A noise pedestrian re-identification method based on feature fusion and label reliability

Abstract

The application discloses a noise pedestrian re-identification method based on feature fusion and label reliability, which adopts a trained network model to extract feature mapping of samples in a training set; obtains fusion features according to the feature mapping; enriches information contained by the fusion features and expands the attention range of the fusion features; screens out first-order K-neighbor images and second-order K-neighbor images of the samples according to the fusion features between the samples; obtains mixed labels through the labels of the first-order K-neighbor images and the labels of the second-order K-neighbor images, and then refines the labels of the samples through the mixed labels, so that the predicted labels and the network model are simultaneously optimized; and the optimized network model can improve the recognition accuracy under the condition of a data set containing noise labels.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision, in particular to a noise pedestrian re-identification method based on feature fusion and label reliability. BACKGROUND

[0002] Pedestrian re-identification is a sub-problem of image retrieval, aiming to match the images of the same pedestrian between multiple non-overlapping cameras. It has great industrial application value in security and surveillance. With the rapid development of deep convolutional neural networks, the pedestrian re-identification technology based on supervised learning has made great progress. However, most of the existing methods rely too much on a large number of annotated training data, but in actual industrial applications, label noise is inevitable due to incorrect samples or incorrect annotations. There are two types of label noise in the data collection process: annotation errors and sampling errors; annotation errors are mainly caused by human annotation errors due to indistinguishable appearance, low-resolution cameras, etc., and sampling errors are mainly caused by image abnormalities due to occlusion or difficulty in detecting people in the image. The above two types of labels will cause a sharp decline in the performance of the pedestrian re-identification method.

[0003] Deep learning with noisy labels has received increasing attention in image classification. Existing classification methods either filter out some noisy samples to reduce the impact of noisy samples, or use small loss selection methods to handle noisy labels, or rely on additional annotated datasets, but these methods are not suitable for pedestrian re-identification tasks with limited resources, where each identity contains only a small number of person images. In addition, some pedestrian re-identification methods for noisy labels have been proposed, such as treating each training sample as a Gaussian distribution, or proposing an online co-refinement framework to reduce the negative impact of noisy labels by using a beneficial selection consistency strategy. However, these methods rely on sample distribution assumptions and multiple networks, limiting the practicality of the instance retrieval task. SUMMARY

[0004] Therefore, it is necessary to provide a pedestrian re-identification method with high accuracy under the setting of containing noisy labels, specifically a noise pedestrian re-identification method based on feature fusion and label reliability, to solve the existing problems.

[0005] The present application provides a noise pedestrian re-identification method based on feature fusion and label reliability, which comprises:

[0006] S1: obtaining a training set with noisy labels; the labels in the training set with noisy labels include noisy labels; using a trained network model to extract the feature mapping of the samples in the training set;

[0007] S2: obtaining a predicted label based on the trained network model from the training set with noisy labels, and calculating a first loss function according to the predicted label;

[0008] S3: Obtain the fusion feature based on the feature mapping; find the first-order K-neighbor image of the sample and the second-order K-neighbor image of the first-order K-neighbor image based on the fusion feature between samples, and calculate the reliability score of the second-order K-neighbor image; take the label of the second-order K-neighbor image with the highest reliability score as the second label of the corresponding first-order K-neighbor image, and form a mixed label with the label of the first-order K-neighbor image.

[0009] S4: Calculate scores based on different mixed label scenarios using 5 evaluation modes, select the label with the highest score as the refined label, and replace the sample label with the refined label;

[0010] S5: Calculate the second loss function based on the refined label;

[0011] S6: Combine the first loss function and the second loss function to obtain the self-adjusting loss function; optimize the network model according to the self-adjusting loss function;

[0012] S7: Input the pedestrian image to be queried into the optimized network model and output the recognition result.

[0013] Preferably, in S1, the samples in the training set are obtained by randomly selecting a percentage of pedestrian images from the dataset and assigning them a different identity than the original image; the training set with noisy labels is denoted as:

[0014] ;

[0015] Where, x i y represents the i-th sample; i The label represents the i-th sample; N represents the number of pedestrian images.

[0016] Preferably, in S2, the training set with noisy labels is input into the trained network model, and the predicted probability of the sample is output; the label with the highest predicted probability is taken as the predicted label; the predicted label is denoted as: ;

[0017] ;

[0018] in, Indicates sample x i The probability of being classified into class y1; Indicates sample x i Divided into y C The probability of the class; y1 represents the first label; y C This represents the Cth type of label; C represents the number of label categories.

[0019] Preferably, in S2, an expression of the first loss function is:

[0020] ;

[0021] wherein L p represents the first loss function; N represents the number of pedestrian images; i represents the i-th; represents the probability of the sample x i being classified into a class; represents a predicted label.

[0022] Preferably, in S3, the process of obtaining the fusion feature is:

[0023] the feature map is evenly divided into K parts according to height; the maximum pooling value of each part is calculated to obtain a fine-grained feature of different attention spans; the feature map is globally averaged pooled to obtain a coarse-grained feature; each feature map is a 3D tensor of WxHxC; W represents width; H represents height; and C represents the number of label categories;

[0024] The fine-grained feature calculation formula is:

[0025] ;

[0026] wherein, represents the fine-grained feature of the k-th part; represents the fine-grained feature of the k-th part of the first classification; represents the fine-grained feature of the k-th part of the c-th classification; represents the feature map of the k-th part; s represents a region with the largest response in each part of the feature map, that is, the maximum value of a pixel point in each channel;

[0027] The coarse-grained feature calculation formula is:

[0028] ;

[0029] wherein, represents the coarse-grained feature; represents the global average pooling value of the WxH feature map of the first classification; represents the global average pooling value of the WxH feature map of the c-th classification; represents a WxH feature map set;

[0030] The coarse-grained feature is added to the fine-grained features of different attention spans to obtain the fusion feature; the calculation formula is:

[0031] ;

[0032] wherein, denotes the fusion feature; denotes the model parameter; x i denotes the i-th sample; y i denotes the label of the i-th sample.

[0033] According to the KNN algorithm, the first-order K-neighbor image of the sample is preferably taken as the sample with the closest fusion feature to the sample; the first-order K-neighbor image comprises

[0034] The expression of the first-order K-neighbor image is:

[0035] ;

[0036] wherein, denotes the set of first-order K-neighbor images; denotes the j-th first-order K-neighbor image; denotes the label of the j-th first-order K-neighbor image; according to the KNN algorithm, the second-order K-neighbor image corresponding to the first-order K-neighbor image is taken as the sample with the closest fusion feature to the first-order K-neighbor image; the second-order K-neighbor image comprises

[0037] The expression of the second-order K-neighbor image is:

[0038] ;

[0039] wherein, denotes the set of second-order K-neighbor images; denotes the l-th second-order K-neighbor image of the j-th first-order K-neighbor image; denotes the label of the l-th second-order K-neighbor image of the j-th first-order K-neighbor image.

[0040] According to the KNN algorithm, the second-order K-neighbor image of the sample is preferably taken as the sample with the closest fusion feature to the sample; the second-order K-neighbor image comprises

[0041] ;

[0042] wherein, denotes the reliability score between the label m of the first-order K-neighbor image and the label n of the second-order K-neighbor image, denotes the number of images in the second-order K-neighbor image with the label , is the Euclidean distance of the fusion feature, and the calculation formula is:

[0043] ;​​​​

[0044] wherein, denotes the jth first-order K-neighbor image; denotes the lth second-order K-neighbor image; denotes the fusion feature of the jth first-order K-neighbor image with label m, denotes the fusion feature of the lth second-order K-neighbor image with label n;

[0045] The reliability scores of the second-order K-neighbor images are sorted, and the label of the second-order K-neighbor image with the highest reliability score is selected as the second label of the corresponding first-order K-neighbor image, and the calculation formula is:

[0046]

[0047] wherein, denotes the label of the second-order K-neighbor image with the first ranked reliability score of the jth first-order K-neighbor image, denotes the reliability score between the label m of the first-order K-neighbor image and the label 1 of the second-order K-neighbor image, denotes the reliability score between the label m of the first-order K-neighbor image and the label of the second-order K-neighbor image; denotes the number of second-order K-neighbor images;

[0048] The second label of the first-order K-neighbor image and the label of the first-order K-neighbor image jointly form a hybrid label.

[0049] Preferably, in S4, the process of obtaining the refined label is: comparing the label of the sample with the hybrid label of the corresponding first-order K-neighbor image of the sample; summarizing the comparison conditions and setting different weights for each condition to obtain five evaluation modes, and the five evaluation modes are denoted as:

[0050]

[0051] wherein, denotes the weight; is a hyperparameter defined in the range of [0.1, 0.9]; t 11 , t 12 , t 21 , t 22 , t 23 respectively denote different comparison conditions; y i denotes the label of the ith sample; denotes the label of the first-order K-neighbor image; denotes the second label of the first-order K-neighbor image;

[0052] ​​According to different weights, the reliability scores of all first-order K-neighbor images corresponding to the sample are calculated, and the calculation formula is:

[0053] ;

[0054] wherein, represents the reliability score between the label b of the sample and the label of the jth first-order K-neighbor image; t represents the comparison condition; T represents the evaluation mode set; represents the reliability score between the label b of the sample and the label of the jth first-order K-neighbor image;

[0055] For the ith sample, all labels are enumerated, and the label with the highest reliability score is selected as the refined label of the ith sample , and the calculation formula is:

[0056] ;

[0057] wherein, represents the reliability score between the label b of the ith sample and the label of the first type; represents the reliability score between the label b of the ith sample and the label of the second type; represents the reliability score between the label b of the ith sample and the label of the Cth type.

[0058] Preferably, in S5, the expression of the second loss function is:

[0059] ;

[0060] wherein, L r represents the second loss function; N represents the number of pedestrian images; i represents the ith; represents the probability that the sample x i is divided into the Cth type; represents the refined label.

[0061] Preferably, in S6, the expression of the self-adjusting loss function is:

[0062] ;

[0063] ;

[0064] wherein, L represents the self-adjusting loss function; is a decreasing weight parameter; L r represents the second loss function; L p represents the first loss function; represents the mean value of L r at the v-1th iteration.

[0065] Beneficial effects: the method extracts the feature mapping of the samples in the training set by using the trained network model; the fusion features are obtained according to the feature mapping; the information contained in the features is enriched and the attention range of the features is expanded through the fusion features; the first-order K-neighbor image and the second-order K-neighbor image of the sample are screened out according to the fusion features between the samples; the mixed label is obtained through the label of the first-order K-neighbor image and the label of the second-order K-neighbor image, and then the label of the sample is refined through the mixed label, so as to optimize the prediction label and the network model at the same time; the optimized network model can improve the recognition accuracy under the data set with noise labels. BRIEF DESCRIPTION OF DRAWINGS

[0066] The exemplary embodiments of this application will be more fully understood from the following description of various embodiments, taken together with the drawings. The drawings are intended to be illustrative, and not limiting of the application. Unless otherwise noted, like reference numerals in the figures indicate like elements or steps throughout the various drawings.

[0067] Figure 1 A flowchart of the method according to an exemplary embodiment of the present application is provided. DETAILED DESCRIPTION

[0068] Exemplary embodiments of the present disclosure will be described more fully hereinafter with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it is to be understood that the present disclosure can be embodied in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the present disclosure to those skilled in the art.

[0069] It should be noted that the technical terms or scientific terms used in the present application should be understood as their general meanings understood by those skilled in the art, unless otherwise specified.

[0070] In addition, the terms "first" and "second" and the like are used to distinguish different objects, and are not used to describe a particular order. Furthermore, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device including a series of steps or units is not limited to the listed steps or units, but can optionally include other steps or units not listed or can optionally include other steps or units inherent to the process, method, product or device.

[0071] The embodiments of the present application provide a noise pedestrian re-identification method based on feature fusion and label reliability, which will be described below with reference to the accompanying drawings.

[0072] Reference is made to Figure 1 , which shows a flowchart of a noise pedestrian re-identification method based on feature fusion and label reliability provided by some embodiments of the present application. As shown in the figure, the method can include the following steps:

[0073] S1: Obtain a training set with noisy labels; the labels in the training set with noisy labels include noisy labels; and a trained network model is used to extract feature maps of samples in the training set;

[0074] Specifically, the samples in the training set are obtained by randomly selecting a percentage (10% or 20% or 30%) of pedestrian images in the clean data set and assigning them different identities from the original images; the training set with noisy labels is denoted as:

[0075] ;

[0076] Where x i represents the i-th sample; y i represents the label of the i-th sample; and N represents the number of pedestrian images.

[0077] In this embodiment, the network model uses an IDE classification model, which treats pedestrian identities as different categories (labels) during training; treats pedestrian re-identification as an image classification task, and uses a soft-max cross-entropy loss function to optimize the network model; and treats pedestrian re-identification as an image retrieval task in the test phase. After training, the learned network model is used as a feature extractor, and cosine similarity is selected as the similarity measure.

[0078] S2: The training set with noisy labels is used to obtain predicted labels based on the trained network model, and a first loss function is calculated according to the predicted labels;

[0079] Specifically, the training set with noisy labels is input into the trained network model, and the predicted probability of the sample is output; the label with the highest predicted probability is taken as the predicted label; and the predicted label is denoted as: ;

[0080] ;

[0081] Where, represents the probability that the sample x i is classified as y1; represents the probability that the sample x i is classified as y C ; y1 represents the first label; y C represents the C-th label; and C represents the number of label categories.

[0082] The expression of the first loss function is:

[0083] ;

[0084] Among them, L p This represents the first loss function; N represents the number of pedestrian images; i represents the i-th image; Indicates sample x i Divided into The probability of a class; This indicates the predicted label.

[0085] S3: Obtain the fusion feature based on the feature mapping; find the first-order K-neighbor image of the sample and the second-order K-neighbor image of the first-order K-neighbor image based on the fusion feature between samples, and calculate the reliability score of the second-order K-neighbor image; take the label of the second-order K-neighbor image with the highest reliability score as the second label of the corresponding first-order K-neighbor image, and form a mixed label with the label of the first-order K-neighbor image.

[0086] Specifically, the feature map is divided into K parts according to its height; the max pooling value of each part is calculated to obtain fine-grained features with different attention spans; the feature map is then subjected to global average pooling to obtain coarse-grained features; each feature map is a W×H×C 3D tensor; W represents the width; H represents the height; and C represents the number of label categories.

[0087] The formula for calculating fine-grained features is:

[0088] ;

[0089] in, This represents the fine-grained features of the k-th part; This represents the fine-grained features of the k-th part of the first category; Represents the fine-grained features of the k-th part of the c-th category; represents the feature map of the k-th part; s represents the region with the largest response in the channel of each feature map part, that is, the maximum value of the pixel in a channel;

[0090] The formula for calculating coarse-grained characteristics is:

[0091] ;

[0092] in, Indicates coarse-grained characteristics; This represents the global average pooling value of the W×H feature map for category 1; This represents the global average pooling value of the W×H feature map for the c-th category; Represents the W×H feature map mapping set;

[0093] The coarse-grained features are added to the fine-grained features of different attention spans to obtain the fused features; the calculation formula is as follows:

[0094] ;

[0095] in, Indicates fusion characteristics; Indicates model parameters; x i Represents the i-th sample; y i This represents the label of the i-th sample.

[0096] According to the KNN algorithm, the feature closest to the sample among the fused features is selected. Each sample is used as a first-order K-neighborhood image; the first-order K-neighborhood image includes indivual;

[0097] The expression for a first-order K-neighborhood image is:

[0098] ;

[0099] in, Represents the set of first-order K-neighborhood images; This represents the j-th first-order K-neighbor image; The label represents the j-th first-order K-neighbor image;

[0100] According to the KNN algorithm, the feature closest to the first-order K-neighbor image is selected from the fused features. Each sample serves as a second-order K-neighbor image corresponding to a first-order K-neighbor image; the second-order K-neighbor image includes... indivual;

[0101] The expression for the second-order K-neighborhood image is:

[0102] ;

[0103] in, Represents the set of second-order K-neighborhood images; Let l represent the l-th second-order K-neighbor image of the j-th first-order K-neighbor image; This represents the label of the l-th second-order K-neighbor image of the j-th first-order K-neighbor image.

[0104] In this embodiment, the reliability score of the second-order K-neighbor image is calculated using the reciprocal of the Euclidean distance of the fused features. The formula for calculating the reliability score of the second-order K-neighbor image is as follows:

[0105] ;

[0106] in, The reliability score represents the difference between the label m of a first-order K-neighbor image and the label n of a second-order K-neighbor image. This indicates that the label in the second-order K-neighborhood image of this image is... Quantity, The Euclidean distance for the fused features is calculated using the following formula:

[0107] ;

[0108] in, This represents the j-th first-order K-neighbor image; This represents the l-th second-order K-neighbor image; Let m represent the fusion feature of the j-th first-order K-neighborhood image with label m. This represents the fusion feature of the l-th second-order K-neighborhood image with label n;

[0109] Sort the reliability scores of the second-order K-neighbor images, and select the label of the second-order K-neighbor image with the highest reliability score as the second label of the corresponding first-order K-neighbor image. The calculation formula is as follows:

[0110] ;

[0111] in, The label representing the second-order K-neighbor image that ranks first in reliability score for the j-th first-order K-neighbor image. The reliability score represents the difference between the label m of a first-order K-neighbor image and the label 1 of a second-order K-neighbor image. Let m represent the label of the first-order K-neighbor image and m represent the label of the second-order K-neighbor image. Reliability scores between; This represents the number of second-order K-neighborhood images;

[0112] The second label of the first-order K-neighbor image, together with the label of the first-order K-neighbor image, forms a mixed label.

[0113] S4: Calculate scores based on different mixed label scenarios using 5 evaluation modes, select the label with the highest score as the refined label, and replace the sample label with the refined label;

[0114] Specifically, the process of obtaining refined labels is as follows: The label of a sample is compared with the mixed label of its corresponding first-order K-neighborhood image; the comparison results are summarized, and different weights are assigned to each case, resulting in five evaluation modes, denoted as:

[0115] ;

[0116] in, Indicates the weight; t is a hyperparameter defined in the range [0.1, 0.9]. 11 t 12 t 21 t 22 t 23 These represent different comparison scenarios; y i The label represents the i-th sample; Labels representing the first-order K-neighborhood image; The second label represents the first-order K-neighborhood image;

[0117] Based on different weights, calculate the reliability score of all first-order K-neighborhood images corresponding to the sample. The calculation formula is as follows:

[0118] ;

[0119] in, The reliability score between the label b of the sample and the label of the j-th first-order K-neighborhood image is represented by t; t represents the comparison case; T represents the set of evaluation modes. The reliability score between the label b of the sample and the label of the j-th first-order K-neighborhood image is represented.

[0120] For the i-th sample, enumerate all labels and select the label with the highest reliability score as the refined label for the i-th sample. The calculation formula is:

[0121] ;

[0122] in, The reliability score between the label b of the i-th sample and the label of class 1 is represented by . The reliability score between the label b of the i-th sample and the label of the second class is represented by . Let b represent the reliability score between the label b of the i-th sample and the label of class C.

[0123] S5: Calculate the second loss function based on the refined label;

[0124] The expression for the second loss function is:

[0125] ;

[0126] Among them, L r This represents the second loss function; N represents the number of pedestrian images; i represents the i-th image; Indicates sample x i Divided into The probability of a class; This indicates a refined label.

[0127] S6: Combine the first loss function and the second loss function to obtain the self-adjusting loss function; optimize the network model according to the self-adjusting loss function;

[0128] The expression for the self-adjusting loss function is:

[0129] ;

[0130] ;

[0131] Where L represents the self-adjusting loss function; L is a decreasing weight parameter. r L represents the second loss function; p Represents the first loss function; L represents the value at the (v-1)th iteration. r The mean of L. As training progresses, L r The error will continue to decrease, while the reliability of the predicted labels will gradually increase, and the network model will also be optimized.

[0132] S7: Input the pedestrian image to be queried into the optimized network model and output the recognition result.

[0133] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application, and they should all be covered within the scope of the claims and specification of this application.

Claims

1. A noise pedestrian re-identification method based on feature fusion and label reliability, characterized in that, Comprise: S1: obtain a training set with noisy labels; the labels in the training set with noisy labels include noise labels; and a trained network model is used to extract feature maps of samples in the training set; S2: the training set with noisy labels is used to obtain predicted labels based on the trained network model, and a first loss function is calculated according to the predicted labels; S3: fusion features are obtained according to the feature maps; According to the fusion features between samples, a first-order K-neighbor image of a sample is found, and a second-order K-neighbor image of the first-order K-neighbor image is found, and the reliability score of the second-order K-neighbor image is calculated; the label of the second-order K-neighbor image with the highest reliability score is taken as the second label of the corresponding first-order K-neighbor image, and the label of the first-order K-neighbor image forms a mixed label; S4: five evaluation modes are obtained according to different situations of the mixed label, the label with the highest score is selected as the refined label, and the label of the sample is replaced by the refined label; The label of the sample is compared with the mixed label of the corresponding first-order K-neighbor image of the sample; the comparison is summarized, and different weights are set for each situation to obtain five evaluation modes, and the five evaluation modes are denoted as: wherein, denotes a weight value; is a hyper-parameter defined in the range [0.1, 0.9]; t 11 , t 12 , t 21 , t 22 , t 23 denote different comparison cases, respectively; y i denotes the label of the i th sample; denotes the label of the first-order K-neighbor image; denotes the second label of the first-order K-neighbor image; S5: a second loss function is calculated according to the refined label; S6: a self-adjusting loss function is obtained by combining the first loss function and the second loss function; and the network model is optimized according to the self-adjusting loss function; S7: input a to-be-queried pedestrian image into the optimized network model to output a recognition result. 2.The noise pedestrian re-identification method based on feature fusion and label reliability according to claim 1, characterized in that, In S1, the samples in the training set are obtained by randomly selecting a percentage of pedestrian images in the data set and assigning them different identities from the original images; the training set with noisy labels is denoted as: wherein, x i denotes the i th sample; y i denotes the label of the i th sample; N denotes the number of pedestrian images. 3.The noise pedestrian re-identification method based on feature fusion and label reliability according to claim 2, characterized in that, In S2, the training set with the noisy label is input to the trained network model, and the prediction probability of the sample is output. The label with the maximum prediction probability is taken as the prediction label. The prediction label is marked as: ; wherein, denotes a sample x i is classified into y 1 class; denotes a sample x i is classified into y C class; y 1 denotes a first label; y C denotes a C second label; C denotes the number of classes of labels.

4. The noise pedestrian re-identification method based on feature fusion and label reliability according to claim 3, characterized in that, In S2, the expression of the first loss function is: wherein, L p represents a first loss function; N represents a number of pedestrian images; i represents a first i represents a sample x i is divided into classes; represents a predicted label.​ 5. The noise pedestrian re-identification method based on feature fusion and label reliability according to claim 4, characterized in that, In S3, the process of obtaining the fusion features is: The feature maps are evenly divided into K parts according to the height; the maximum pooling value of each part is calculated to obtain fine-grained features with different attention spans; the feature maps are globally averaged to obtain coarse-grained features; each feature map is a 3D tensor with a size of WxHxC; W represents the width; H represents the height; and C represents the number of label categories; The fine-grained feature calculation formula is: wherein, represents the first k part of the fine-grained feature; represents the first classification first k part of the fine-grained feature; represents the first c classification first k part of the fine-grained feature; represents the feature map of the kth part; s represents a region with the maximum response in a channel of the feature map of each part, i.e., the maximum value of a pixel point in a channel. The coarse-grained feature calculation formula is: wherein, denotes coarse-grained features; denotes the global average pooling value of the 1st classification WxH feature map; denotes the global average pooling value of the 1st classification WxH feature map; c denotes the global average pooling value of the 1st classification WxH feature map; denotes the WxH feature map mapping set; The coarse-grained feature and the fine-grained feature with different attention spans are added to obtain the fusion feature; the calculation formula is: wherein, represents a fusion feature; represents a model parameter; x i represents a first i sample; y i represents a label of a first i sample.

6. The noise pedestrian re-identification method based on feature fusion and label reliability according to claim 5, characterized in that, In S3, according to the KNN algorithm, the sample nearest to the sample in the fusion feature is taken as a first-order K-neighbor image of the sample; the first-order K-neighbor image includes one sample; and one sample. The expression of the first-order K-neighbor image is: in, Represents the set of first-order K-neighborhood images; Indicates the first j A first-order K-neighbor image; Indicates the first j The labels of the first-order K-neighbor images are determined; according to the KNN algorithm, the fused features closest to the first-order K-neighbor images are selected. Each sample serves as a second-order K-neighbor image corresponding to a first-order K-neighbor image; the second-order K-neighbor image includes... indivual; The expression of the second-order K-neighbor image is: wherein, denotes a set of second order K-neighbor images; denotes the first j second order K-neighbor image of the first l first order K-neighbor image; denotes the first j second order K-neighbor image of the first l label of the first second order K-neighbor image.

7. The noise pedestrian re-identification method based on feature fusion and label reliability according to claim 6, characterized in that, In S3, the reliability score of the second-order K-neighbor image is calculated using the reciprocal of the Euclidean distance of the fusion features, and the reliability score calculation formula of the second-order K-neighbor image is: wherein, represents the reliability score between the label m of the first-order K-neighbor image and the label n of the second-order K-neighbor image, represents the number of the label m in the second-order K-neighbor image, is the Euclidean distance of the fusion feature, and the calculation formula is:​ wherein, denotes the jth first-order K-neighbor image; denotes the fusion feature of the jth first-order K-neighbor image with label m, denotes the fusion feature of the lth second-order K-neighbor image with label n; The reliability scores of the second-order K-neighbor images are sorted, and the label of the second-order K-neighbor image with the highest reliability score is taken as the second label of the corresponding first-order K-neighbor image, and the calculation formula is: wherein, a label representing a second order K-neighbor image ranked first in reliability score for the jth first order K-neighbor image, a reliability score between a label of the first order K-neighbor image being m and a label of the second order K-neighbor image being 1, a reliability score between a label of the first order K-neighbor image being m and a label of the second order K-neighbor image being ; denotes a number of second order K-neighbor images; The second label of the first-order K-neighbor image and the label of the first-order K-neighbor image form a mixed label.

8. The noise pedestrian re-identification method based on feature fusion and label reliability according to claim 7, characterized in that, In S4, the process of obtaining the refined label is: According to different weights, the reliability scores of all first-order K-neighbor images corresponding to the sample are calculated, and the calculation formula is: wherein, represents the reliability score between the label of the sample and the label of the first j order K-neighbor image; t represents the comparison case; T represents the set of evaluation modes; represents the reliability score between the label of the sample and the label of the first j order K-neighbor image; For the i-th sample, enumerate all labels, and select the label with the highest reliability score as the refined label of the i-th sample The calculation formula is: wherein, represents a reliability score between the label of the i-th sample being b and the 1st class label; represents a reliability score between the label of the i-th sample being b and the 2nd class label; represents a reliability score between the label of the i-th sample being b and the 1st class label; C class label.

9. The noise pedestrian re-identification method based on feature fusion and label reliability according to claim 8, characterized in that, In S5, the expression of the second loss function is: wherein, L r represents a second loss function; N represents a number of pedestrian images; i represents a first i represents a sample x i is divided into classes; represents a refined label.​ 10. The noise pedestrian re-identification method based on feature fusion and label reliability according to claim 9, characterized in that, In S6, the expression of the self-adjusting loss function is: wherein, L denotes a self-adjusting loss function; is a decreasing weight parameter; L r denotes a second loss function; L p denotes a first loss function; denotes a v at the -1th iteration L r the mean of.

Images