CN113505642A - Method, device, equipment and storage medium for improving target re-identification generalization - Google Patents
Method, device, equipment and storage medium for improving target re-identification generalization Download PDFInfo
- Publication number
- CN113505642A CN113505642A CN202110627547.4A CN202110627547A CN113505642A CN 113505642 A CN113505642 A CN 113505642A CN 202110627547 A CN202110627547 A CN 202110627547A CN 113505642 A CN113505642 A CN 113505642A
- Authority
- CN
- China
- Prior art keywords
- domain
- loss function
- field
- target
- sample image
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 52
- 230000002776 aggregation Effects 0.000 claims abstract description 27
- 238000004220 aggregation Methods 0.000 claims abstract description 27
- 230000006870 function Effects 0.000 claims description 70
- 238000012549 training Methods 0.000 claims description 43
- 238000012360 testing method Methods 0.000 claims description 14
- 238000013507 mapping Methods 0.000 claims description 7
- 238000004364 calculation method Methods 0.000 claims description 5
- 238000000605 extraction Methods 0.000 claims description 5
- 238000010586 diagram Methods 0.000 description 9
- 238000004590 computer program Methods 0.000 description 5
- 238000005259 measurement Methods 0.000 description 5
- 238000004891 communication Methods 0.000 description 4
- 230000009286 beneficial effect Effects 0.000 description 3
- 238000010606 normalization Methods 0.000 description 2
- 230000003287 optical effect Effects 0.000 description 2
- 230000008569 process Effects 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000000802 evaporation-induced self-assembly Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000012163 sequencing technique Methods 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a method, a device, equipment and a storage medium for improving the target re-identification generalization, wherein the method comprises the following steps: inputting sample image data of unknown fields into a pre-trained multi-field expert mixed model to obtain the exclusive characteristics of each field and the query characteristics of the sample images; calculating the average similarity of the query feature and the exclusive feature, and determining the correlation between the sample image and each field according to the average similarity; carrying out weighted aggregation on the exclusive characteristics of each field according to the field correlation to obtain aggregation characteristics; and performing target re-identification on the sample image according to the aggregation characteristics. According to the method for improving the target re-recognition generalization, the features of the known domain data are dynamically aggregated by utilizing the meta-learning and multi-domain expert mixed model to generate the features with strong generalization capability on the unknown domain data, so that the generalization of the target recognition domain is improved.
Description
Technical Field
The invention relates to the technical field of target identification, in particular to a method, a device, equipment and a storage medium for improving the generalization of target re-identification.
Background
Research on object re-identification has attracted extensive attention in both academic and industrial fields, with the objective of identifying the same object (e.g., pedestrian, vehicle, etc.) at different camera perspectives. The target re-identification has great application value in the security protection field and the construction of smart cities. Although good performance can be achieved when the target re-identification model is trained and tested on a data set in the same domain, performance can be significantly degraded when the model is applied directly to data sets in other domains for testing due to domain bias. In a real scene, data of a target domain is often difficult to collect and label, so that it is very important to improve the generalization capability of a target re-identification model.
In the prior art, methods for improving the target re-identification recognition generalization all follow the same framework: they combine all the active domain data sets into one data set, train a single model on it, and apply the trained model directly to the data test of unknown domain. Such a framework may present two potential problems: (1) learning common feature spaces of different fields without different treatment on each source field, neglecting the characteristics of each field, and using the characteristics to improve the generalization capability of the model; (2) the prior art often ignores the inherent correlation between the target domain data and the source domain data, which can limit the generalization capability of the model.
Disclosure of Invention
The embodiment of the disclosure provides a method, a device, equipment and a storage medium for improving target re-identification generalization. The following presents a simplified summary in order to provide a basic understanding of some aspects of the disclosed embodiments. This summary is not an extensive overview and is intended to neither identify key/critical elements nor delineate the scope of such embodiments. Its sole purpose is to present some concepts in a simplified form as a prelude to the more detailed description that is presented later.
In a first aspect, an embodiment of the present disclosure provides a method for improving target re-recognition generalization, including:
inputting sample image data of unknown fields into a pre-trained multi-field expert mixed model to obtain the exclusive characteristics of each field and the query characteristics of the sample images;
calculating the average similarity of the query feature and the exclusive feature, and determining the correlation between the sample image and each field according to the average similarity;
carrying out weighted aggregation on the exclusive characteristics of each field according to the field correlation to obtain aggregation characteristics;
and carrying out target re-identification on the sample image according to the aggregation characteristics.
In an alternative embodiment, the multi-domain expert hybrid model includes a backbone network module, a voting network module, and a plurality of domain expert network modules, wherein the voting network module and the plurality of domain expert network modules branch after the backbone network module.
In an optional embodiment, inputting sample image data of an unknown domain into a pre-trained multi-domain expert mixed model to obtain a specific feature of each domain and a query feature of a sample image, including:
inputting sample image data of an unknown field into a pre-trained multi-field expert mixed model;
extracting shared characteristics among different fields through a backbone network module;
mapping the shared characteristics through each domain expert network module to obtain the exclusive characteristics of each domain;
and mapping the shared features through a voting network module to obtain the query features of the sample image.
In an optional embodiment, before inputting the sample image data of the unknown domain into the pre-trained multi-domain expert mixed model, the method further comprises:
and training the multi-field expert mixed model in a meta-learning training mode.
In an optional embodiment, training the multi-domain expert mixed model by a meta-learning training mode comprises:
at the beginning of each training iteration, randomly dividing K source domain training data into K-1 meta training domains and 1 meta testing domain;
calculating a domain loss function and a relation alignment loss function by using the meta-training domain data, updating parameters of a backbone network module and parameters of a plurality of domain expert network modules by using the domain loss function, and updating parameters of a voting network module according to a first-order gradient of the relation alignment loss function;
calculating a domain loss function and a relationship alignment loss function again by using the meta-test domain data;
updating parameters of the backbone network module and parameters of a plurality of domain expert network modules according to the domain loss function obtained by recalculation;
updating parameters of the voting network module according to the second-order gradient of the relation alignment loss function obtained by recalculation;
and (5) iteratively training the meta training domain and the meta testing domain for multiple times until the network converges.
In an alternative embodiment, the domain loss function is comprised of a metric loss function and a decorrelation loss function, the metric loss function being comprised of a classification loss function, a triplet loss function, and a center loss function.
In an optional embodiment, the relationship alignment loss function is obtained by computing a binary cross entropy loss of domain specific feature and aggregated feature metric relationships.
In a second aspect, an embodiment of the present disclosure provides an apparatus for improving generalization of target re-recognition, including:
the characteristic extraction module is used for inputting sample image data of unknown fields into a pre-trained multi-field expert mixed model to obtain the exclusive characteristic of each field and the query characteristic of the sample image;
the correlation calculation module is used for calculating the average similarity of the query feature and the exclusive feature and determining the correlation between the sample image and each field according to the average similarity;
the characteristic aggregation module is used for carrying out weighted aggregation on the exclusive characteristic of each field according to the field correlation to obtain an aggregation characteristic;
and the target re-identification module is used for carrying out target re-identification on the sample image according to the aggregation characteristics.
In a third aspect, an embodiment of the present disclosure provides an apparatus for improving target re-recognition generalization, including a processor and a memory storing program instructions, where the processor is configured to execute the method for improving target re-recognition generalization provided in the foregoing embodiment when executing the program instructions.
In a fourth aspect, the present disclosure provides a computer-readable medium, on which computer-readable instructions are stored, where the computer-readable instructions are executable by a processor to implement a method for improving target re-recognition generalization provided in the foregoing embodiments.
The technical scheme provided by the embodiment of the disclosure can have the following beneficial effects:
according to the method for improving the target re-recognition generalization provided by the embodiment of the disclosure, the special characteristics of each field can be reserved by the multi-field expert mixed model, so that the characteristics of multiple fields can be complemented to form the characteristics with generalization capability; the relevance between the unknown target domain and the existing multiple source domains can be calculated in a self-adaptive mode through a voting network module in the model, the multi-domain exclusive characteristics are aggregated through the relevance, the characteristics of the known source domains are aggregated for the unknown target domain in a self-adaptive mode, and therefore the discrimination of the characteristics of the unknown domains is improved. According to the method, the relation between the known domain and the unknown domain can be well simulated through the meta-learning training mode, and the generalization of target recognition is further improved. Compared with the prior art, the target re-recognition model has larger generalization improvement.
It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention, as claimed.
Drawings
The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and together with the description, serve to explain the principles of the invention.
FIG. 1 is a flow diagram illustrating a method for improving target re-identification generalization, according to an exemplary embodiment;
FIG. 2 is a flow diagram illustrating a method for training a multi-domain expert hybrid model in accordance with an exemplary embodiment;
FIG. 3 is a block diagram illustrating a multi-domain expert hybrid model in accordance with an exemplary embodiment;
FIG. 4 is a diagram illustrating a training process for a multi-domain expert hybrid model in accordance with an exemplary embodiment;
FIG. 5 is a block diagram illustrating an apparatus for improving generalization of target re-identification, according to an exemplary embodiment;
FIG. 6 is a diagram illustrating an apparatus for improving target re-recognition generalization, according to an example embodiment;
FIG. 7 is a schematic diagram illustrating a computer storage medium in accordance with an exemplary embodiment.
Detailed Description
The following description and the drawings sufficiently illustrate specific embodiments of the invention to enable those skilled in the art to practice them.
It should be understood that the described embodiments are only some embodiments of the invention, and not all embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
When the following description refers to the accompanying drawings, like numbers in different drawings represent the same or similar elements unless otherwise indicated. 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 certain aspects of the invention, as detailed in the appended claims.
In the description of the present invention, it is to be understood that the terms "first," "second," and the like are used for descriptive purposes only and are not to be construed as indicating or implying relative importance. The specific meanings of the above terms in the present invention can be understood in specific cases to those skilled in the art. In addition, in the description of the present invention, "a plurality" means two or more unless otherwise specified. "and/or" describes the association relationship of the associated objects, meaning that there may be three relationships, e.g., a and/or B, which may mean: a exists alone, A and B exist simultaneously, and B exists alone. The character "/" generally indicates that the former and latter associated objects are in an "or" relationship.
The method for improving the target re-recognition generalization provided by the embodiment of the present application will be described in detail below with reference to fig. 1 to 4. Referring to fig. 1, the method specifically includes the following steps.
S101, inputting sample image data of unknown fields into a pre-trained multi-field expert mixed model to obtain the exclusive characteristics of each field and the query characteristics of the sample images.
In a possible implementation mode, sample image data in an unknown field is obtained, and the sample image data in the unknown field is input into a multi-field expert mixed model which is trained in advance to perform feature extraction. The sample image data of the unknown domain refers to domain data that is not found in the training set.
As shown in fig. 4, the multi-domain expert hybrid model in the embodiment of the present disclosure includes a backbone network module, a voting network module, and a plurality of domain expert network modules, where the voting network module and the plurality of domain expert network modules are connected as branches after the backbone network module.
Specifically, sample image data of an unknown field is input into a pre-trained multi-field expert mixed model, the sample image data is input into a backbone network module, and the backbone network module can extract shared features of different fields. Further, each domain expert network module is connected as a branch behind the backbone network module, and the domain sharing features are mapped to the measurement space exclusive to each domain. For each sample, domain-specific features with characteristics of each domain are available through all domain expert networks. As shown in fig. 3, the source domain a has a corresponding domain a expert network, the source domain B has a corresponding domain B expert network, the source domain C has a corresponding domain C expert network, and the data set of each domain has its own model for extracting the domain-specific features having the characteristics of each domain, and the number of the domain expert networks corresponds to the number of the source domain types one to one.
Furthermore, the voting network module is also connected behind the backbone network module as a branch, and the shared features are mapped by the voting network module to obtain the query features of the sample image. The output feature dimensions of the voting network are the same as those of the domain expert network.
S102, calculating the average similarity of the query feature and the exclusive feature, and determining the correlation between the sample image and each field according to the average similarity.
Specifically, for a sample image in an unknown domain, the correlation between the sample image and each domain is obtained by calculating the average similarity between the query features extracted by the voting network module and the exclusive features of each domain. The exclusive feature of each domain is obtained by calculating the central feature of each domain category. And calculating and averaging the similarity of the query feature of each sample and all the exclusive features in each domain, and taking the obtained average similarity as the correlation between the sample and each domain.
S103, carrying out weighted aggregation on the exclusive characteristics of each field according to the field correlation to obtain aggregation characteristics.
In one possible implementation, the correlation between the sample and each domain is used as a weight, and each domain-specific feature is weighted and aggregated, for example, by feature concatenation or feature addition, to obtain an aggregated feature.
And S104, carrying out target re-identification on the sample image according to the aggregation characteristics.
In one possible implementation, the known domain data features are dynamically aggregated, and an aggregated feature with strong generalization capability on the unknown domain data is generated, which can be used for final target re-identification of the unknown target domain.
Specifically, sample images of all candidate targets are input into a trained model to obtain weighted aggregation features corresponding to each target, and the features are subjected to L2 normalization. And inputting any interested target picture into a trained model to obtain a weighted aggregation feature, performing L2 normalization on the feature, calculating cosine similarity between the feature and the features of all targets, sequencing the cosine similarity from large to small, and taking the candidate target with the largest cosine similarity with the interested target picture as the identified final target.
According to the method for improving the target re-recognition generalization, the multi-domain expert mixed model can reserve the exclusive characteristics of each domain, so that the characteristics of a plurality of domains can be complemented to form the characteristics with more generalization capability; the relevance between the unknown target domain and the existing multiple source domains can be calculated in a self-adaptive mode through a voting network module in the model, the multi-domain exclusive characteristics are aggregated through the relevance, the characteristics of the known source domains are aggregated for the unknown target domain in a self-adaptive mode, and therefore the discrimination of the characteristics of the unknown domains is improved. The recognition capability of the target re-recognition model in the unknown field is greatly improved.
In an optional embodiment, before inputting the sample image data of the unknown domain into the pre-trained multi-domain expert mixed model, training the multi-domain expert mixed model by a meta-learning training mode is further included.
Fig. 2 is a schematic diagram illustrating a training method of a multi-domain expert hybrid model according to an exemplary embodiment, where as shown in fig. 2, the training method includes:
s201, at the beginning of each training iteration, randomly dividing K source domain training data into K-1 element training domains and 1 element testing domain;
s202, calculating a domain loss function and a relation alignment loss function by using meta-training domain data, updating parameters of a backbone network module and parameters of a plurality of domain expert network modules by the domain loss function, and updating parameters of a voting network module according to a first-order gradient of the relation alignment loss function;
s203, recalculating the domain loss function and the relationship alignment loss function by using the meta-test domain data;
s204, updating parameters of the backbone network module and parameters of a plurality of domain expert network modules according to the domain loss function obtained by recalculation;
s205, updating parameters of the voting network module according to the second-order gradient of the relationship alignment loss function obtained through recalculation;
s206, the meta training field and the meta testing field are iteratively trained for multiple times until the network converges.
The training method of the multi-domain expert mixed model is described in detail below with reference to fig. 4.
Specifically, as shown in fig. 4, at the beginning of each training, K source domain data are subjected to random domain division to obtain K-1 meta-training domains and 1 meta-testing domain.
The data of the K domain in the K source domain data passes through a backbone network FψExtracting to obtain a domain sharing feature, wherein the domain sharing feature passes through K domain expert networks
The mapping of (a) obtains K domain-specific features, the domain-shared features passing through the voting network QθA query feature is obtained.
Further, a domain loss function and a relation alignment loss function are calculated by using the meta-training domain data, wherein the domain loss function is composed of a measurement loss function and a decorrelation loss function, and the measurement loss function is composed of a classification loss function, a triplet loss function and a center loss function.
In particular, a metric loss function L is applied to the kth domain-specific features of the kth domain samplemetricWherein the metric loss function is represented by the classification loss LclsTriple loss LtriAnd central loss LcentThe method comprises the following steps:
Lmetric＝Lcls+Ltri+λLcent
where λ is a weighting coefficient, set to 0.0005 in one optional embodiment.
In a possible implementation mode, a classification loss function is calculated, the sample image of each field is subjected to backbone network extraction sharing characteristics, the sample class classification probability value of the sample image is obtained through mapping of the expert network of the field, and the classification probability value and the cross entropy loss function of the sample label are calculated to obtain the classification loss function.
In a possible implementation manner, a triple loss function is calculated, shared features of each field sample image are extracted through a backbone network, the field-specific features of the sample are obtained through mapping of the field expert network, and the triple loss function is calculated for the field-specific features of each batch of training samples.
In one possible implementation, a central loss function is calculated, which is calculated by minimizing the distance of the sample domain specific features from the class central features of each domain.
In one possible implementation, a decorrelation loss function is calculated, which is obtained by minimizing the correlation between the kth domain-specific feature and the remaining K-1 non-related domain-specific features, the correlation being obtained by calculating the L2 norm after multiplication of the kth domain-specific feature with the corresponding element of each of the remaining non-related domain-specific features.
Further, a domain loss function L is calculated according to the obtained metric loss function and decorrelation loss functiondomain：
Ldomain＝Lmetric+Ldecor
And updating parameters of the backbone network module and parameters of each domain expert network module through the obtained domain loss function.
Computing a relational alignment loss function L for data of each of K-1 meta-training fieldsrelationAnd updating the parameters of the dynamic voting network module by calculating the first-order gradient of the relation alignment loss function.
Wherein the relationship alignment loss function LrelationThe k-th domain exclusive feature and the weighted aggregation feature metric relation are calculated to obtain the k-th domain exclusive feature and the weighted aggregation feature metric relation. The measurement relation of the characteristics is obtained by calculating the proportional relation between the positive sample distance and the sum of the positive sample distance and the negative sample distance, and the measurement relation can reflect the discriminant relation of the samples in the high-dimensional characteristic space. The metric relationship of the features is specifically calculated as follows:
wherein v denotes a kth-domain-specific feature of the kth-domain sample, v+To representPositive sample characteristics, v, of samples in a small batch of training data-Representing negative example features of examples in a small batch of training data.
Further, the domain loss function and the relation alignment loss function are calculated again by using the meta-test domain data, the parameters of the backbone network module and the parameters of the plurality of domain expert network modules are updated according to the field loss function obtained by calculation again, and the parameters of the voting network module are updated according to the second-order gradient of the relation alignment loss function obtained by calculation again.
And (5) carrying out iterative training on the element training domain and the element testing domain for multiple times until the network converges to obtain the trained multi-field expert mixed model.
According to the embodiment of the disclosure, through a training mode of meta-learning, the relationship between the known domain and the unknown domain can be well simulated, and further the generalization of the model is improved.
The embodiment of the present disclosure further provides a device for improving the generalization of target re-identification, where the device is configured to execute the method for improving the generalization of target re-identification in the foregoing embodiment, as shown in fig. 5, the device includes:
the feature extraction module 501 is configured to input sample image data in an unknown domain into a pre-trained multi-domain expert mixed model, so as to obtain an exclusive feature of each domain and a query feature of a sample image;
a correlation calculation module 502, configured to calculate an average similarity between the query feature and the dedicated feature, and determine a correlation between the sample image and each field according to the average similarity;
a feature aggregation module 503, configured to perform weighted aggregation on the dedicated feature of each field according to the field correlation, so as to obtain an aggregated feature;
and the target re-identification module 504 is configured to perform target re-identification on the sample image according to the aggregation features.
According to the device for improving the generalization of the target re-recognition, the meta-learning and multi-domain expert mixed model is utilized to dynamically aggregate the characteristics of the known domain data to generate the aggregated characteristics with strong generalization capability on the unknown domain data, so that the generalization of the target recognition domain is improved.
It should be noted that, when the apparatus for improving the target re-recognition generalization provided in the foregoing embodiment executes the method for improving the target re-recognition generalization, only the division of the functional modules is illustrated, and in practical applications, the functions may be distributed by different functional modules according to needs, that is, the internal structure of the device may be divided into different functional modules, so as to complete all or part of the functions described above. In addition, the apparatus for improving the generalization of the target re-identification provided in the above embodiments and the method embodiment for improving the generalization of the target re-identification belong to the same concept, and details of the implementation process are found in the method embodiment and are not described herein again.
The embodiment of the present disclosure further provides an electronic device corresponding to the method for improving the target re-recognition generalization provided in the foregoing embodiment, so as to execute the method for improving the target re-recognition generalization.
Please refer to fig. 6, which illustrates a schematic diagram of an electronic device according to some embodiments of the present application. As shown in fig. 6, the electronic apparatus includes: the processor 600, the memory 601, the bus 602 and the communication interface 603, wherein the processor 600, the communication interface 603 and the memory 601 are connected through the bus 602; the memory 601 stores a computer program that can be executed on the processor 600, and the processor 600 executes the method for improving the target re-recognition generalization provided by any of the foregoing embodiments when executing the computer program.
The Memory 601 may include a high-speed Random Access Memory (RAM) and may further include a non-volatile Memory (non-volatile Memory), such as at least one disk Memory. The communication connection between the network element of the system and at least one other network element is realized through at least one communication interface 603 (which may be wired or wireless), and the internet, a wide area network, a local network, a metropolitan area network, and the like can be used.
The electronic device provided by the embodiment of the application and the method for improving the target re-identification generalization provided by the embodiment of the application have the same beneficial effects as the method adopted, operated or realized by the electronic device.
Referring to fig. 7, the computer-readable storage medium is an optical disc 700, on which a computer program (i.e., a program product) is stored, and when the computer program is executed by a processor, the computer program performs the method for improving the target re-recognition generalization provided in any of the foregoing embodiments.
It should be noted that examples of the computer-readable storage medium may also include, but are not limited to, phase change memory (PRAM), Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), other types of Random Access Memory (RAM), Read Only Memory (ROM), Electrically Erasable Programmable Read Only Memory (EEPROM), flash memory, or other optical and magnetic storage media, which are not described in detail herein.
The computer-readable storage medium provided by the above-mentioned embodiment of the present application and the method for improving the target re-recognition generalization provided by the embodiment of the present application have the same inventive concept, and have the same beneficial effects as the method adopted, run, or implemented by the application program stored in the computer-readable storage medium.
The technical features of the above embodiments can be arbitrarily combined, and for the sake of brevity, all possible combinations of the technical features in the above embodiments are not described, but should be considered as the scope of the present specification as long as there is no contradiction between the combinations of the technical features.
The above examples only show some embodiments of the present invention, and the description thereof is more specific and detailed, but not construed as limiting the scope of the present invention. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the inventive concept, which falls within the scope of the present invention. Therefore, the protection scope of the present patent shall be subject to the appended claims.
Claims (10)
1. A method for improving the generalization of target re-recognition is characterized by comprising the following steps:
inputting sample image data of unknown fields into a pre-trained multi-field expert mixed model to obtain the exclusive characteristics of each field and the query characteristics of the sample images;
calculating the average similarity of the query feature and the exclusive feature, and determining the correlation between the sample image and each field according to the average similarity;
carrying out weighted aggregation on the exclusive characteristics of each field according to the field correlation to obtain aggregation characteristics;
and performing target re-identification on the sample image according to the aggregation characteristics.
2. The method of claim 1, wherein the multi-domain expert hybrid model comprises a backbone network module, a voting network module, and a plurality of domain expert network modules, wherein the voting network module and the plurality of domain expert network modules branch after the backbone network module.
3. The method of claim 2, wherein inputting sample image data of unknown domains into a pre-trained multi-domain expert hybrid model, obtaining the unique features of each domain and the query features of the sample image, comprises:
inputting sample image data of an unknown field into a pre-trained multi-field expert mixed model;
extracting shared characteristics among different fields through the backbone network module;
mapping the shared characteristics through each domain expert network module to obtain the exclusive characteristics of each domain;
and mapping the shared features through the voting network module to obtain the query features of the sample image.
4. The method of claim 1, wherein prior to inputting sample image data of an unknown domain into a pre-trained multi-domain expert hybrid model, further comprising:
and training the multi-field expert mixed model in a meta-learning training mode.
5. The method of claim 4, wherein training the multi-domain expert hybrid model by meta-learning training comprises:
at the beginning of each training iteration, randomly dividing K source domain training data into K-1 meta training domains and 1 meta testing domain;
calculating a domain loss function and a relation alignment loss function by using the meta-training domain data, updating parameters of a backbone network module and parameters of a plurality of domain expert network modules by using the domain loss function, and updating parameters of a voting network module according to a first-order gradient of the relation alignment loss function;
calculating a domain loss function and a relationship alignment loss function again by using the meta-test domain data;
updating parameters of the backbone network module and parameters of a plurality of domain expert network modules according to the domain loss function obtained by recalculation;
updating parameters of the voting network module according to the second-order gradient of the relation alignment loss function obtained by recalculation;
and iterating the meta-training domain and the meta-testing domain for multiple times until the network converges.
6. The method of claim 5, wherein the domain loss function is comprised of a metric loss function and a decorrelation loss function, and wherein the metric loss function is comprised of a classification loss function, a triplet loss function, and a center loss function.
7. The method according to claim 5, wherein the relationship alignment loss function is obtained by computing a binary cross-entropy loss of domain-specific feature and aggregate feature metric relationships.
8. The utility model provides a promote device of target re-identification generalization which characterized in that includes:
the characteristic extraction module is used for inputting sample image data of unknown fields into a pre-trained multi-field expert mixed model to obtain the exclusive characteristic of each field and the query characteristic of the sample image;
the correlation calculation module is used for calculating the average similarity of the query feature and the exclusive feature and determining the correlation between the sample image and each field according to the average similarity;
the characteristic aggregation module is used for carrying out weighted aggregation on the exclusive characteristic of each field according to the field correlation to obtain an aggregation characteristic;
and the target re-identification module is used for carrying out target re-identification on the sample image according to the aggregation characteristics.
9. An apparatus for improving target re-recognition generalization comprising a processor and a memory having stored thereon program instructions, the processor being configured to, when executing said program instructions, perform the method for improving target re-recognition generalization of any one of claims 1 to 7.
10. A computer readable medium having computer readable instructions stored thereon which are executable by a processor to implement a method of improving target re-identification generalization according to any one of claims 1 to 7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110627547.4A CN113505642B (en) | 2021-06-04 | 2021-06-04 | Method, device, equipment and storage medium for improving target re-identification generalization |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110627547.4A CN113505642B (en) | 2021-06-04 | 2021-06-04 | Method, device, equipment and storage medium for improving target re-identification generalization |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113505642A true CN113505642A (en) | 2021-10-15 |
| CN113505642B CN113505642B (en) | 2023-10-24 |
Family
ID=78009063
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110627547.4A Active CN113505642B (en) | 2021-06-04 | 2021-06-04 | Method, device, equipment and storage medium for improving target re-identification generalization |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113505642B (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111209832A (en) * | 2019-12-31 | 2020-05-29 | 华瑞新智科技（北京）有限公司 | Auxiliary obstacle avoidance training method, equipment and medium for transformer substation inspection robot |
| CN111401281A (en) * | 2020-03-23 | 2020-07-10 | 山东师范大学 | Unsupervised pedestrian re-identification method and system based on deep clustering and sample learning |
| CN111724083A (en) * | 2020-07-21 | 2020-09-29 | 腾讯科技（深圳）有限公司 | Training method and device for financial risk recognition model, computer equipment and medium |
| CN111738172A (en) * | 2020-06-24 | 2020-10-02 | 中国科学院自动化研究所 | Cross-domain target re-identification method based on feature counterstudy and self-similarity clustering |
| CN112232241A (en) * | 2020-10-22 | 2021-01-15 | 华中科技大学 | Pedestrian re-identification method and device, electronic equipment and readable storage medium |
-
2021
- 2021-06-04 CN CN202110627547.4A patent/CN113505642B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111209832A (en) * | 2019-12-31 | 2020-05-29 | 华瑞新智科技（北京）有限公司 | Auxiliary obstacle avoidance training method, equipment and medium for transformer substation inspection robot |
| CN111401281A (en) * | 2020-03-23 | 2020-07-10 | 山东师范大学 | Unsupervised pedestrian re-identification method and system based on deep clustering and sample learning |
| CN111738172A (en) * | 2020-06-24 | 2020-10-02 | 中国科学院自动化研究所 | Cross-domain target re-identification method based on feature counterstudy and self-similarity clustering |
| CN111724083A (en) * | 2020-07-21 | 2020-09-29 | 腾讯科技（深圳）有限公司 | Training method and device for financial risk recognition model, computer equipment and medium |
| CN112232241A (en) * | 2020-10-22 | 2021-01-15 | 华中科技大学 | Pedestrian re-identification method and device, electronic equipment and readable storage medium |
Non-Patent Citations (2)
| Title |
|---|
| SHIJIE YU 等: "COCAS: A Large-Scale Clothes Changing Person Dataset for Re-identification", IEEE * |
| YUNPENG ZHAI 等: "Multiple Expert Brainstorming for Domain Adaptive Person Re-identi cation", ARXIV * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113505642B (en) | 2023-10-24 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110033026B (en) | Target detection method, device and equipment for continuous small sample images | |
| CN110807472B (en) | Image recognition method and device, electronic equipment and storage medium | |
| CN111932577B (en) | Text detection method, electronic device and computer readable medium | |
| CN113065525B (en) | Age identification model training method, face age identification method and related device | |
| CN111310743B (en) | Face recognition method and device, electronic equipment and readable storage medium | |
| CN111179270A (en) | Image co-segmentation method and device based on attention mechanism | |
| TWI803243B (en) | Method for expanding images, computer device and storage medium | |
| CN114596546A (en) | Vehicle weight recognition method and device, computer and readable storage medium | |
| CN115424053A (en) | Small sample image identification method, device and equipment and storage medium | |
| CN109447943B (en) | Target detection method, system and terminal equipment | |
| Wang et al. | Contrastive GNN-based traffic anomaly analysis against imbalanced dataset in IoT-based its | |
| CN116612500B (en) | Pedestrian re-recognition model training method and device | |
| CN113505642B (en) | Method, device, equipment and storage medium for improving target re-identification generalization | |
| CN112069269A (en) | Big data and multidimensional feature-based data tracing method and big data cloud server | |
| CN116821087A (en) | Power transmission line fault database construction method, device, terminal and storage medium | |
| CN115984671A (en) | Model online updating method and device, electronic equipment and readable storage medium | |
| CN108400887B (en) | Method and device for identifying structure of symbol network | |
| CN115019296A (en) | Cascading-based license plate detection and identification method and device | |
| Cheng et al. | Image-to-gps verification through a bottom-up pattern matching network | |
| CN112528068A (en) | Voiceprint feature storage method, voiceprint feature matching method and device and electronic equipment | |
| CN111461255A (en) | Siamese network image identification method and system based on interval distribution | |
| CN111340092B (en) | Target association processing method and device | |
| CN117573655B (en) | Data management optimization method and system based on convolutional neural network | |
| US20230298326A1 (en) | Image augmentation method, electronic device and readable storage medium | |
| CN117576518B (en) | Image distillation method, apparatus, electronic device, and computer-readable storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |