A cross-resolution pedestrian re-identification method based on trusted reference guidance

Abstract

The application discloses a cross-resolution pedestrian re-identification method of trusted reference booting, which does not take generating high-definition images as a main target, but establishes a search-driven trusted reference compensation mechanism, forms a candidate reference set from a reference image set, and then judges whether the candidate reference is reliable; only for the trusted reference, a local feature corresponding relationship between a search image and the trusted reference image is established, and a limited gate compensation is used to compensate for the missing identity clues of the search image; when the reference is unreliable, the basic identity search bypass is automatically rejected, thereby reducing the feature pollution caused by the false reference. Through the search closed-loop training, the reference source in the training stage is as close as possible to the test stage, that is, the current model representation is used to actually search the candidate reference from the training set, instead of always manually specifying the same identity reference, so that the network explicitly experiences three states of correct reference, false reference and no reliable reference, so as to reduce the reference distribution difference between the training and test stages.

Description

Technical Field

[0001] This invention relates to a trusted reference-guided method for cross-resolution pedestrian re-identification, belonging to the field of cross-resolution pedestrian re-identification technology. Background Technology

[0002] Person re-identification (ReID) refers to the technique of retrieving images of the same person from a database of images captured by other cameras, given a query image of a pedestrian from one camera. In real-world surveillance networks, there are significant differences in camera installation height, focal length, shooting distance, and bitrate. Distant targets often appear with low resolution, lack of detail, and indistinguishable local textures, while the database may simultaneously contain higher-resolution images of the same person with more complete appearance details. This resolution mismatch causes the query image to miss clothing textures, edge structures, and fine-grained identification clues, significantly increasing the difficulty of cross-camera matching.

[0003] Currently, pedestrian re-identification technologies mainly include the following three categories: The first category comprises super-resolution or reconstruction-driven cross-resolution re-identification methods. Works such as SING, CAD-Net, INTACT, MRJL, PS-HRNet, and resolution adaptive representations typically alleviate the lack of low-resolution information through image reconstruction, adversarial learning, distillation, dual-resolution modeling, or resolution adaptive representation. These methods mainly revolve around recovering or learning resolution-invariant features from the low-resolution single image itself, and usually do not directly utilize high-quality candidate images actually existing in the test image library as external identity references.

[0004] The second category involves cross-image interaction or matching methods between query images and images in the image library. For example, the Transformer-based image matching method (TransMatcher) studies cross-image attention and inter-image matching mechanisms between query images and images in the image library. A Chinese patent application (Publication No. CN116486433A, Publication Date 2023.07.25) discloses a cross-attention mechanism for cross-camera samples of the same identity during the re-identification training phase, and solves the problem of insufficient paired samples during the testing phase through cross-self-distillation. However, relying solely on the cross-attention mechanism is insufficient to fully address the reliability selection and utilization of real image library references during the testing phase.

[0005] The third category comprises coarse-to-fine retrieval, reordering, and query expansion methods. k-reciprocal reordering, neighbor-based query expansion, and attention / memory reordering methods utilize neighbor information from the initial ranking to optimize the final distance. A Chinese patent application, "Image Retrieval Method and System Based on Global and Local Feature Reordering" (Publication No. CN112966137A, Publication Date 2021.06.15), also discloses a process of first obtaining a Top-N list based on global features, and then reordering based on local feature matching. However, a coarse retrieval followed by a secondary sorting is insufficient to address the local detail compensation problem of low-resolution query images in cross-resolution scenarios.

[0006] The above technology has the following problems: (1) Most existing cross-resolution methods rely on low-resolution single images to recover or learn resolution-invariant representations themselves, without making full use of the high-resolution identity details that objectively exist in the map library during the deployment phase; (2) There is a significant risk in directly using a Top-1 or Top-K image library sample as a reference: the coarse search results may be incorrect. If texture injection is performed without reliability judgment, information such as incorrect clothing and local structure will be written into the query image features, causing reference pollution. (3) During the training phase, image-reference image pairs with the same identity can usually be manually constructed, while during the testing phase, the reference image comes from the model's own retrieval, and the two distributions are inconsistent. Existing solutions often cancel the cross-attention mechanism during testing by distillation, rather than explicitly learning how to select, reject, and use real image library references during testing; (4) Traditional query expansion and reordering methods mostly utilize neighborhood information at the global embedding or similarity level, making it difficult to establish a one-to-one compensation path for the lost local textures in low-resolution query images. (5) Simple scalar gating can only control the strength of reference information as a whole, and it is difficult to distinguish between compensable areas and rejected areas for different spatial locations or channels; (6) When there is no sufficiently reliable reference map, the existing fusion scheme lacks an explicit bypass mechanism, which can easily lead to negative gain in uncertain scenarios.

[0007] Therefore, in cross-resolution pedestrian re-identification, how to reliably select, evaluate and utilize real high-resolution identity details in the image library to spatially and channel-controllably compensate for the missing local information in low-resolution query images without introducing erroneous references, and maintain robustness when there are no reliable references, has become an urgent technical problem to be solved. Summary of the Invention

[0008] The technical problem to be solved by this invention is to provide a reliable reference-guided cross-resolution person re-identification method to solve the problem that low-resolution query images lack identity details and high-resolution information in the image library is difficult to be reliably utilized in cross-resolution person re-identification.

[0009] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A trusted reference-guided cross-resolution person re-identification method includes the following steps: Step 1: Divide the pedestrian re-identification benchmark dataset into a training set and a query set and image library set for the testing phase; preprocess the high-resolution images in the training set to obtain a set of preprocessed high-resolution reference images; at the same time, construct low-resolution query images using the high-resolution images in the training set. The low-resolution query images and the preprocessed high-resolution reference images together constitute the cross-resolution pedestrian re-identification training set. Step 2: Construct a cross-resolution pedestrian re-identification model, including a feature extraction module, a candidate reference construction module, a reference credibility evaluation module, a local cross-image compensation module, a security gating module, and a secondary retrieval module; The feature extraction module is used to extract intermediate local features and global identity features from the low-resolution query image, and also to extract intermediate local features and global identity features from each preprocessed high-resolution reference image. The candidate reference building module is used to perform a coarse search using global identity features to select a candidate reference set from the preprocessed high-resolution reference image set; The reference credibility assessment module is used to calculate the reference credibility of each candidate reference image in the candidate reference set, and to form a credible reference set by selecting candidate reference images with a reference credibility not lower than the reference acceptance threshold. The local cross-graph compensation module is used to establish a local cross-graph correspondence between the trusted reference images in the trusted reference set and the low-resolution query image, and to perform trusted weighted aggregation on the candidate compensation features generated by the trusted reference images in the trusted reference set according to the reference credibility and local correspondence determinism, so as to obtain the reference compensation features of the low-resolution query image. The security gating module is used to update the intermediate local features of the low-resolution query image with reference compensation features through amplitude-limited security gating, and input the updated intermediate local features into the back-end identity representation part of the feature extraction module to obtain the enhanced global identity features. The secondary retrieval module is used to perform a secondary fine retrieval using the enhanced global identity features to obtain the re-identification results; Step 3: Train the cross-resolution person re-identification model constructed in Step 2 using the cross-resolution person re-identification training set to obtain the trained model; Step 4: Input the query images from the test phase query set and the image library images from the image library set into the trained model to obtain the cross-resolution pedestrian re-identification results of the query images.

[0010] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects: 1. This invention uses low-resolution query images as the objects to be identified and high-resolution pedestrian images in the image library as potential references. Instead of performing pixel-level super-resolution reconstruction, it completes secondary retrieval in a unified feature space, including candidate reference retrieval, reference credibility judgment, local identity information compensation, security gating update, and enhanced representation. This reduces the interference of erroneous references on the representation of the query identity while utilizing high-quality identity information from the image library.

[0011] 2. This invention directly utilizes existing high-resolution image library information for reliable compensation in the feature space, eliminating the need for complex pixel-level super-resolution reconstruction and large-scale replacement of existing surveillance camera equipment. Furthermore, image library features generated during the testing phase can be pre-extracted and cached after the model parameters are fixed, reducing redundant feature calculations and thus lowering computational overhead during system deployment and operation. This invention can improve low-resolution pedestrian recognition performance while fully utilizing existing video surveillance infrastructure, reducing the time and labor costs associated with equipment upgrades, manual retrieval, and redundant calculations.

[0012] 3. This invention can be applied to practical scenarios such as intelligent security, public video surveillance, personnel retrieval, and pedestrian tracking. Addressing the issues of low image resolution and missing identity details in pedestrian images under long-distance monitoring conditions, it provides technical support for cross-camera personnel identification and target tracking. By more accurately retrieving target pedestrians from large-scale surveillance image databases, it helps improve the efficiency of personnel retrieval and identity verification in public places, providing beneficial support for applications such as smart cities, public safety, and personnel management in key areas. Attached Figure Description

[0013] Figure 1 This is a flowchart of a trusted reference-guided cross-resolution pedestrian re-identification method proposed in this invention; Figure 2 This is a network structure diagram of a trusted reference-guided cross-resolution person re-identification model proposed in this invention; Figure 3 This is a schematic diagram of the reference credibility assessment module proposed in this invention; Figure 4 This is a schematic diagram of the structure of the local cross-graph compensation module and the security gating module proposed in this invention; Figure 5 This is a flowchart of the training process for the cross-resolution pedestrian re-identification model proposed in this invention. Detailed Implementation

[0014] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0015] This invention addresses the problem of missing identity details in low-resolution query images during real-world surveillance scenarios, while high-resolution image information from image libraries is difficult to reliably utilize. It proposes a cross-resolution pedestrian re-identification method guided by trusted references. This method uses low-resolution query images as the object to be identified and high-resolution pedestrian images from the image library as potential references. Instead of pixel-level super-resolution reconstruction, it performs secondary retrieval within a unified feature space, including candidate reference retrieval, reference credibility assessment, local identity information compensation, security gating updates, and enhanced representation. This approach utilizes high-quality identity information from the image library while reducing interference from erroneous references on the identity representation of the query.

[0016] like Figure 1 As shown, the method of the present invention includes the following steps: Step 1: Divide the pedestrian re-identification benchmark dataset into a training set and a query set and image library set for the testing phase; preprocess the high-resolution images in the training set, construct low-resolution query images based on the original high-resolution images, and retain the high-resolution images as candidate reference sources for the training phase. The process involves reading a high-resolution image and its identification tag, resizing the image to a uniform input size, and preprocessing it with random flipping, random erasing, random cropping, and normalization. This preprocessed image serves as a reference. To simulate low-resolution degradation caused by a long-distance camera, the original high-resolution image is downsampled and then restored to a uniform input size, maintaining the image spatial dimensions while losing fine-grained texture, edge, and local structural information. Let the original high-resolution image be... The degradation rate is Low-resolution query images Represented as: , in, Indicates according to the multiplier Perform a downsampling operation. This represents an upsampling operation that restores a degraded image to the model's uniform input size. The reference image maintains a higher resolution or image quality, thus creating a cross-resolution retrieval condition between the low-resolution query image and the high-resolution reference image.

[0017] Step 2: Use the feature extraction module with shared parameters to extract the intermediate local features and global identity features of the low-resolution query image and the reference image respectively; The query image and reference image are input into a feature extraction module with shared parameters to obtain intermediate local features and global identity features in a unified feature space. The two types of features of the query image are denoted as follows: and ;No. The two types of features of the reference image are denoted as follows: and , This represents the reference image index. Intermediate local features retain a certain spatial structure for subsequent local correspondence and identity information compensation; global identity features are used for coarse retrieval and final similarity calculation. During the testing phase, intermediate local features and global identity features of all images in the test set are pre-calculated and cached, and this feature cache is constructed only from images in the image library to reduce redundant calculations.

[0018] Step 3: Complete the first coarse search based on global identity features, and select one or more highly similar images from the reference images to form a candidate reference set; A first coarse search is performed using basic global identity features to query images related to the first... The initial identity similarity between the two reference images is represented as follows: , in, Indicates the initial identity similarity. This indicates the calculation of cosine similarity. According to... The reference images are sorted from highest to lowest quality, and the top-K reference images are selected to form a candidate reference set. This indicates the number of candidate reference images. Unlike directly using the Top-1 search results as the determining reference, this invention only uses the coarse search results as references to be verified, and then continues to determine whether they contain reliable and usable high-resolution information of the same identity.

[0019] Step 4: Calculate the credibility of candidate references by combining identity similarity, image quality, reciprocal neighborhood relationship and lightweight local matching consistency, and reject low-credibility candidates based on credibility. The credibility of the candidate reference image is evaluated, for the first... For each candidate reference image, calculate its image quality or effective resolution evaluation value, the reciprocal neighborhood consistency between the query image and the candidate image, and the local correspondence consistency.

[0020] To measure the clarity of usable identity details in a candidate reference, in one specific implementation, an image quality evaluation value can be calculated based on the high-frequency edge response of the candidate reference image: , in, Indicates the first One candidate reference image, Represents the first in the Top-K candidate reference set One candidate reference image; This represents the Laplacian operator, used to extract edges and high-frequency responses in an image; This represents variance calculation. This represents a small positive number used to avoid zero denominators and improve numerical stability. The clearer the texture and edges in the candidate reference, the better. The larger the value, the richer the available identity details contained in the reference.

[0021] Building upon this, to avoid relying solely on the one-to-one similarity between the query image and a single candidate for judgment, this invention further utilizes the global feature neighborhood relationships formed during the coarse retrieval stage to evaluate the matching stability of the two. Let... Indicates sample In the front of the global identity feature space A set of nearest neighbor samples, when sample lie in The former Among the nearest neighbors, and At the same time located The former When considering two neighbors as mutually beneficial neighbors, then the sample... The set of mutually beneficial nearest neighbors is represented as: , Furthermore, query the image and the first The reciprocal neighborhood consistency among candidate reference images is represented as follows: , in, This indicates a low-resolution query image. Indicates the number of nearest neighbors. This indicates the number of elements contained in the set. The larger the value, the more similar the query image and the candidate reference image are in the global identity feature space.

[0022] In addition to the global neighborhood structure, this invention further utilizes the intermediate local features extracted in Step 2, unfolding them along the spatial dimension into local feature locations, to perform a lightweight local consistency evaluation on the Top-K candidate reference images. (Query image and the...) Local correspondence consistency between candidate reference images It can be represented as: , in, Represents the intermediate local features of the query image Local position indexes expanded along the spatial dimension. Indicates the first Intermediate local features of candidate reference images Local position indexes expanded along the spatial dimension. express The total number of local locations obtained after unfolding along the spatial dimension. express The total number of local locations obtained after unfolding along the spatial dimension. Indicates the Query number The feature vector corresponding to each local location Indicates the first The candidate reference image The feature vectors corresponding to each local location. For each local location in the query image, the local location with the highest similarity is found in the candidate references, and then the best matching results of all local locations in the query images are averaged. The larger the value, the more local regions in the query image can find stable corresponding identity information in the candidate reference.

[0023] No. The reference confidence of each candidate reference image is represented as follows: , in, Indicates the first The reference credibility of each candidate reference image This represents the credibility estimation function. This represents a normalization function that maps the output to the range of 0-1. This indicates the splicing or combination of multiple evaluation quantities.

[0024] Set reference acceptance threshold , will satisfy The candidate indexes form a trusted reference set If all candidate references are below the threshold, the reference compensation process enters a bypass state, and the global identity features of the image are directly queried using low resolution. Complete the search to avoid forcibly merging low-reliability references.

[0025] Step 5: Establish local cross-image correspondence between trusted references and query images, and selectively migrate and perform multi-reference trusted weighted aggregation of high-resolution identity information based on reference credibility and local correspondence certainty. Establish a local cross-image correspondence between the trusted reference image and the query image, and... and the Intermediate local features of a reliable reference image Expand into local feature sequences and And obtained through learnable linear mapping: , in, , and These represent the query matrix, key matrix, and value matrix, respectively. , and The corresponding learnable mapping parameters; the output channel dimension of the value feature and the intermediate local features of the query image. To ensure that the channel dimensions remain consistent, so that the subsequently generated candidate compensation features are consistent with... To integrate.

[0026] Further calculations are performed on the local correspondence matrix between the query image and the trusted reference, as well as the candidate compensation features: , in, This represents the local cross-graph correspondence matrix. The feature dimensions of the query matrix and the key matrix are represented. Indicates by the first Candidate compensation features obtained from a reliable reference image.

[0027] Considering that occlusion, pose changes, or mismatches may cause instability in some local correspondences, this invention calculates the determinism of local correspondences based on the concentration of the correspondence probability distribution: , in, Indicates the first The trusted reference image in the query Determinism of correspondence at local locations This represents the probability distribution corresponding to that position. Represents the information entropy function. Indicates the first Intermediate local features of a reliable reference image The total number of local positions after unfolding along the spatial dimension.

[0028] Information entropy is defined as: , in, Indicates the query image number The local location and the first The first reliable reference image The corresponding probabilities between local locations.

[0029] because After Softmax normalization, a probability distribution is formed, and its maximum information entropy is: Therefore, the more concentrated the corresponding probabilities, The larger the value, the more certain the cross-graph correspondence at that local location; the more dispersed the correspondence probability. The smaller the value, the less the impact of uncertain local matching on reference compensation.

[0030] When multiple reliable reference images exist, this invention controls the participation level of reference information at two levels: reference-level reliability and local-level corresponding reliability. First, based on the reliability of candidate references… Obtain the reference-level normalized weights: , in, Indicates the first Reference-level weights for a trusted reference image; Temperature parameter indicating the degree of smoothness in controlling the weight distribution; This represents the index of the trusted reference image in the trusted reference set.

[0031] Subsequently, with Control the first The overall level of participation of a trusted reference image, in order to Control the trusted reference image in the query image. The degree of participation at each local location is obtained as follows: , in, Indicates the query image number The final reference compensation features obtained at each local location; Indicates the first Candidate compensation features generated at this location from a trusted reference image; This represents a small positive number used to prevent the denominator from being zero and to improve numerical stability.

[0032] Therefore, only when both the overall reference is reliable and the current location corresponds to clear reference information will it be included in the final compensation feature with a higher weight.

[0033] Step 6: Update the aggregated reference compensation features to the query features through amplitude-limited safety gating to obtain the enhanced query identity representation; Aggregate the local correspondence determinism and reference confidence of all reliable reference images: , in, Indicates the query image number Determinism of aggregation at local locations, This represents the overall reference confidence level of all trusted reference images.

[0034] Based on the local features in the middle of the queried image, the aggregated compensation features, the feature differences between the two, and the aggregate deterministic generation of amplitude-limited safety gating: , in, Indicates security gating; Indicates the composition of each local location The aggregate compensation features of the composition; Indicates absolute difference on an element-by-element basis; Indicates by The locally deterministic information is broadcast along the channel dimension when needed. This represents a learnable gated mapping function.

[0035] Enhanced local features in the query image are obtained by combining the overall reference confidence: , in, This represents the enhanced local features in the middle of the query image. This represents element-wise multiplication. When the trusted reference image and the query image have high consistency in identity semantics and local structure, missing identity clues can be enhanced; when there is a conflict between the two or the overall credibility of the reference is low, security gating and... Together, we can reduce the magnitude of reference updates, thereby reducing the interference of unreliable information on the original identity representation.

[0036] The enhanced intermediate features are further input into the backend identity representation part to obtain global identity features for secondary retrieval: , in, This represents the enhanced global identity characteristics of the query; This represents the backend identity representation part of the feature extraction module, which is used to further perform feature mapping and global aggregation.

[0037] Step 7: Optimize the cross-resolution person re-identification model using a closed-loop retrieval training method. The model architecture is as follows: Figure 2 As shown, during the test, the final pedestrian re-identification results are output in the order of coarse search, trusted reference screening, trusted compensation, and secondary fine search; First, the trusted reference compensation process is not enabled; only the feature extraction module is trained to learn a stable cross-resolution identity representation. Then, same-identity references, different-identity references, difficult negative references, and empty references are introduced, and reference credibility assessment, local cross-graph compensation, and security gating processes are enabled to enable the model to learn trusted reference compensation, erroneous reference rejection, and bypass without reliable references. Furthermore, candidate references are constructed using the current model's own coarse retrieval results to ensure that the method of obtaining candidate references in the training phase is consistent with that in the actual testing phase.

[0038] The model employs a joint optimization of identity classification loss, triplet metric loss, reference credibility supervision loss, and erroneous reference security constraint loss. The overall loss function is expressed as: , Among them, identity classification loss Triples measure loss They are defined as follows: , .

[0039] For the Given candidate reference images, their credibility supervision labels are defined as follows: , Then refer to the credibility supervision loss for: .

[0040] Error reference safety constraint loss for: , in, and Used for basic identity discrimination learning; Used to monitor the credibility of candidate references; Used to limit changes in query representation caused by erroneous or low-trust references. This indicates the real identity label of the queried image. This represents the probability that the classifier predicts the query image belongs to a real person; and These represent the global identity features of samples with the same identity as the query image and samples with different identities, respectively. Represents the feature distance. This represents the triplet interval parameter. ; Indicates the first Credibility supervision labels for each candidate reference image; , and These represent the weight coefficients of the corresponding loss terms. When the trusted reference set is empty, the basic identity representation is directly used as a bypass, without performing reference compensation and security gating updates, and the erroneous reference security constraint loss is not calculated for the corresponding samples.

[0041] During the testing phase, a coarse search is first performed using the basic global identity features of the image to be queried and the global identity features in the image feature cache of the image library. If a reliable reference exists, the corresponding intermediate local features are read from the image cache of the image library. The enhanced global identity features are obtained by sequentially performing local cross-graph compensation, multi-reference aggregation, and security-gated update. . With the The secondary retrieval similarity between images in the image library is represented as: , in, This indicates the similarity between secondary searches. According to... The images in the image library are reordered and the final pedestrian re-identification results are output. If no reliable reference exists, the local compensation and security gating processes are skipped, and the global identity features of the image to be queried are used directly to complete the retrieval.

[0042] Therefore, this invention transforms high-resolution pedestrian images in the image library from traditional search targets into reference information sources that, after credibility verification, are used for feature compensation of low-resolution query images. This forms a closed-loop technical solution that connects coarse retrieval, reference credibility judgment, local cross-image identity information compensation, multi-reference credibility weighted aggregation, security gating update, and secondary retrieval. Simultaneously, through reference rejection, local correspondence determinism, and restricted update mechanisms, it reduces identity feature contamination caused by erroneous candidates and local mismatches. Thus, without relying on pixel-level super-resolution reconstruction, it helps improve the reliability of pedestrian identity representation and retrieval under low-resolution query conditions.

[0043] The implementation of the technical solution of the present invention will be further described in detail below with reference to the accompanying drawings.

[0044] S11: Preprocess the pedestrian images, uniformly adjust the images to 256×128 pixels, and construct query images (Query) of different resolutions through downsampling and upsampling methods, while the reference image (Gallery) maintains its original clarity.

[0045] S12: Input the query image and reference image into the feature extraction module with shared parameters to extract intermediate local features and global identity features.

[0046] S13: Calculate the cosine similarity between the global identity features of the query image and the reference images, sort them from high to low similarity, and select the top K reference images to form a candidate reference set.

[0047] S14: Evaluate the credibility of the candidate reference images obtained in step S13, such as... Figure 3As shown in the figure. For each candidate reference image, the corresponding reference credibility is obtained by combining its global identity similarity with the query image, candidate image quality, mutual neighborhood consistency, and local feature correspondence consistency.

[0048] Among them, global identity similarity measures the proximity between the candidate reference and the query image in the identity feature space; image quality measures the richness of usable identity details in the candidate reference; reciprocal neighborhood consistency determines the matching stability between the query image and the candidate reference in the nearest neighbor structure; and local correspondence consistency is calculated based on the correlation between the intermediate local features of the query image and the candidate reference. These information are fused and a confidence level ranging from 0 to 1 is obtained through a confidence estimation network.

[0049] S15: Filter candidate reference images based on reference credibility. Set a reference acceptance threshold and retain candidates with credibility not lower than the threshold as credible reference images; if all candidates are lower than the threshold, no subsequent reference compensation is performed, and the global identity features of the query image are used directly for retrieval.

[0050] S16: For the trusted reference images obtained in step S15, establish a local cross-image feature correspondence between the query image and the trusted reference images, such as... Figure 4 As shown, the intermediate features of the query image and the trusted reference image are expanded into a sequence of local features. The local features of the query image are used as the query, and the local features of the trusted reference image are used as the key and value. The local correlation between the two is calculated through cross-graph attention, and the corresponding candidate compensation features are obtained from the high-resolution trusted reference image.

[0051] Considering that occlusion, pose changes, and local region misalignment may lead to unreliable correspondences, this invention further calculates local correspondence certainty based on cross-graph attention distribution, giving higher compensation weights to clearly matched local regions and lower compensation weights to unclearly matched regions. When multiple reliable references exist, the compensation features generated by different references are weighted and aggregated according to the reference reliability and local correspondence certainty to obtain the final reference compensation features.

[0052] S17: The reference compensation features obtained in step S16 are fused into the intermediate local features of the query image using a secure gating method. The intermediate local features of the query image, the reference compensation features, and the feature differences between the two are input into a lightweight gating network, and the gating amplitude is limited by the tanh function.

[0053] When the reference information matches the features of the query image, the compensation level for the corresponding identity information is increased; when there is a significant conflict between the two, the fusion strength of the reference information is reduced; when the overall credibility of the reference is low, the compensation magnitude is reduced accordingly. After security-gated fusion, the intermediate features of the enhanced query image are obtained, and the global identity features of the enhanced query image are further extracted.

[0054] S18: Recalculate the similarity between the enhanced global identity features of the query image and the global identity features of the library images, and sort them from high to low similarity to obtain the final pedestrian re-identification result.

[0055] S19: In application scenarios with limited computing resources, a second-stage fine matching can be performed only on the first few candidate images obtained from the first coarse search, so as to reduce the computational overhead of the second search.

[0056] Figure 5 This paper demonstrates the closed-loop training process of the present invention. First, only the feature extraction module is trained to learn a stable cross-resolution identity representation. Then, same-identity references, different-identity references, difficult negative references, and empty references are constructed. The current model is then used to coarsely retrieve Top-K candidate references from the training reference images, performing reference credibility assessment, local cross-image compensation, multi-reference aggregation, and security-gated forward propagation. Model parameters are updated based on the joint loss, and the candidate retrieval and model optimization processes are repeated until the preset training termination condition is met. After training, the optimal model parameters are loaded, and candidate reference retrieval, credibility filtering, local feature compensation, security-gated fusion, and secondary retrieval are sequentially performed on the query images during the testing phase, outputting the final person re-identification result.

[0057] This invention achieves excellent recognition performance on two mainstream cross-resolution person re-identification datasets, MLR-Market-1501 and MLR-CUHK03. The comparison results with other cross-resolution person re-identification methods are shown in Table 1. Rank-1, Rank-5, and Rank-10 represent the proportions of the correct identity appearing in the 1st, top 5, and top 10 positions of the search results, respectively. Table 1. Accuracy comparison of this method with other cross-resolution person re-identification methods.

[0058] Based on the same inventive concept, embodiments of this application provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the aforementioned trusted reference-guided cross-resolution pedestrian re-identification method.

[0059] Based on the same inventive concept, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the aforementioned trusted reference-guided cross-resolution pedestrian re-identification method.

[0060] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0061] This invention is described with reference to flowchart illustrations of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each step in the flowchart, and combinations of steps in the flowchart, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, generate instructions for implementing the steps in the flowchart. Figure 1 A device for a function specified in one or more processes.

[0062] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 The function specified in one or more processes.

[0063] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 Steps of a specified function in one or more processes.

[0064] The above embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solutions based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

Claims

1. A reliable reference-guided cross-resolution person re-identification method, characterized in that, Includes the following steps: Step 1: Divide the pedestrian re-identification benchmark dataset into a training set and a query set and image library set for the testing phase; preprocess the high-resolution images in the training set to obtain a set of preprocessed high-resolution reference images; at the same time, construct low-resolution query images using the high-resolution images in the training set. The low-resolution query images and the preprocessed high-resolution reference images together constitute the cross-resolution pedestrian re-identification training set. Step 2: Construct a cross-resolution pedestrian re-identification model, including a feature extraction module, a candidate reference construction module, a reference credibility evaluation module, a local cross-image compensation module, a security gating module, and a secondary retrieval module; The feature extraction module is used to extract intermediate local features and global identity features from the low-resolution query image, and also to extract intermediate local features and global identity features from each preprocessed high-resolution reference image. The candidate reference building module is used to perform a coarse search using global identity features to select a candidate reference set from the preprocessed high-resolution reference image set; The reference credibility assessment module is used to calculate the reference credibility of each candidate reference image in the candidate reference set, and to form a credible reference set by selecting candidate reference images with a reference credibility not lower than the reference acceptance threshold. The local cross-graph compensation module is used to establish a local cross-graph correspondence between the trusted reference images in the trusted reference set and the low-resolution query image, and to perform trusted weighted aggregation on the candidate compensation features generated by the trusted reference images in the trusted reference set according to the reference credibility and local correspondence determinism, so as to obtain the reference compensation features of the low-resolution query image. The security gating module is used to update the intermediate local features of the low-resolution query image with reference compensation features through amplitude-limited security gating, and input the updated intermediate local features into the back-end identity representation part of the feature extraction module to obtain the enhanced global identity features. The secondary retrieval module is used to perform a secondary fine retrieval using the enhanced global identity features to obtain the re-identification results; Step 3: Train the cross-resolution person re-identification model constructed in Step 2 using the cross-resolution person re-identification training set to obtain the trained model; Step 4: Input the query images from the test phase query set and the image library images from the image library set into the trained model to obtain the cross-resolution pedestrian re-identification results of the query images.

2. The trusted reference-guided cross-resolution person re-identification method according to claim 1, characterized in that, In step 1, each high-resolution image and its corresponding identity label in the training set are read, the size of the high-resolution image is adjusted to a preset size, and preprocessing is performed by random flipping, random erasing, random cropping and standardization to obtain a set of preprocessed high-resolution reference images. Constructing a low-resolution query image using a high-resolution image, represented as: ， in, Indicates according to the multiplier Perform a downsampling operation. This indicates an upsampling operation that restores the degraded image obtained from downsampling to a preset size. Represents a high-resolution image. This indicates a low-resolution image query.

3. The trusted reference-guided cross-resolution person re-identification method according to claim 1, characterized in that, The specific process of constructing the candidate reference module is as follows: Calculate the first [image] in the set of low-resolution query images and preprocessed high-resolution reference images. Initial identity similarity between reference images: ， in, Indicates the initial identity similarity. This represents the global identity features of a low-resolution query image. Indicates the first Global identity features of a reference image This indicates the calculation of cosine similarity. according to The reference images are sorted from highest to lowest quality, and the top-K reference images are selected to form a candidate reference set. This indicates the number of candidate reference images.

4. The trusted reference-guided cross-resolution person re-identification method according to claim 1, characterized in that, The specific process of the reference credibility assessment module is as follows: The credibility of candidate reference images in the candidate reference set is evaluated. Reference confidence of each candidate reference image Represented as: ， in, This represents a normalization function that maps the output to the range of 0-1. This represents the credibility estimation function. This indicates a splicing or combination operation. Indicates the first Initial identity similarity of candidate reference images, , Indicates the number of candidate reference images; Indicates the first Image quality or effective resolution evaluation value of each candidate reference image Indicates low-resolution query image and the first Reciprocal neighborhood consistency of candidate reference images Indicates the first The local correspondence consistency of each candidate reference image is calculated using the following formula: ， ， ， in, , They represent the first , One candidate reference image, Represents the Laplace operator. This represents variance calculation. This represents a positive number used to prevent the denominator from being zero and to improve numerical stability; Represents a low-resolution query image reciprocal neighbor set Indicates the first A set of mutually beneficial nearest neighbors for each candidate reference image. Indicates the number of nearest neighbors; This indicates that the intermediate local features of the low-resolution query image will be retrieved. Local position index after unfolding along the spatial dimension. express The total number of local locations obtained after unfolding along the spatial dimension; Indicates the first Intermediate local features of candidate reference images Local position indexes expanded along the spatial dimension. express The total number of local locations obtained after unfolding along the spatial dimension; Indicates the low-resolution query image number. The feature vector corresponding to each local location Indicates the first The candidate reference image Feature vectors corresponding to local locations; Set reference acceptance threshold , will satisfy The candidate reference images form a reliable reference set; if the reference reliability of all candidate reference images is lower than 1, the reference reliability of all candidate reference images is lower than 1. If the compensation process is bypassed, the global identity features of the low-resolution query image are directly used. Search complete.

5. The trusted reference-guided cross-resolution person re-identification method according to claim 4, characterized in that, The specific process of the local cross-graph compensation module is as follows: Intermediate local features of low-resolution query images and the first in the trusted reference set Intermediate local features of a reliable reference image Expand into local feature sequences and And obtained through learnable linear mapping: ， in, , and These represent the query matrix, key matrix, and value matrix, respectively. , and These are the corresponding learnable mapping parameters; Calculate the local cross-image correspondence matrix and candidate compensation features between the low-resolution query image and the trusted reference image: ， in, This represents the local cross-graph correspondence matrix. This represents the feature dimensions of the query matrix and the key matrix. Indicates matrix transpose. Indicates by the first Candidate compensation features obtained from a reliable reference image; Calculate local correspondence determinism based on probability distribution: ， in, Indicates the first The trusted reference image in the low-resolution query image. Determinism of correspondence at local locations Indicates the first The probability distribution corresponding to each local location Represents the information entropy function; ， in, Indicates the low-resolution query image number. The local location and the first The first reliable reference image The corresponding probability between local locations; According to the The reliability of a reliable reference image Calculate the reference-level normalized weights: ， in, Indicates the first Reference-level normalized weights for a trusted reference image. The temperature parameter represents the degree of smoothness of the weight distribution. Represents a trustworthy reference set The trusted reference image index in For the first The reliability of a reliable reference image; by Control the first The overall level of participation of a trusted reference image, in order to Control the first The trusted reference image in the low-resolution query image. The degree of participation at each local location is used to obtain reference compensation features: ， in, Indicates the low-resolution query image number. The final reference compensation features obtained at each local location; Indicates the first The trusted reference image in the low-resolution query image. Candidate compensation features generated at local locations.

6. The trusted reference-guided cross-resolution person re-identification method according to claim 5, characterized in that, The specific process of the security gating module is as follows: Aggregate the local correspondence determinism and reference confidence of all trusted reference images in the trusted reference set: ， in, Indicates the low-resolution query image number. Determinism of aggregation at local locations, This represents the overall reference confidence level of all trusted reference images in the trusted reference set; Based on the intermediate local features of the low-resolution query image Aggregate compensation features and safety gating with limited amplitude generation for aggregate deterministic generation: ， in, Indicates security gating, Indicates that it consists of all local locations The aggregate compensation features of the composition Indicates absolute difference on an element-by-element basis. Indicates by The composition of local deterministic information, This represents a learnable gated mapping function. Represents the hyperbolic tangent function; The enhanced intermediate local features of the low-resolution query image are obtained by combining the overall reference confidence: ， in, This represents the intermediate local features of a low-resolution query image after enhancement. This represents element-wise multiplication; The enhanced intermediate local features of the low-resolution query image are input into the back-end identity representation part of the feature extraction module to obtain global identity features for secondary retrieval: ， in, This represents the global identity features after enhancement of a low-resolution query image. This indicates a feature normalization operation. This represents the backend identity representation part of the feature extraction module.

7. The trusted reference-guided cross-resolution person re-identification method according to claim 1, characterized in that, In step 3, the model training uses an identity classification loss term. Triples measure loss term Reference credibility monitoring loss item and error reference safety constraint loss item Perform joint optimization, the overall loss function Represented as: ， in, , and These represent the weighting coefficients of the corresponding loss terms.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the trusted reference-guided cross-resolution pedestrian re-identification method as described in any one of claims 1 to 7.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the trusted reference-guided cross-resolution pedestrian re-identification method as described in any one of claims 1 to 7.

Images