A re-identification processing method and device for image search

Abstract

The application discloses a re-identification processing method and device for picture search, relates to the technical field of big data analysis and application, and can solve the problem of long time consumption of the pedestrian re-identification processing method in the prior art. The specific technical scheme is as follows: firstly, an initial feature vector of a target re-identification data set is acquired, the initial feature vector comprises an N-dimensional vector, at least one human picture is included in the target re-identification data set, and N is greater than or equal to 1; then, the initial similarity between the initial feature vector and all feature vectors in a preset picture database is calculated, a first similarity result list is obtained, and the current storage capacity of the picture database is N_G; finally, it is judged whether the first similarity result list is all lower than a first preset threshold value; if yes, the initial feature vector is added to the picture database, and N_G is equal to N_G plus 1. The application is used for the re-identification processing of picture search.

Description

Technical Field

[0001] This invention relates to the field of big data analysis and application technology, specifically to a re-identification processing method and apparatus for image search. Background Technology

[0002] With the rapid development of computing power, artificial intelligence computer vision technology, especially deep learning technology, has developed and become widespread. In fields such as security, pedestrian re-identification technology is increasingly widely used. Currently, pedestrian re-identification mainly utilizes computer vision technology to determine the presence of a specific pedestrian in an image or video sequence. Given a monitored pedestrian image, it can achieve multi-angle, blind-spot-free monitoring and tracking across devices, combining with pedestrian detection / tracking technologies. Pedestrian re-identification tasks generally include steps such as pedestrian image extraction, pedestrian feature extraction, and pedestrian feature comparison. Post-processing methods for pedestrian re-identification can significantly improve the accuracy of the comparison results; commonly used methods include Re-rank, Query Expansion, and PCA. However, existing re-ranking methods such as Re-rank are very time-consuming and cannot be used for online comparison. Summary of the Invention

[0003] This invention provides a method and apparatus for re-identification processing in image search, which solves the problem of long processing time in existing pedestrian re-identification methods. The technical solution is as follows:

[0004] According to a first aspect of the present invention, a re-identification processing method for image search is provided, the method comprising:

[0005] Obtain the initial feature vector of the target re-identification dataset, wherein the initial feature vector includes an N-dimensional vector, and the target re-identification dataset includes at least one human figure image, wherein N≥1;

[0006] Calculate the initial similarity between the initial feature vector and all feature vectors in the preset image database to obtain a first similarity result list. The current storage size of the image database is N_G.

[0007] Determine whether all results in the first similarity list are below a first preset threshold;

[0008] If so, the initial feature vector is added to the image database, then N_G = N_G + 1.

[0009] The image re-identification processing method provided in this invention first obtains an initial feature vector of the target re-identification dataset. The initial feature vector includes an N-dimensional vector, and the target re-identification dataset includes at least one human-shaped image, where N≥1. Then, it calculates the initial similarity between the initial feature vector and all feature vectors in a preset image database to obtain a first similarity result list. The current storage size of the image database is N_G. Finally, it determines whether all elements in the first similarity result list are below a first preset threshold. If so, the initial feature vector is added to the image database, and N_G = N_G+1. The re-identification processing method provided by this invention can quickly identify images, significantly improving image processing efficiency.

[0010] As a further aspect of the present invention: before calculating the initial similarity between the initial feature vector and all feature vectors in a preset image database, the method further includes:

[0011] The image database is created in advance;

[0012] Add the first human-shaped image to the image database.

[0013] As a further aspect of the present invention: the initial feature vector of the target re-identification dataset includes:

[0014] Obtain the target re-identification dataset;

[0015] The target re-identification dataset is trained with a multi-neural network structure to extract the initial feature vector.

[0016] As a further aspect of the present invention: after determining whether all results in the first similarity result list are below a first preset threshold, the method further includes:

[0017] If not, then the top N1 feature vectors with the highest initial similarity results in the image database are denoted as {T_a}, and the similarity of element T_i in {T_a} is updated to obtain an updated similarity result list, wherein N1 < N_G;

[0018] Determine whether the entire updated similarity result list is below the second preset threshold;

[0019] If so, the initial feature vector is added to the image database, then N_G = N_G + 1.

[0020] As a further aspect of the present invention: the step of updating the similarity of element T_i in {T_a} to obtain an updated similarity result list includes:

[0021] Calculate the similarity between T_i and other feature vectors in the image database other than T_i, and obtain a second similarity result list;

[0022] Obtain the corresponding values ​​of the top N2 feature vectors with the highest results in the second similarity result list in the first similarity result list to obtain the third similarity result list, where N2 < N1;

[0023] The similarity between T_i and the initial feature vector is corrected to obtain the updated similarity.

[0024] The initial similarity corresponding to vector T_i in the first similarity result list is changed to the updated similarity to obtain the updated similarity result list.

[0025] As a further aspect of the present invention: after determining whether all results in the updated similarity result list are below a second preset threshold, the method further includes:

[0026] If not, calculate the updated feature vector based on the highest initial feature vector in the updated similarity result list;

[0027] The initial feature vector corresponding to the image database is updated to the updated feature vector.

[0028] As a further aspect of the present invention: the updated similarity is calculated using a first formula, which is:

[0029] x_ti_=w1×Mean(x3)+w2×x_ti;

[0030] Where x_ti_ represents the updated similarity; w1 and w2 are preset weighted weights, with values ​​between 0 and 1; x3 is the third similarity result list; and Mean is the average value. x_ti represents the similarity between T_i and the initial feature vector.

[0031] As a further aspect of the present invention: the updated feature vector is calculated using a second formula, which is:

[0032] p_new = w3 × tq + w4 × p;

[0033] Where p_new is the updated feature vector; w3 and w4 are preset weighted weights, with values ​​between 0 and 1; tq is the highest initial feature vector in the updated similarity result list in the image database; and p is the initial feature vector.

[0034] As a further aspect of the present invention: the initial similarity is calculated using the following third formula, wherein the third formula is:

[0035]

[0036] Where s is the initial similarity; A is the initial feature vector; B is the feature vector in the image database; Ai and Bi represent the components of A and B, respectively.

[0037] According to a second aspect of the present invention, a re-identification processing apparatus for image search is provided, comprising an acquisition module, a calculation module, a first judgment module, and an addition module.

[0038] The acquisition module is used to acquire the initial feature vector of the target re-identification dataset. The initial feature vector includes an N-dimensional vector. The target re-identification dataset includes at least one human figure image, and N≥1.

[0039] The calculation module is used to calculate the initial similarity between the initial feature vector and all feature vectors in the preset image database, and obtain a first similarity result list. The current storage size of the image database is N_G.

[0040] The first judgment module is used to determine whether all results in the first similarity result list are below a first preset threshold.

[0041] The adding module is used to add the initial feature vector to the image database when the judgment result of the first judgment module is yes, then N_G = N_G + 1.

[0042] The image re-identification processing apparatus provided in this invention includes an acquisition module, a calculation module, a first judgment module, and an adding module. The acquisition module acquires an initial feature vector from a target re-identification dataset. The initial feature vector includes an N-dimensional vector, and the target re-identification dataset includes at least one human-shaped image, where N ≥ 1. The calculation module calculates the initial similarity between the initial feature vector and all feature vectors in a preset image database, obtaining a first similarity result list. The current storage size of the image database is N_G. The first judgment module determines whether all elements in the first similarity result list are below a first preset threshold. When the judgment result of the first judgment module is yes, the adding module adds the initial feature vector to the image database, then N_G = N_G + 1. The re-identification processing method provided by this invention can quickly identify images, significantly improving image processing efficiency.

[0043] According to a third aspect of the present invention, a re-identification processing apparatus for image search is provided, the re-identification processing apparatus for image search including a processor and a memory, the memory storing at least one computer instruction, the instruction being loaded and executed by the processor to implement the steps performed in the re-identification processing method for image search described above.

[0044] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, the storage medium storing at least one computer instruction, the instruction being loaded and executed by a processor to perform the steps performed in the re-identification processing method for image search as described in any of the preceding claims.

[0045] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0046] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0047] Figure 1 This is a flowchart of a re-identification processing method for image search provided in an embodiment of the present invention;

[0048] Figure 2 This is a structural diagram of a re-identification processing device for image search provided in an embodiment of the present invention. Detailed Implementation

[0049] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of systems and methods consistent with some aspects of the invention as detailed in the appended claims.

[0050] This invention provides a re-identification processing method for image search, such as... Figure 1 As shown, it includes the following steps:

[0051] Step 101: Obtain the initial feature vector of the target re-identification dataset.

[0052] Specifically, the target re-identification dataset includes at least one human image, and the initial feature vector consists of an N-dimensional vector, where N≥1;

[0053] In one embodiment, obtaining the initial feature vector of the target re-identification dataset includes:

[0054] Obtain the target re-identification dataset;

[0055] The target re-identification dataset is trained with a multi-neural network structure to extract initial feature vectors.

[0056] In practical applications, the multiple neural network structure can be multiple ResNet, ResNext, Resnest, etc. As long as it involves multiple convolutional layers or multiple fully connected layers plus classification, the desired effect can be achieved. Specifically, this embodiment can use the ResNet50 neural network structure, which, after training on a pedestrian re-identification dataset, can be used to extract pedestrian features. The general implementation involves inputting a pedestrian image, processing it through the neural network, and outputting an N-dimensional vector representing the initial feature vector of the pedestrian.

[0057] Step 102: Calculate the initial similarity between the initial feature vector and all feature vectors in the preset image database to obtain the first similarity result list.

[0058] Specifically, the current storage size of the image database is N_G. The initial similarity is calculated using the following third formula:

[0059]

[0060] Where s is the initial similarity; A is the initial feature vector; B is the feature vector in the image database; Ai and Bi represent the components of A and B, respectively.

[0061] In one embodiment, before calculating the initial similarity between the initial feature vector and all feature vectors in a preset image database, the method further includes:

[0062] Pre-create an image database;

[0063] Add the first human-shaped image to the image database.

[0064] Specifically, image data is created in advance, which can be a gallery database. At this time, the database is empty, and the first human-shaped image is stored directly. Then, it is stored according to the re-identification method of the present invention.

[0065] Step 103: Determine whether all results in the first similarity list are below the first preset threshold.

[0066] In this embodiment, the first preset threshold can be set to 0.6.

[0067] Step 104: If the result of step 103 is yes, add the initial feature vector to the image database, then N_G = N_G + 1.

[0068] Specifically, if all results in the first similarity list are below the first preset threshold, it is considered that no match was found and that it is a new pedestrian. The pedestrian feature vector is then inserted into the image database, and the current storage size is incremented by one.

[0069] In one embodiment, after determining whether the entire list of first similarity results is below a first preset threshold, the method further includes:

[0070] If not, then denote the top N1 feature vectors with the highest initial similarity results in the image database as {T_a}, update the similarity of element T_i in {T_a}, and obtain the updated similarity result list, where N1 < N_G;

[0071] Determine whether the entire list of updated similarity results is below the second preset threshold;

[0072] If so, add the initial feature vector to the image database, then N_G = N_G + 1.

[0073] Specifically, the updated similarity result list represents the weighted similarity results between the initial feature vector and the image database. The second preset threshold can be set to 0.8. If all results in the updated similarity result list are below the second preset threshold, it is considered that no match was found, and it is a new pedestrian. The pedestrian's feature vector is then inserted into the image database, and the current storage size N_G is incremented by one.

[0074] In one embodiment, the similarity of element T_i in {T_a} is updated, and the resulting list of updated similarity results includes:

[0075] Calculate the similarity between T_i and other feature vectors in the image database besides T_i, and obtain a list of second similarity results;

[0076] Obtain the corresponding values ​​of the top N2 feature vectors with the highest results in the second similarity result list in the first similarity result list to obtain the third similarity result list, where N2 < N1;

[0077] The similarity between T_i and the initial feature vector is adjusted to obtain the updated similarity.

[0078] Change the initial similarity corresponding to vector T_i in the first similarity result list to the updated similarity to obtain the updated similarity result list.

[0079] In practical use, when not all results in the first similarity list are below the first preset threshold, the top N1 feature vectors with the highest similarity (N1 is much smaller than N_G) are taken from the image database and denoted as {T_a}. For example, N1 can be 3. The following loop is performed on the element T_i in {T_a}:

[0080] 1. Calculate the similarity between T_i and other elements in the image database except T_i. The calculation method is the same as the initial similarity calculation method in step 102, and obtain the second similarity result list.

[0081] 2. Take the top N2 feature vectors with the highest results in the second similarity result list. For example, N2 can be 2. Check the corresponding values ​​of these N2 feature vectors in the first similarity result list to obtain the third similarity result list.

[0082] 3. Correct the similarity between T_i and the initial feature vector to update the similarity;

[0083] 4. Update the similarity of vector T_i in the first similarity result list to the updated similarity.

[0084] In this embodiment, compared with existing re-ranking methods, the computational load is reduced from an N:N feature matrix to an n:N feature matrix, where n is much smaller than N. This method not only significantly improves the pedestrian re-identification and comparison results but also maintains ranking efficiency and speed, enabling real-time operation.

[0085] Specifically, the updated similarity is calculated using the first formula, which is:

[0086] x_ti-=w_1×Mean(x3)+w2×x_ti;

[0087] Where x_ti_ represents the updated similarity; w1 and w2 are preset weighted weights, with values ​​between 0 and 1; x3 is the third similarity result list; and Mean is the average value. x_ti is the similarity between T_i and the initial feature vector. In this embodiment, w1 can be 0.2, and w2 can be 0.8.

[0088] In one embodiment, after determining whether the entire updated similarity result list is below a second preset threshold, the above method further includes:

[0089] If not, calculate the updated feature vector based on the highest initial feature vector in the updated similarity result list;

[0090] Update the initial feature vector in the image database to the updated feature vector.

[0091] In this embodiment, if not all of the updated similarity results in the list are below the second preset threshold, it is considered that a match has been successfully made. The highest feature vector in the updated similarity results list is taken from the image database, the updated feature vector of the pedestrian is recalculated, and the corresponding initial feature vector in the image database is updated as the updated feature vector.

[0092] In one embodiment, the updated feature vector is calculated using a second formula, which is:

[0093] p_new = w3 × tq + w4 × p;

[0094] Where p_new is the updated feature vector; w3 and w4 are preset weighting weights, with values ​​between 0 and 1; tq is the highest initial feature vector in the updated similarity result list of the image database; and p is the initial feature vector. Specifically, w3 and w4 can both be set to 0.5.

[0095] The image re-identification processing method provided in this invention first obtains an initial feature vector of the target re-identification dataset. The initial feature vector includes an N-dimensional vector, and the target re-identification dataset includes at least one human-shaped image, where N≥1. Then, it calculates the initial similarity between the initial feature vector and all feature vectors in a preset image database to obtain a first similarity result list. The current storage size of the image database is N_G. Finally, it determines whether all elements in the first similarity result list are below a first preset threshold. If so, the initial feature vector is added to the image database, and N_G = N_G+1. The re-identification processing method provided by this invention can quickly identify images, significantly improving image processing efficiency.

[0096] Based on the above Figure 1 The image search re-identification processing method described in the corresponding embodiments is described below as a system embodiment of the present invention, which can be used to execute the method embodiment of the present invention.

[0097] This invention provides a re-identification processing device for image search, such as... Figure 2 As shown, the device includes an acquisition module 201, a calculation module 202, a first judgment module 203, and an addition module 204;

[0098] The acquisition module 201 is used to acquire the initial feature vector of the target re-identification dataset. The initial feature vector includes an N-dimensional vector. The target re-identification dataset includes at least one human figure image, and N≥1.

[0099] The calculation module 202 is used to calculate the initial similarity between the initial feature vector and all feature vectors in the preset image database, and obtain the first similarity result list. The current storage size of the image database is N_G.

[0100] The first judgment module 203 is used to judge whether the entire first similarity result list is lower than the first preset threshold.

[0101] Add module 204, which is used to add the initial feature vector to the image database when the judgment result of the first judgment module 203 is yes, then N_G = N_G + 1.

[0102] The image re-identification processing apparatus provided in this embodiment of the invention includes an acquisition module 201, a calculation module 202, a first judgment module 203, and an addition module 204. The acquisition module 201 acquires the initial feature vector of the target re-identification dataset, the initial feature vector being an N-dimensional vector, and the target re-identification dataset including at least one human-shaped image, where N≥1. The calculation module 202 calculates the initial similarity between the initial feature vector and all feature vectors in a preset image database, obtaining a first similarity result list, where the current storage size of the image database is N_G. The first judgment module 203 determines whether all elements in the first similarity result list are below a first preset threshold. When the judgment result of the first judgment module 203 is yes, the addition module 204 adds the initial feature vector to the image database, then N_G = N_G + 1. The re-identification processing method provided by this invention can quickly identify images, significantly improving image processing efficiency.

[0103] In one embodiment, the above-described apparatus further includes a creation module 205 for pre-creating an image database;

[0104] Add module 204, which is also used to add the first human-shaped image to the image database.

[0105] In one embodiment, the acquisition module 201 includes an acquisition unit 2011 and a training unit 2012;

[0106] Acquisition unit 2011 is used to acquire the target re-identification dataset;

[0107] Training unit 2012 is used to train a multi-neural network structure on the target re-identification dataset and extract initial feature vectors.

[0108] In one embodiment, the above-mentioned device further includes a processing module 206, which is used to, when the judgment result of the first judgment module 203 is negative, denote the top N1 feature vectors with the highest initial similarity results in the image database as {T_a}, update the similarity of element T_i in {T_a}, and obtain an updated similarity result list, wherein N1 < N_G;

[0109] The above-mentioned device also includes a second judgment module 207, which is further used to judge whether the updated similarity result list is all below a second preset threshold;

[0110] The addition module 204 is also used to add the initial feature vector to the image database when the judgment result of the second judgment module 207 is yes, then N_G = N_G + 1.

[0111] In one embodiment, the processing module 206 includes a calculation unit 2061, an acquisition unit 2062, a correction unit 2063, and a modification unit 2064;

[0112] The calculation unit 2061 is used to calculate the similarity between T_i and other feature vectors in the image database except for T_i, and obtain a second similarity result list;

[0113] The acquisition unit 2062 is used to acquire the corresponding values ​​of the top N2 feature vectors with the highest results in the second similarity result list in the first similarity result list, and obtain the third similarity result list, where N2 < N1;

[0114] The correction unit 2063 is used to correct the similarity between T_i and the initial feature vector to obtain the updated similarity.

[0115] The modification unit 2064 is used to change the initial similarity corresponding to vector T_i in the first similarity result list to the updated similarity, so as to obtain the updated similarity result list.

[0116] In one embodiment, the calculation module 202 is further configured to calculate the updated feature vector based on the highest initial feature vector in the updated similarity result list when the judgment result of the second judgment module 207 is negative.

[0117] The aforementioned device also includes an update module 208, used to update the corresponding initial feature vector in the image database to an updated feature vector.

[0118] In one embodiment, the updated similarity is calculated using a first formula, which is:

[0119] x_ti_=w_1×Mean(x3)+w2×x_ti;

[0120] Where x_ti_ represents the updated similarity; w1 and w2 are preset weighted weights, with values ​​between 0 and 1; x3 is the third similarity result list; Mean is the mean value; and x_ti represents the similarity between T_i and the initial feature vector.

[0121] In one embodiment, the updated feature vector is calculated using a second formula, which is:

[0122] p_new = w3 × tq + w4 × p;

[0123] Where p_new is the updated feature vector; w3 and w4 are preset weighted weights, with values ​​between 0 and 1; tq is the highest initial feature vector in the updated similarity result list in the image database; and p is the initial feature vector.

[0124] In one embodiment, the initial similarity is calculated using the following third formula:

[0125]

[0126] Where s is the initial similarity; A is the initial feature vector; B is the feature vector in the image database; Ai and Bi represent the components of A and B, respectively.

[0127] Based on the above Figure 1 In addition to the image search re-identification processing method described in the corresponding embodiments, another embodiment of the present invention also provides an image search re-identification processing device, which includes a processor and a memory. The memory stores at least one computer instruction, which is loaded and executed by the processor to implement the above-described method. Figure 1 The corresponding embodiment describes a re-identification processing method for image search.

[0128] Based on the above Figure 1 In addition to the image search re-identification processing method described in the corresponding embodiments, this invention also provides a computer-readable storage medium. For example, a non-transitory computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a CD-ROM, magnetic tape, a floppy disk, or an optical data storage device. This storage medium stores at least one computer instruction for executing the above-described... Figure 1 The re-identification processing method for image search described in the corresponding embodiments will not be repeated here.

[0129] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the following claims.

[0130] It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A re-identification processing method for image search, characterized in that, The method includes: Obtain the initial feature vector of the target re-identification dataset, wherein the initial feature vector includes an N-dimensional vector, and the target re-identification dataset includes at least one human figure image, wherein N≥1; Calculate the initial similarity between the initial feature vector and all feature vectors in the preset image database to obtain a first similarity result list. The current storage size of the image database is N_G. Determine whether all results in the first similarity list are below a first preset threshold; If so, the initial feature vector is added to the image database, then N_G = N_G + 1; If not, then the top N1 feature vectors with the highest initial similarity results in the image database are denoted as {T_a}, and the similarity of element T_i in {T_a} is updated to obtain an updated similarity result list, wherein N1 < N_G; Determine whether the entire updated similarity result list is below the second preset threshold; If so, the initial feature vector is added to the image database, then N_G = N_G + 1; The step of updating the similarity of element T_i in {T_a} to obtain an updated similarity result list includes: Calculate the similarity between T_i and other feature vectors in the image database other than T_i, and obtain a second similarity result list; Obtain the corresponding values ​​of the top N2 feature vectors with the highest results in the second similarity result list in the first similarity result list to obtain the third similarity result list, where N2 < N1; The similarity between T_i and the initial feature vector is corrected to obtain the updated similarity. The initial similarity corresponding to vector T_i in the first similarity result list is changed to the updated similarity to obtain the updated similarity result list; The updated similarity is calculated using a first formula, which is: ; Where x_ti_ is the updated similarity; w1 and w2 are preset weighted weights, with values ​​between 0 and 1; x3 is the third similarity result list, Mean is the mean value; and x_ti is the similarity between T_i and the initial feature vector.

2. The re-identification processing method for image search according to claim 1, characterized in that, Before calculating the initial similarity between the initial feature vector and all feature vectors in a preset image database, the method further includes: The image database is created in advance; Add the first human-shaped image to the image database.

3. The re-identification processing method for image search according to claim 1, characterized in that, The initial feature vector for obtaining the target re-identification dataset includes: Obtain the target re-identification dataset; The target re-identification dataset is trained with a multi-neural network structure to extract the initial feature vector.

4. The re-identification processing method for image search according to claim 1, characterized in that, After determining whether all results in the updated similarity result list are below a second preset threshold, the method further includes: If not, calculate the updated feature vector based on the highest feature vector in the updated similarity result list; The initial feature vector corresponding to the image database is updated to the updated feature vector.

5. The re-identification processing method for image search according to claim 4, characterized in that, The updated feature vector is calculated using a second formula, which is: ; Where p_new is the updated feature vector; w3 and w4 are preset weighted weights, with values ​​between 0 and 1; tq is the highest feature vector in the updated similarity result list in the image database; and p is the initial feature vector.

6. The re-identification processing method for image search according to claim 1, characterized in that, The initial similarity is calculated using the following third formula, which is: ； Where s is the initial similarity; A is the initial feature vector; B is the feature vector in the image database; Ai and Bi represent the components of A and B, respectively.

7. A re-identification processing device for image search, characterized in that, include: The module includes an acquisition module, a calculation module, a first judgment module, an addition module, and an update module. The acquisition module is used to acquire the initial feature vector of the target re-identification dataset. The initial feature vector includes an N-dimensional vector. The target re-identification dataset includes at least one human figure image, and N≥1. The calculation module is used to calculate the initial similarity between the initial feature vector and all feature vectors in the preset image database, and obtain a first similarity result list. The current storage size of the image database is N_G. The first judgment module is used to determine whether all results in the first similarity result list are below a first preset threshold. The adding module is used to add the initial feature vector to the image database when the judgment result of the first judgment module is yes, then N_G = N_G + 1; The update module is configured to, when the judgment result of the first judgment module is negative, denote the top N1 feature vectors with the highest initial similarity results in the image database as {T_a}, update the similarity of elements T_i in {T_a} to obtain an updated similarity result list, wherein N1 < N_G; determine whether all elements in the updated similarity result list are below a second preset threshold; if yes, add the initial feature vectors to the image database, then N_G = N_G + 1; wherein, updating the similarity of elements T_i in {T_a} to obtain an updated similarity result list... The similarity result list includes: calculating the similarity between T_i and other feature vectors in the image database (excluding T_i) to obtain a second similarity result list; obtaining the corresponding values ​​of the top N2 feature vectors with the highest results in the second similarity result list in the first similarity result list to obtain a third similarity result list, where N2 < N1; correcting the similarity between T_i and the initial feature vector to obtain an updated similarity; and changing the initial similarity corresponding to the T_i vector in the first similarity result list to the updated similarity to obtain the updated similarity result list. The updated similarity is calculated using a first formula, which is: ; Where x_ti_ is the updated similarity; w1 and w2 are preset weighted weights, with values ​​between 0 and 1; x3 is the third similarity result list, Mean is the mean value; and x_ti is the similarity between T_i and the initial feature vector.

Images