacy areas or the non-privacy areas based on a visual password scheme. The invention can solve the problems of difficult privacy protection and irreversible traditional desensitization of the color image in the prior art, and realize safe, efficient and reversible privacy protection of the color image.

Inventors

CHEN HUI
MA LIN
HE PENG
SONG JING
YAO YUXIN
HUANG HAIJIAO

Assignees

湖北大学

Dates

Publication Date: 20260508
Application Date: 20260108

Claims (5)

1. The intelligent network-connected automobile privacy image desensitization system is characterized by comprising an image preprocessing unit, an encryption unit, a transmission and storage unit and a decryption unit; the image preprocessing unit is configured to recognize a privacy zone in an image through a lightweight object detection model and generate a binary mask, then separate a color image on three color channels of red, green and blue, and perform self-adaptive binarization processing on the privacy zone and the non-privacy zone of each channel to generate three sets of binary matrixes; The encryption unit is configured to generate two sharing shares according to different rules for each pixel in the binary matrix of each color channel based on a visual cryptography scheme and the privacy area or the non-privacy area, and respectively combine the sharing shares of three channels into two complete color noise pictures, the encryption unit drives selection logic by adopting a pseudo-random number generator based on a linear feedback shift register aiming at the randomization process of the pixels in the privacy area when generating the sharing shares, and the pseudo-random number generator generates a random sequence which is uniformly distributed to ensure that the selection probability of the combination of the two sharing pairs is 50% for the original binary pixel value; The transmission and storage unit is configured to store the two color noise pictures in the vehicle-mounted unit and the remote server respectively, so as to realize physical isolation; The decryption unit is configured to synthesize two color noise pictures through calculation and reconstruction operation, and restore an original color image; the computing and reconstructing operation in the decryption unit introduces a sharing part alignment checking mechanism based on image gradients before performing exclusive OR operation, and comprises the specific processes of respectively computing gradient amplitude maps of two sharing part images, judging whether the sharing part is misplaced in the transmission and storage process by comparing the similarity of the gradient amplitude maps in a non-privacy background area, and triggering an automatic realignment process if the similarity is lower than a threshold value; In a feature fusion stage, a lightweight object detection model in an image preprocessing unit adopts a multi-scale feature enhancement module based on an attention mechanism, and the module carries out self-adaptive adjustment on fusion features by calculating importance weights of feature images of different scales; The mask post-processing algorithm identifies and filters abnormal areas generated by false detection of a detection model by calculating the area, perimeter and external rectangular aspect ratio of each connected area, wherein the specific judgment rule is that if the area of the connected area is smaller than a specific proportion of the total area of the image or the aspect ratio of the connected area exceeds a preset range, the area is removed from the privacy mask; The self-adaptive binarization processing adopts a mixed threshold algorithm combining global and local statistical characteristics, and the specific process is that for each color channel, firstly, an Otsu algorithm is used for calculating a global threshold, then, in a privacy area, a gray average value and a standard deviation are calculated in a local window taking a pixel point as a center, and further, a local self-adaptive threshold is obtained, and finally, the binarization threshold of the pixel point is determined by weighting the global threshold and the local self-adaptive threshold.
2. The intelligent network-connected privacy image desensitizing system according to claim 1, wherein said transmission and storage unit performs lossless compression operation on the color noise picture before transmitting the share to the remote server; The compression operation adopts context-based adaptive arithmetic coding, firstly carries out prediction residual calculation on picture pixel values, and then carries out arithmetic coding on residual sequences so as to reduce storage and transmission bandwidth requirements.
3. The intelligent network-connected automobile privacy image desensitizing system according to claim 2, wherein the computing reconstruction operation adopts morphological closing operation to process the restored privacy area after finishing exclusive or operation to restore each channel binary image; The morphological closing operation is performed after the expansion operation is performed, and is used for eliminating tiny holes and burrs caused by slight dislocation of sharing parts or noise in the restored image.
4. The intelligent networked car privacy image desensitization system according to claim 3, further comprising a metadata management module for generating and packaging auxiliary information related to the privacy zone as metadata during encryption; The metadata at least comprises binary mask information defining the position of the privacy zone, color space and channel separation parameters and threshold parameters for binarization processing, and the metadata is stored together with the share and read in the decryption process for guiding the accurate recovery of the non-privacy background zone and the correct reconstruction of the color image.
5. The intelligent network-connected vehicle privacy image desensitizing system according to claim 4, wherein the image preprocessing unit and the decryption unit are integrated with dynamic performance adjusters; The dynamic performance regulator monitors the frame rate and delay of a system processing pipeline in real time, and when the performance index is lower than a preset threshold value, the processing resolution of the target detection model is automatically reduced or the detection of partial frames is skipped.

Description

Intelligent network-connected automobile privacy image desensitization system Technical Field The invention relates to the technical field of intelligent network-connected automobile data security, in particular to an intelligent network-connected automobile privacy image desensitization system. Background With the popularity of intelligent networking automobiles, vehicle-mounted cameras capture a large amount of image data containing private information of pedestrians, vehicles and other traffic participants. The data have serious privacy leakage risks in the processes of vehicle-road coordination, accident analysis, cloud storage and the like. Therefore, how to use these image data while protecting personal privacy is a critical issue to be solved. The traditional image desensitization method, such as mosaic, blurring and the like, is irreversible, so that original information is permanently lost, and legal requirements of post-hoc traceability, responsibility identification and the like cannot be met. The partial reversible desensitization method has the problems of insufficient security or pixel expansion (which leads to the increase of the image size) and increases the burden of storage and transmission. The visual password is taken as an important security technology, the secret image can be decomposed into multiple shares, no information is revealed in a single share, and the secret can be recovered through human eyes by only overlapping the multiple shares. However, the conventional visual cryptography scheme is mainly aimed at binary images, is difficult to be directly applied to color images generated by intelligent network-connected automobiles, and generally has the problem of pixel expansion, so that the application of the visual cryptography scheme in a vehicle-mounted environment with limited resources is limited. Accordingly, one skilled in the art would provide an intelligent networked car privacy image desensitization system to address the problems set forth in the background above. Disclosure of Invention The invention aims to provide an intelligent network-connected automobile privacy image desensitization system, which solves the problems of difficult color image privacy protection and irreversible traditional desensitization in the prior art, and realizes safe, efficient and reversible color image privacy protection. In order to achieve the above purpose, the present invention provides the following technical solutions: The intelligent network-connected automobile privacy image desensitization system comprises an image preprocessing unit, an encryption unit, a transmission and storage unit and a decryption unit; the image preprocessing unit is configured to recognize a privacy zone in an image through a lightweight object detection model and generate a binary mask, then separate a color image on three color channels of red, green and blue, and perform self-adaptive binarization processing on the privacy zone and the non-privacy zone of each channel to generate three sets of binary matrixes; the encryption unit is configured to generate two sharing parts by applying different rules to each pixel in the binary matrix of each color channel based on a visual password scheme according to the privacy area or the non-privacy area, and respectively combining the sharing parts of the three channels into two complete color noise pictures; The transmission and storage unit is configured to store the two color noise pictures in the vehicle-mounted unit and the remote server respectively, so as to realize physical isolation; The decryption unit is configured to synthesize two color noise pictures through stacking operation or calculation reconstruction operation, and restore an original color image. In the feature fusion stage, a multi-scale feature enhancement module based on an attention mechanism is adopted by a lightweight target detection model in the image preprocessing unit, and the module carries out self-adaptive adjustment on fusion features by calculating importance weights of feature graphs of different scales; The specific process includes that for a plurality of scale feature images to be fused, a channel descriptor is firstly obtained through global average pooling, then a channel attention weight is generated through a full connection layer and an activation function, and finally, the fused feature image after enhancement is obtained through weighted summation, and the formula is expressed as follows: ; wherein F i represents the feature map of the ith scale, GAP represents global average pooling operation, MLP represents multi-layer perceptron, sigma represents Sigmoid activation function, and F enhanced is the enhanced fusion feature map. After generating the binary mask, the image preprocessing unit optimizes the mask area by adopting a mask post-processing algorithm based on connectivity analysis; The mask post-processing algorithm identifies and filters abnormal regions generat