A privacy protection method for pedestrian re-identification anonymization keeping identity invariability

Abstract

The application discloses a pedestrian re-identification anonymization privacy protection method for maintaining identity invariability, which constructs a framework comprising an anonymization image generation module and an identity reservation module: a desensitization image generated by a traditional method is used as initial privacy protection supervision, an original pedestrian image is converted into a full-body anonymization image through an anonymization generator, and privacy compliance of the generated image is constrained through an adversarial loss and an L1 loss; the identity reservation module mines the internal relationship of all images with the same identity, adopts an AGW baseline model as an ReID model, jointly inputs the original image and the anonymization image, integrates identity classification loss, center loss and weighted regularization triplet loss for joint training, and iteratively optimizes a joint objective function of the anonymization generator and the ReID model to minimize the influence of anonymization on ReID performance. The generated anonymization image protects the privacy of pedestrians while significantly improving the first hit rate and accuracy of the ReID task.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision and privacy protection, face recognition, etc., in particular to a privacy protection method suitable for person re-identification (ReID), which protects the privacy of pedestrian images while maintaining data utility through anonymization technology to meet the identity recognition needs in intelligent transportation systems, public security monitoring and other scenarios. BACKGROUND

[0002] The person re-identification technology aims to identify the same pedestrian through cross-camera appearance features such as clothes, gait, body shape, etc., and is a core technology for smart cities and public security. With the development of deep learning such as convolutional neural networks and Transformers, the ReID performance has been significantly improved, and advanced results have been achieved on benchmark datasets such as Market-1501 and DukeMTMC. However, the original pedestrian images contain sensitive biological information such as facial features and body posture, which poses a risk of being misused to track individual trajectories or reconstruct identities, and faces challenges from strict privacy protection regulations such as the European Union's GDPR.

[0003] Traditional anonymization methods such as blurring, cropping and pixelization protect privacy by removing personally identifiable information (PII), but inevitably result in loss of semantic information, severely reducing the usability of data for downstream ReID tasks. For example, face blurring can destroy identity-related subtle features, leading to a decrease in the performance of recognition models. In recent years, deep learning-based anonymization techniques such as generative adversarial networks (GANs) have attempted to achieve privacy protection while maintaining image quality, but by introducing artificial distortion or disturbance, they have inadvertently changed the identity features of pedestrians such as the continuity of walking patterns, making anonymized data incompatible with ReID models and unable to meet the requirements of identity consistency for recognition tasks.

[0004] There are two major core defects in the existing technology. First, the imbalance between privacy protection and data utility. Traditional methods excessively sacrifice semantic information, while deep learning methods destroy identity features, both of which cannot achieve a balance between privacy protection and recognition performance in ReID. Second, the unintended changes to identity features. Existing anonymization techniques do not consider the special requirements of ReID for identity invariance, resulting in anonymized images that cannot be effectively recognized by existing ReID models.

[0005] In addition, existing ReID research focuses on performance improvement and ignores privacy protection, while general data de-identification (DeID) methods overemphasize privacy at the expense of data utility required by ReID, creating a technical gap. SUMMARY

[0006] To solve the technical problems in the background art, the present application provides a pedestrian re-identification anonymization privacy protection method with identity invariance,

[0007] Based on the first main aspect of the present application, a pedestrian re-identification anonymization privacy protection method with identity invariance is provided, comprising the following steps:

[0008] An anonymization framework is constructed, which includes an anonymization image generation module and an identity preservation module;

[0009] The desensitization image generated by the traditional method is used as the initial privacy protection supervision, and the original pedestrian image is converted into a full-body anonymous image through the anonymization generator;

[0010] The identity preservation module is used to mine the internal relationship of the same identity image, to preserve the identity invariant feature in the anonymization process, and to jointly train the anonymization generator and the pedestrian re-identification (ReID) model to minimize the impact of anonymization on the ReID performance;

[0011] The anonymization image generation module is constructed based on the conditional generative adversarial network (GAN), and the mapping relationship from the original image to the anonymization image is learned through adversarial training. The identity preservation module ensures that the identity features of the anonymization image are consistent with the original image by jointly optimizing the identity-related loss of the ReID model.

[0012] As a further preferred scheme, in the aforementioned pedestrian re-identification anonymization privacy protection method with identity invariance, the anonymization image generation module adopts the pix2pix framework, which includes a generator H X and a discriminator D Y The generator H X is configured to map the original image x i ∈X p to the anonymization image H X (x i ), and the discriminator D Y is configured to distinguish between the real supervision image y i ∈Y and the generated anonymization image H X (x i ), and is trained through the following adversarial loss function:

[0013]

[0014] Where n is the number of batch training samples.

[0015] As a further preferred scheme, in the aforementioned pedestrian re-identification anonymization privacy protection method with identity invariance, the anonymization image generation module also introduces an L1 loss to constrain the pixel-level difference between the generated image and the supervision image, and the total loss function is:

[0016]

[0017] wherein, λ1 is a hyperparameter, used to adjust the pixel loss weight.

[0018] As a further preferred solution, in the aforementioned identity invariance maintaining pedestrian re-identification anonymization privacy protection method, the identity reservation module adopts an AGW baseline model as the ReID model, the loss function of the ReID model integrates identity classification loss L id , center loss L c and weighted regularization triplet loss L wrt , and the original image and the anonymized image are mixed as input for joint training, and the total loss function is:

[0019]

[0020] wherein, is the AGW loss of the original image, is the AGW loss of the anonymized image.

[0021] As a further preferred solution, in the aforementioned identity invariance maintaining pedestrian re-identification anonymization privacy protection method, the traditional desensitization method includes blurring, pixelation or noise enhancement processing, and the generated initial supervised image Y includes the desensitized image after the above processing, which is used to guide the initial training of the anonymized image generation module.

[0022] As a further preferred solution, in the aforementioned identity invariance maintaining pedestrian re-identification anonymization privacy protection method, the original pedestrian image preprocessing step includes adjusting the image size to 256x128 pixels, and generating training samples through a data enhancement method, which includes but is not limited to random cropping, flipping or color jittering.

[0023] As a further preferred solution, in the aforementioned identity invariance maintaining pedestrian re-identification anonymization privacy protection method, in the joint training process, the image generation loss of the anonymized generator and the identity recognition loss of the ReID model are optimized at the same time, and the total objective function is:

[0024]

[0025] wherein, is the total loss of the anonymized generator, is the total loss of the ReID model.

[0026] As a further preferred scheme, in the aforementioned pedestrian re-identification anonymization privacy protection method with identity invariance preservation, the performance is evaluated by the following indexes: the peak signal-to-noise ratio (PSNR) and the structural similarity (SSIM) are used to measure the privacy protection effect, and the cumulative matching characteristics (CMC) and the average precision mean (mAP) are used to measure the pedestrian re-identification performance, so as to ensure that the anonymized image reaches a balance between privacy protection and data utility.

[0027] Based on the second main aspect of the present application, an electronic device is provided, comprising: at least one processor; a memory connected in communication with the at least one processor; the memory stores a computer program, which, when executed by the at least one processor, enables the at least one processor to implement the aforementioned pedestrian re-identification anonymization privacy protection method with identity invariance preservation.

[0028] Based on the third main aspect of the present application, a computer readable storage medium is provided, which stores a computer program, which, when executed, implements the aforementioned pedestrian re-identification anonymization privacy protection method with identity invariance preservation.

[0029] Advantages and beneficial effects of the present application:

[0030] Firstly, the present application breaks through the limitation of traditional separation of anonymization and re-identification technology, and constructs an integrated framework of the anonymization image generation module and the identity preservation module. By taking the desensitization image generated by the traditional desensitization method as the initial supervision signal, the privacy protection mapping relationship is learned by the conditional generative adversarial network (GAN), and the identity preservation module is introduced to mine the internal feature correlation of the same identity image, so as to ensure the visual invariance of the pedestrian identity in the anonymization process. The framework first realizes the systematic unity of privacy protection and data utility, and solves the core contradiction that the semantic information is lost due to privacy protection and the identity feature is destroyed by the deep learning method in the traditional method.

[0031] Secondly, unlike the defect of independent design of anonymization and re-identification model in the prior art, the present application proposes a cross-modal joint training mechanism, that is, by inputting the original image and the anonymized image into the ReID model at the same time, integrating the identity classification loss, the center loss and the weighted regularization triplet loss, the model is forced to learn the cross-modal consistent identity feature representation. Combined with the adversarial training of conditional GAN and the L1 pixel loss constraint, the visual privacy compliant image is generated while ensuring that the feature vector of the anonymized image and the original image remain approximately in the identity subspace. This technology breaks the unconscious change of identity features by traditional anonymization methods, and provides a bottom layer guarantee of identity invariance for the ReID task.

[0032] Thirdly, the application designs a dynamic optimization mechanism based on the dual indicators of privacy and performance. In the training process, the privacy protection effect is quantified by the peak signal-to-noise ratio (PSNR) and the structural similarity (SSIM), and the ReID performance is evaluated by the cumulative matching accuracy (CMC) and the average precision mean (mAP), forming a two-way supervision of the anonymization generator and the ReID model. By iteratively updating the desensitization strength of the supervised image and the model parameters, the optimal balance point of privacy protection and data utility is gradually approached, avoiding the performance drop caused by the "one-size-fits-all" desensitization in traditional methods, and realizing the adaptive adjustment to different application scenarios.

[0033] From the aspect of privacy protection effectiveness, the whole-body anonymization image generated by the application effectively removes sensitive biological features such as face and body (PSNR≤20, SSIM≤0.15), meets the requirements of strict privacy protection regulations such as GDPR, and blocks the risk of identity reconstruction from the source.

[0034] Moreover, the application has made a breakthrough in data utility. On benchmark datasets such as Market-1501, the ReID performance of anonymization images (Rank-1 accuracy improved by 7.2%-14.2%, mAP improved by 4.6%-24.8%) is significantly better than traditional desensitization methods, and even close to the recognition effect of original images, proving that identity-related semantic features such as clothing and body shape are effectively preserved.

[0035] In addition, the model of the application has strong compatibility for mixed input (original image and anonymization image), and the feature space distribution is highly consistent, supporting cross-modal re-identification tasks in complex scenarios, and providing a landing solution that takes into account privacy and efficiency for intelligent transportation, public security and other fields.

[0036] In summary, through technical architecture innovation and joint optimization mechanism, the application realizes the "win-win" of privacy protection and data value in the field of pedestrian re-identification for the first time, and has significant technical advancement and practical application value. BRIEF DESCRIPTION OF DRAWINGS

[0037] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description only some embodiments of the application, and for those skilled in the art, according to these drawings, other drawings can be obtained without creative labor, which still belong to the scope of the application.

[0038] Figure 1 An anonymization image generation process in an embodiment of the application is shown;

[0039] Figure 2 An identity protection process in an embodiment of the application is shown. Detailed Implementation

[0040] The preferred embodiments of the present invention will be described in detail below to provide a clearer understanding of the purpose, features, and advantages of the invention. It should be understood that the following embodiments are not intended to limit the scope of the invention, but are merely illustrative of the essential spirit of the technical solution of the invention.

[0041] In the following description, certain specific details are set forth for the purpose of illustrating various disclosed embodiments in order to provide a thorough understanding of the various disclosed embodiments. However, those skilled in the art will recognize that embodiments may be practiced without one or more of these specific details. In other instances, well-known techniques associated with this application may not have been shown or described in detail to avoid unnecessarily obscuring the description of the embodiments.

[0042] Throughout this specification, references to "an embodiment" or "an embodiment" indicate that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment. Therefore, the appearance of "in an embodiment" or "an embodiment" in various places throughout the specification does not necessarily refer to the same embodiment. Furthermore, a particular feature, structure, or characteristic may be combined in any manner in one or more embodiments.

[0043] Person re-identification (ReID) is a specialized retrieval task designed to identify the same person across non-overlapping camera views or different time instances from the same camera. Building a ReID system involves five key steps: raw data acquisition, bounding box generation, training data annotation, model training, and re-identification inference. However, privacy risks persist throughout these stages. For example, during data acquisition, ReID data may be stolen from centralized storage systems; in distributed learning frameworks, attackers can reconstruct private training data by sharing gradients.

[0044] This invention addresses ReID scenarios by integrating encryption / decryption architecture with a custom cryptographic protocol, solving privacy issues at the infrastructure level while enabling seamless access to protected data. The innovation lies in breaking down the "privacy-utility" dichotomy, achieving synergistic optimization between the two through technological integration, and providing a feasible solution for ReID applications in scenarios with high privacy requirements.

[0045] The following content will elaborate on the anonymized image generation method for person re-identification (ReID) according to this invention. For example... Figure 1 and Figure 2 As shown, the framework of this invention comprises two key components.

[0046] Part One ( Figure 1) Focus on anonymized image generation: use the image-to-image translation capability of conditional adversarial networks to generate anonymized images in a learnable way. In addition, to improve the ReID performance, the supervised images are progressively optimized based on the dual indicators of privacy protection and ReID effectiveness of the anonymized output.

[0047] The second part ( Figure 2 ) solves the identity preservation problem: excavates the internal relationship of all images of the same identity, so that the images of the same identity can maintain consistent visual features and similar appearance characteristics while protecting privacy.

[0048] The above framework modularizes its arrangement, that is, the present application constructs an anonymization framework, which includes an anonymized image generation module and an identity preservation module.

[0049] The method for generating privacy-compliant images of the present application is limited to image-to-image translation technology, which converts input images into a specific output form through a learnable mapping. The present application aims to convert the original input image into a protected format, using blurred, pixelated or noise-enhanced images as the initial supervision signal to guide the generation of privacy-protected images.

[0050] Specifically, the present application uses the pix2pix image translation framework based on generative adversarial networks (GAN) to achieve this goal.

[0051] The training samples contain a set of original images labeled The set of supervised images is denoted as ( which can be further updated).

[0052] The goal of the anonymized generator H X is to learn the mapping H: X p → Y. To this end, the generator H X is trained adversarially with the discriminator D Y : H X strives to generate images H i (x X ) that approximate y i , while D Y tries to distinguish between real supervised images y i and generated images H X (x i ).

[0053] The adversarial loss function is defined as:

[0054]

[0055] where n is the number of batch training samples, H X minimizes this objective, while the discriminator D Y maximizes it.​

[0056] In addition, in H X (x i ) and y i , the L1 loss is introduced between x i and y i , which ensures that the learned mapping converts the input x Y to the desired output y X .

[0057] The total loss function of the anonymization generator combines the above components:

[0058]

[0059] Where λ1 is a hyperparameter used to reduce the artifacts of the generated image.

[0060] In the above scheme, the core design of the framework of the present application is to realize the privacy-utility balance through the cooperation of the double components. Figure 1 The anonymization image generation module of the present application is shown in the above scheme. Based on the image translation technology (such as pix2pix) of conditional generative adversarial network (GAN), the present application maps the original image to the anonymization image, and the core is to replace the traditional “physical disturbance” (such as fixed blur kernel) with “learnable privacy transformation”, so that the anonymization process has both privacy protection and semantic reservation ability.

[0061] The initial supervised image is generated by using the traditional desensitization method (blurring, pixelization, noise enhancement processing), which is used as the “baseline signal” for privacy protection, guiding the generator to learn the image transformation that meets the minimum privacy requirement, and at the same time, through subsequent iterative optimization (based on the dual indicators of privacy and ReID), the supervision signal is dynamically adjusted to avoid over-desensitization or information loss.

[0062] The generator training mechanism of the present application is to use the joint constraint of adversarial loss and pixel loss.

[0063] Formula (1) embodies the method of the present application on the adversarial loss, and the discriminator D Y forces the generator H X to generate an anonymization result that is visually close to the real supervised image (such as a blurred image), ensuring that sensitive information (such as faces and body details) is effectively blurred or disturbed, meeting the visual requirements of privacy protection. The loss embodies the core idea of “adversarial learning”, that is, the game between the generator and the discriminator promotes the generated image to visually approach the “real sample” that meets privacy compliance, avoiding unnatural artifacts caused by artificial disturbance.

[0064] Formula (2) embodies the L1 pixel loss. The difference between the image generated from the pixel-level constraint and the supervised image reduces the loss of high-frequency information (such as clothing texture and body contour) caused by the adversarial training in the generation process, and ensures that the anonymized image retains sufficient structural information to provide usable semantic features for the subsequent ReID model. The λ1 hyperparameter balances the weights of the adversarial loss and the pixel loss, avoiding excessive pursuit of visual authenticity at the expense of structural similarity, or vice versa.

[0065] Compared with the prior art, the traditional desensitization (such as a fixed blur kernel) is a "one-size-fits-all" approach that cannot adapt to the importance of different image features (such as a stronger blur for a front view of a pedestrian and more contour information for a side view of a pedestrian), resulting in the deletion of key identity features (such as clothing color and carried items).

[0066] The present application generates an end-to-end learning of conditional GAN, and the generator can dynamically adjust the anonymization strategy according to the specific features (such as posture and viewing angle) of the input image, while protecting sensitive areas (such as the face) and retaining non-sensitive but identity-related features (such as shirt stripes and backpack shape).

[0067] On the other hand, the "upgradability" of the supervised image is the core design, that is, the initial supervision signal provides a bottom-line constraint for privacy protection, and subsequent performance feedback (such as reduced desensitization intensity when recognition accuracy decreases) is used to achieve a dynamic balance between privacy protection and data utility, breaking the fixed trade-off between "privacy-utility" in traditional methods.

[0068] In addition, the above scheme and the prior art have the following key differences. Existing image translation techniques (such as CycleGAN) are mainly used for style transfer or cross-domain image generation, and do not explicitly introduce "privacy protection indicators" and "identity invariance constraints", and the generated image may change the identity of the pedestrian (such as replacing the clothing color), which is not suitable for ReID tasks.

[0069] The supervision signal of the present application anchors the privacy protection target (initially a traditional desensitized image), ensuring that the privacy compliance of the generated image is quantifiable (such as controllable PSNR and SSIM indicators); subsequent joint training with the identity preservation module ( Figure 2 ) that is "joint training of anonymization generator and ReID model", the generator is reversely constrained by the ReID loss function to avoid destroying identity-related features during the generation process, achieving "identity invariance under privacy protection".

[0070] The anonymous image generation method proposed in the above scheme realizes the key capabilities of "controllable privacy protection strength and adjustable semantic information reservation" through the learnable mapping of conditional GAN, double loss function constraint and dynamic supervision signal optimization. The core value lies in breaking the "rigid processing" limitation of traditional desensitization methods, providing high-quality anonymized image input for the subsequent identity reservation module, and jointly constituting the technical cornerstone of privacy-utility balance.

[0071] Further, the present application strictly analyzes the internal correlation between all images of the same identity. From the perspective of human vision, the images in the subset X p have similar contour features; from the perspective of machine vision, these images should produce the same recognition output. This double intuitive observation prompts us to construct an approximate identity-specific feature subspace for image generation.

[0072] The framework of the present application embeds the ReID model into a joint learning architecture, using the advanced method AGW in ReID research as the baseline model. Although anonymization is a privacy protection scheme in the ReID system, directly applying traditional blurring methods to desensitized images will severely reduce the recognition performance.

[0073] Therefore, the present application proposes a hybrid image batch (combining original images and anonymized images) to jointly train the ReID model and the anonymization generator. The loss function integrates the three losses commonly used in ReID tasks, namely identity classification loss (L id ), center loss (L c ) and weighted regularization triplet loss (L wrt ) for optimization.

[0074] To ensure the compatibility of the ReID model for the original scene and the privacy protection scene, the present application simultaneously inputs the original image and the anonymized image, so the total loss function of the ReID model is defined as:

[0075]

[0076] In summary, the final objective function of the ReID anonymization model for hybrid images is:

[0077]

[0078] Compared with the prior art, the design core of the above scheme is from "visual similarity" to "identity invariance" mechanism. From the human visual aspect, the images of the same identity pedestrian have consistent global features such as contour and body shape (such as the same height, clothing style), which are the key clues of ReID. From the machine vision aspect, in the ideal case, the images of the same identity should be mapped to the same feature vector by the ReID model, to ensure the consistency of recognition. Based on this, the application constructs an "identity-specific feature subspace", and requires that the feature vectors of the anonymized images and the original images remain close in the subspace, that is, the identity invariance.

[0079] The specific technical implementation core of the application is joint training and mixed input strategy. First, the advanced model AGW in the ReID field is adopted, which integrates three loss functions. Among them, the identity classification loss (L id ) forces the feature vector to correctly classify the identity, ensuring the class separability; the center loss (L c ) reduces the distance between the feature vector of the same identity and the class center, enhancing the intra-class compactness; and the weighted regularization triplet loss (L wrt ) expands the inter-class distance between the feature vectors of different identities through difficult sample mining, improving the discriminability.

[0080] When performing mixed batch training, the original image and the anonymized image are input into the ReID model to force the model to learn the cross-modal consistent feature representation. The total loss simultaneously constrains the feature extraction of the original image and the anonymized image, ensuring that the anonymization operation does not destroy the identity-related features (such as clothing color and shape of carried objects).

[0081] The physical meaning of formula (3) is that the ReID model requires the feature extraction results of the original image x and the anonymized image H X (x) to meet the requirements of identity classification, intra-class compactness and inter-class separability, that is, the feature vectors of the two are highly approximate in the identity subspace. If the feature of the anonymized image deviates from the original feature, will increase, and the back propagation will prompt the anonymization generator H X to adjust the generation strategy to avoid destroying the identity features.

[0082] The physical meaning of formula (4) is that the total objective function integrates the anonymization generator loss (privacy protection oriented) and the ReID model loss (identity preservation oriented), forming a two-way constraint, to ensure that the anonymized image meets the privacy visual requirements (such as blurred face); and ensure that the features of the anonymized image are compatible with the original image, maintaining the ReID performance.

[0083] Therefore, the present application "binds" the ReID model with the anonymization generator for training, so that the generator "perceives" the importance of identity features (such as automatically retaining clothing patterns and only blurring faces) while protecting privacy. By mixing the input, the model is forced to adapt to the feature distribution difference between the original and anonymized images, avoiding the "domain shift" problem and improving the robustness of the model.

[0084] The present application explicitly introduces an "identity preservation constraint" to guide the anonymization process in the reverse direction through the ReID loss function, forming a closed-loop optimization of "privacy protection-feature preservation"; and uses the internal correlation of the same identity image to construct a feature subspace constraint, so that the anonymization operation has an "identity perception" ability rather than blind disturbance.

[0085] The identity preservation module is the key pillar of the present application to achieve "privacy-utility balance". By jointly training the ReID model and the anonymization generator, and combining the collaborative optimization of multiple loss functions, the anonymized image is ensured to protect privacy visually while its feature vector remains highly consistent with the original image in the identity subspace. This design breaks through the limitations of traditional anonymization methods that "prioritize privacy over recognition", providing a practical anonymization solution for the ReID task, especially suitable for public security scenarios that have strict requirements for both privacy and recognition accuracy.

[0086] The present application conducts experiments on widely used pedestrian re-identification (ReID) datasets. The Market-1501 (Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. 2015. Scalable person re-identification: A benchmark. In ICCV. 1116-1124.) dataset contains 32,668 annotated bounding boxes captured by 6 cameras. The dataset is divided into: 12,936 training images (750 identities), 3,368 query images (750 identities), and 15,913 gallery images (751 identities). The present application uses image quality indicators and re-identification indicators to evaluate the model. For privacy protection and recovery, the present application uses two widely used indicators: peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM). For ReID performance, the present application measures cumulative matching characteristics (CMC, i.e. Rank-k matching accuracy) and mean average precision (mAP).

[0087] The specific implementation details are as follows:

[0088] First, the original training set is divided into a new training set and a validation set according to a ratio of 4:1. Then, the validation set is further divided into a gallery set and a query set according to the same ratio. All performance indicators in the training process are obtained by evaluating the validation set.

[0089] In all experiments, the three models are jointly trained for 120 epochs with a batch size of 64. All input images are adjusted to 256x128 pixels and desensitized by blurring (12x12 kernel), pixelization (24x24 blocks), and adding noise enhancement N(0,0.5) with a mean of 0 and a variance of 0.5. The results are shown in Table 1.

[0090] Table 1. Comparison of ReID performance of common AGW models on protected images: "BASE" represents traditional desensitized images.

[0091]

[0092] The results show that the anonymization framework of the present application is adaptable to privacy protection mechanisms. In addition, the performance of the model is comparable to the baseline configuration, confirming its suitability for pedestrian re-identification tasks involving mixed image sets containing original images and processed images. This indicates that the feature space distribution of anonymized images maintains significant similarity with the feature space distribution of unmodified images. It is worth noting that the method of the present application performs better in privacy protection scenarios compared to cross-domain evaluation, which may be due to the remaining domain differences between the original image and the anonymized image modalities.

[0093] As can be seen from Table 1, the SSIM value between the original image and the protected image generated using desensitized images is higher than the image generated by the method of the present application, while the PSNR value is lower. It is inferred that the identity feature subspace assigns weight coefficients to the identity feature vectors, so the constructed mask can more effectively change the structure of the original image.

[0094] As shown in Table 2, compared with the baseline methods of blurring, pixelization, and adding noise, the Rank-1 accuracy is increased by 7.2% and the mAP is increased by 15.6% in terms of blurring robustness, indicating that the retrieval consistency is enhanced. In terms of pixelization resistance, the Rank-1 accuracy is moderately increased by 3.6%, and the mAP is increased by 4.6%, indicating that stable ranking matching can be maintained even under rough spatial quantization.

[0095] In terms of noise resistance, the Rank-1 accuracy is significantly increased by 14.2%, and the mAP is increased by 24.8% (from 50.6% to 75.4%), which is the largest absolute increase in all indicators.

[0096] Table 2. ReID performance evaluation on three ReID datasets: "BASE" denotes the AGW model trained on desensitized images. The table mainly shows two indicators: Rank-r accuracy (%) and mAP (%).

[0097]

[0098] In summary, the present application proposes a reversible method that balances the privacy and data utility of pedestrian images in pedestrian re-identification (ReID). This method creates anonymous full-body images while hardly changing the accuracy of ReID, and uses a step-by-step training process to improve performance. Comprehensive testing shows that this method effectively meets three key requirements: protecting privacy, restoring the original identity, and maintaining reliable ReID results even in the case of mixed normal images and anonymized images.

[0099] It should be noted that the above content relates to the specific implementation of the present application, and the above content which is not described in detail is the known technology of those skilled in the art.

[0100] The above shows and describes the basic principles and main features of the present application and the advantages of the present application. Those skilled in the art should understand that the present application is not limited by the above examples, and the above examples and descriptions in the specification are only to illustrate the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A pedestrian re-identification anonymization and privacy protection method that maintains identity invariance, characterized in that, Includes the following steps: An anonymization framework is constructed, which includes an anonymized image generation module and an identity preservation module; The desensitized image generated by traditional methods is used as the initial privacy protection supervision, and the original pedestrian image is converted into a full-body anonymized image through an anonymization generator; The identity preservation module mines the inherent relationships between images with the same identity, preserves identity-invariant features during the anonymization process, and jointly trains the anonymization generator and the pedestrian re-identification (ReID) model to minimize the impact of anonymization on ReID performance. The anonymized image generation module is built based on a conditional adversarial network (GAN) and learns the mapping relationship from the original image to the anonymized image through adversarial training. The identity preservation module ensures that the identity features of the anonymized image are consistent with the original image by jointly optimizing the identity-related loss of the ReID model. The anonymized image generation module uses the pix2pix framework, including a generator. and discriminator The generator Configured to use the original image Mapping to anonymized images The discriminator Configured to distinguish real supervised images With the generated anonymized image And it is trained using the following adversarial loss function: ； Where n is the number of training samples in the batch; The identity preservation module uses the AGW baseline model as the ReID model, and the loss function of the ReID model integrates the identity classification loss. Central loss and weighted regularized triplet loss The system is jointly trained by mixing the original input image with the anonymized image, and the total loss function is: ； in, The AGW loss of the original image, The AGW loss is used for anonymized images.

2. The pedestrian re-identification anonymization and privacy protection method that maintains identity invariance according to claim 1, characterized in that, The anonymized image generation module also introduces L1 loss to constrain the pixel-level differences between the generated image and the supervised image. The total loss function is: ； in, This is a hyperparameter used to adjust the pixel loss weight.

3. The pedestrian re-identification anonymization and privacy protection method that maintains identity invariance according to claim 1, characterized in that, Traditional desensitization methods include blurring, pixelation, or noise enhancement. The generated initial supervision image Y includes the desensitized image processed by the above methods and is used to guide the initial training of the anonymization image generation module.

4. The pedestrian re-identification anonymization and privacy protection method that maintains identity invariance according to claim 1, characterized in that, The preprocessing steps for the original pedestrian image include: adjusting the image size to 256×128 pixels, and generating training samples through data augmentation methods, including but not limited to random cropping, flipping, or color jittering.

5. The pedestrian re-identification anonymization and privacy protection method that maintains identity invariance according to claim 1, characterized in that, During the joint training process, the image generation loss of the anonymization generator and the identity recognition loss of the ReID model are optimized simultaneously. The overall objective function is: ； in, The total loss of the anonymization generator. This represents the total loss of the ReID model.

6. The pedestrian re-identification anonymization and privacy protection method that maintains identity invariance according to claim 1, characterized in that, The method evaluates its performance using the following metrics: Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) are used to measure the privacy protection effect, and Cumulative Matching Characteristics (CMC) and Mean Average Precision (mAP) are used to measure the pedestrian re-identification performance, ensuring that anonymized images achieve a balance between privacy protection and data utility.

7. An electronic device, characterized in that, include: At least one processor; A memory that is communicatively connected to the at least one processor; The memory stores a computer program that, when executed by the at least one processor, enables the at least one processor to implement the pedestrian re-identification anonymization privacy protection method that maintains identity invariance as described in any one of claims 1-6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed, it implements the pedestrian re-identification anonymization privacy protection method that maintains identity invariance as described in any one of claims 1-6.

Images