CN113989627B - City prevention and control image detection method and system based on asynchronous federal learning - Google Patents

City prevention and control image detection method and system based on asynchronous federal learning Download PDF

Info

Publication number
CN113989627B
CN113989627B CN202111633287.8A CN202111633287A CN113989627B CN 113989627 B CN113989627 B CN 113989627B CN 202111633287 A CN202111633287 A CN 202111633287A CN 113989627 B CN113989627 B CN 113989627B
Authority
CN
China
Prior art keywords
training
model
local
data
image data
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.)
Active
Application number
CN202111633287.8A
Other languages
Chinese (zh)
Other versions
CN113989627A (en
Inventor
袁戟
常可欣
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shenzhen Wanwuyun Technology Co ltd
Original Assignee
Shenzhen Wanwuyun Technology Co ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Shenzhen Wanwuyun Technology Co ltd filed Critical Shenzhen Wanwuyun Technology Co ltd
Priority to CN202111633287.8A priority Critical patent/CN113989627B/en
Publication of CN113989627A publication Critical patent/CN113989627A/en
Application granted granted Critical
Publication of CN113989627B publication Critical patent/CN113989627B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F21/00—Security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
    • G06F21/60—Protecting data
    • G06F21/602—Providing cryptographic facilities or services
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L67/00—Network arrangements or protocols for supporting network services or applications
    • H04L67/01—Protocols
    • H04L67/06—Protocols specially adapted for file transfer, e.g. file transfer protocol [FTP]
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L9/00—Cryptographic mechanisms or cryptographic arrangements for secret or secure communications; Network security protocols
    • H04L9/008—Cryptographic mechanisms or cryptographic arrangements for secret or secure communications; Network security protocols involving homomorphic encryption

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Software Systems (AREA)
  • Signal Processing (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Computer Security & Cryptography (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Bioethics (AREA)
  • Computer Hardware Design (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses an urban prevention and control image detection method based on asynchronous federal learning, which comprises the following steps that a cloud server initializes a global model; each end device carries out scene labeling on the urban prevention and control image data, divides the urban prevention and control image data into a training set and a testing set, and screens out a data set which is distributed close to a sample of the testing set from the training set as a to-be-trained set; acquiring a global model from a cloud server, and performing local training on a to-be-trained set to obtain a local model; each end device uploads the local model to a cloud server after homomorphic encryption; and the cloud server performs global model training on the local model by adopting an asynchronous federated calculation strategy to obtain an updated global model.

Description

Urban prevention and control image detection method and system based on asynchronous federal learning
Technical Field
The invention relates to the technical field of image recognition, in particular to an urban prevention and control image detection method and system based on asynchronous federal learning.
Background
At present, the image recognition scenes of urban prevention and control mainly include application scenes such as road surface construction damage, vehicle parking violation, river surface floating objects, garbage dumping and the like. Since the above application scenarios are widely distributed in various parts of a city, collaboration between different companies is required. For example, company a uses a camera to acquire image data, company B uses an unmanned aerial vehicle mounted camera to acquire images, and since pictures acquired by different companies cannot be directly shared due to data security, privacy protection and the like, a federal training network needs to be established to extract data features locally in each company, data is pushed to a cloud for global training by using a homomorphic encryption technology, and trained network parameters are respectively transmitted back to company a and company B for subsequent model application after decryption.
Due to the fact that the frequency of image updating and model training of different companies is different, global synchronous federal calculation needs to consume more manpower and material resources to carry out unified management on side equipment, and efficiency is low; and the image angle that acquires through unmanned aerial vehicle and fixed camera etc. can show the difference, lead to not being accurate in the training.
Disclosure of Invention
The invention aims to provide an urban prevention and control image detection method and system based on asynchronous federal learning, and aims to solve the problems that in the prior art, a global synchronous federal calculation strategy is adopted, more manpower and material resources are consumed to carry out unified management on edge-end equipment, the efficiency is low, and training is not accurate due to differences of image angles and the like.
In a first aspect, an embodiment of the present invention provides an urban prevention and control image detection method based on asynchronous federal learning, including:
s101, a cloud server initializes a global model;
s102, each end device carries out scene labeling on city prevention and control image data, divides the city prevention and control image data into a training set and a testing set, and screens out a data set which is distributed close to a sample of the testing set from the training set as a to-be-trained set;
s103, acquiring the global model from a cloud server, and performing local training on the to-be-trained set to obtain a local model;
s104, each end device uploads the local model to a cloud server after homomorphic encryption;
s105, the cloud server performs global model training on the local model by adopting an asynchronous federated computing strategy to obtain an updated global model;
and S106, turning to the step S102 until the global model on the cloud server reaches the expected performance.
In a second aspect, an embodiment of the present invention provides an urban prevention and control image detection system based on asynchronous federal learning, which is characterized by including: the system comprises a cloud server and end devices;
the cloud server is used for initializing a global model; carrying out global model training on the local model by adopting an asynchronous federal calculation strategy to obtain an updated global model;
the end equipment is used for carrying out scene labeling on the urban prevention and control image data, dividing the urban prevention and control image data into a training set and a testing set, and screening out a data set which is distributed close to a sample of the testing set from the training set to serve as a to-be-trained set; acquiring the global model from a cloud server, and carrying out local training on the set to be trained to obtain a local model; and uploading the local model to a cloud server after homomorphic encryption.
According to the embodiment of the invention, the safety is well guaranteed by limiting the urban prevention and control image data in each end device for marking and training; the cloud server performs weight calculation on the local models trained by the end devices through an asynchronous federal calculation strategy, selects the local model with larger weight to perform global training, so that the local models participating in training are more closely associated and have higher relevancy, and compared with a synchronous federal calculation strategy, the cloud server does not need to perform global training after all the end devices train the local models of the current round, so that the efficiency is higher, and the end devices which are not used can perform other work and cannot be always occupied;
meanwhile, the training sets in the urban prevention and control image data are screened through the equipment at each end, a to-be-trained set which is close to the data distribution of the test set is obtained, and the training accuracy of the model is improved.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings needed to be used in the description of the embodiments are briefly introduced below, and it is obvious that the drawings in the following description are some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on these drawings without creative efforts.
Fig. 1 is a schematic flow chart of a city prevention and control image detection method based on asynchronous federal learning according to an embodiment of the present invention;
fig. 2 is a system block diagram of an urban prevention and control image detection system based on asynchronous federal learning according to an embodiment of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some, not all, embodiments of the present invention. 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.
It will be understood that the terms "comprises" and/or "comprising," when used in this specification and the appended claims, specify the presence of stated features, integers, steps, operations, elements, and/or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and/or groups thereof.
It is also to be understood that the terminology used in the description of the invention herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. As used in the specification of the present invention and the appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise.
It should be further understood that the term "and/or" as used in this specification and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.
Referring to fig. 1, a city prevention and control image detection method based on asynchronous federal learning includes:
s101, a cloud server initializes a global model;
s102, each end device carries out scene labeling on city prevention and control image data, divides the city prevention and control image data into a training set and a testing set, and screens out a data set which is distributed close to a sample of the testing set from the training set as a to-be-trained set;
s103, acquiring the global model from a cloud server, and performing local training on the to-be-trained set to obtain a local model;
s104, each end device uploads the local model to a cloud server after homomorphic encryption;
s105, the cloud server performs global model training on the local model by adopting an asynchronous federated computing strategy to obtain an updated global model;
and S106, turning to the step S102 until the global model on the cloud server reaches the expected performance.
In the embodiment, the safety is well guaranteed by limiting the urban prevention and control image data in each end device for marking and training; and through asynchronous federal calculation strategy, the cloud server carries out weight calculation on the local model trained by each end device, and selects the local model with larger weight to carry out global training, so that the connection between the local models participating in training is tighter, the correlation degree is higher, and compared with the synchronous federal calculation strategy, the cloud server can carry out global training without waiting for all the end devices to train out the local models of the round, the efficiency is higher, the end devices which are not in use can carry out other works, and the cloud server cannot be occupied all the time.
Meanwhile, the training sets in the urban prevention and control image data are screened through the equipment at each end, a to-be-trained set which is close to the data distribution of the test set is obtained, and the training accuracy of the model is improved.
The city prevention and control image types comprise road surface construction damage, vehicle illegal parking, river surface floating objects, garbage dumping and the like.
The end equipment can be processors distributed in different companies or processors distributed in the same company; can be distributed in different corners of the same city, or can be centralized in a machine room.
Because each end device can obtain a large number of images through each image obtaining terminal, each end device needs to label the images after obtaining the images, and executes the step S102, and executes the steps S102-S105 in a circulating manner; the time interval from S105 to the beginning of the step S102 may be a preset time, or a certain preset condition is reached, so that the whole step operation may not be performed when any image is obtained, which wastes the running resources of the end device, and it is not necessary, and only after enough city prevention and control image data is obtained, the labeling and local training are performed, which may be set according to the actual situation.
In an embodiment, the screening out, from the training set, a data set that is distributed close to the test set samples as a to-be-trained set includes:
calculating a data distribution of the test set
Figure 100002_DEST_PATH_IMAGE002
；
Calculating the feature distribution KL divergence of the training set and the test set by adopting the following formula:
Figure 100002_DEST_PATH_IMAGE004
；
judging whether the sample distribution of the training set is close to that of the test set according to the KL divergence;
if yes, sampling is not needed to be carried out on the training set;
and if not, sampling the training set to obtain the set to be trained.
Wherein the content of the first and second substances,
Figure 100002_DEST_PATH_IMAGE006
representing the corresponding normalized features of the partially sampled data in the kth end device,
Figure 100002_DEST_PATH_IMAGE008
a data distribution representing the training set is determined,
Figure 100002_DEST_PATH_IMAGE010
representing the normalized data characteristics collected by an end device.
In this embodiment, the KL divergence is used to measure the distance between the distributions;
wherein the distribution data is distributed
Figure 965187DEST_PATH_IMAGE002
The model is obtained by calculation through a test set and can be determined through a probability density equation pdf (probability density function) in the calculation as a module for displaying kdeplot in matplotlib.
Judging whether the training set is close to the sample distribution of the test set according to the KL divergence, wherein a threshold value is mainly set, if the KL divergence is smaller than the threshold value, the two samples are judged to be close enough, otherwise, the samples are judged to be required to be sampled, and the samples close enough are obtained and used as a set to be trained; the threshold value needs to be adjusted according to the actual distribution form and the training result on the test set, and is not a specific value.
Specifically, samples in the training set are subjected to screening and sampling through a metroplis algorithm.
In an embodiment, the sampling the training set to obtain the set to be trained includes:
data distribution from the training set
Figure 100002_DEST_PATH_IMAGE012
Randomly sampling image data features
Figure 100002_DEST_PATH_IMAGE014
；
According to the suggested distribution
Figure 100002_DEST_PATH_IMAGE016
Compliance
Figure 100002_DEST_PATH_IMAGE018
Sampling to obtain image data features
Figure 100002_DEST_PATH_IMAGE020
；
The first intermediate value is calculated using the following formula
Figure 100002_DEST_PATH_IMAGE022
：
Figure 100002_DEST_PATH_IMAGE024
；
From
Figure 100002_DEST_PATH_IMAGE026
Sampling in the uniform distribution to obtain a second intermediate value u;
judging whether the first intermediate value is greater than or equal to a second intermediate value;
if so, accepting the image data feature
Figure 450657DEST_PATH_IMAGE020
Into the set to be trained, i.e.
Figure 100002_DEST_PATH_IMAGE028
；
If not, receiving image data characteristics
Figure 325072DEST_PATH_IMAGE014
Into the set to be trained, i.e.
Figure 100002_DEST_PATH_IMAGE030
；
Repeating the above operations to characterize all image data
Figure 108352DEST_PATH_IMAGE014
Screening is carried out, and obedience data distribution is obtained after T rounds of iteration
Figure 650191DEST_PATH_IMAGE002
The set of image data features of;
carrying out reverse iteration on the image data feature set to obtain the to-be-trained set;
wherein the content of the first and second substances,
Figure 100002_DEST_PATH_IMAGE032
to represent
Figure 959426DEST_PATH_IMAGE020
The distribution of the data of (a) is,
Figure 100002_DEST_PATH_IMAGE034
to represent
Figure 773667DEST_PATH_IMAGE014
The data distribution of (2).
In this example, after obtaining the sample
Figure 100002_DEST_PATH_IMAGE036
Due to the fact that
Figure 891795DEST_PATH_IMAGE036
To generate samples, a reverse iteration is required to obtain samples
Figure 100002_DEST_PATH_IMAGE038
Neutralization
Figure 652857DEST_PATH_IMAGE036
Best matched sample
Figure 100002_DEST_PATH_IMAGE040
；
Calculated using the following formula
Figure 463818DEST_PATH_IMAGE040
：
Figure 250509DEST_PATH_IMAGE040
Figure 100002_DEST_PATH_IMAGE042
；
Wherein the content of the first and second substances,
Figure 100002_DEST_PATH_IMAGE044
to represent
Figure 703487DEST_PATH_IMAGE044
And (4) norm.
In an embodiment, the obtaining the global model from the cloud server, and performing local training on the to-be-trained set to obtain a local model includes:
and each end device adopts a spark R-CNN model to carry out local training on the marked urban prevention and control image data.
In this embodiment, the single-stage model of YOLO is adopted in the prior art, and compared with the single-stage model of YOLO, the Sparse R-CNN model has higher precision.
In an embodiment, the training the local model by using an asynchronous federated calculation strategy to obtain an updated global model includes:
and the cloud server trains the local model by adopting a spark R-CNN model to obtain an updated global model.
In this embodiment, the single-stage model of YOLO is adopted in the prior art, and compared with the single-stage model of YOLO, the Sparse R-CNN model has higher precision.
In an embodiment, the training of the local model by using an asynchronous federated calculation strategy to obtain an updated global model (step S104) includes:
s201, the cloud server counts the number of local models participating in the global model training, and judges whether each local model participates in past training in a specified time window;
s202, if the local model is judged to be negative, the local model is not added with the global model training;
and S203, if the weight is judged to be positive, giving corresponding weight to the local model, and adding global model training.
In this embodiment, each local model is determined to see whether it has participated in past training in a specified time window (within the specified time window, the relationship between the local models is closer, the correlation is higher, and the accuracy of the trained global model is higher), that is, the local model trained within a specific number of rounds of global training is different from the local model trained in the current round of global training.
And if so, giving corresponding weight to the local model, wherein the weight is larger as the number of rounds different from the global training of the current round is closer.
Wherein the specified window may be the number of rounds of global training.
In one embodiment, the determining whether each of the local models has engaged in past training in a specified time window comprises:
determining whether each local model has engaged in past training in a specified time window using the following formula:
Figure 100002_DEST_PATH_IMAGE046
；
wherein s is the number of rounds of the current global model update,
Figure 100002_DEST_PATH_IMAGE048
for the number of rounds a local model has taken part in training,
Figure 100002_DEST_PATH_IMAGE050
and the local model time window is a hyper-parameter and is used for setting the local model time window of how many versions can be accommodated in the global model training.
In this embodiment, generally, the number of rounds in which the local model participates in training is equal to the number of rounds updated by the global model in the previous round.
In an embodiment, if yes, assigning a corresponding weight to the local model, and adding global model training, including:
s301, corresponding weight is given to the local model by adopting the following formula:
Figure 100002_DEST_PATH_IMAGE052
；
s302, calculating the influence weight of each local model in the S-th round on the global model by adopting the following formula:
Figure 100002_DEST_PATH_IMAGE054
；
wherein the content of the first and second substances,
Figure 100002_DEST_PATH_IMAGE056
for the amount of data in a local model that the same round participates in updating,
Figure 100002_DEST_PATH_IMAGE058
the total amount of data for all local models participating in the update for the same round,
Figure 100002_DEST_PATH_IMAGE060
is the weight value of the local model for the s-th round,
Figure 100002_DEST_PATH_IMAGE062
the weight values of the global model for round s,
Figure DEST_PATH_IMAGE064
for each set of local model codes for the s-th round,
Figure DEST_PATH_IMAGE066
in order to obtain a learning rate,
Figure DEST_PATH_IMAGE068
to determine the hyper-parameters that influence the decay rate of the weights.
In this embodiment, if
Figure 764459DEST_PATH_IMAGE046
The local model can participate in the global model training, and the training participation degree given to the local model by calculating a proper weight value through a formula is achieved to achieve higher precision.
In an embodiment, if the determination is yes, the local model is given corresponding weight, a global model training is added, and then the method includes:
s401, after the updated global model is obtained through training, the cloud server sends the global model to each end device;
s402, the step S102 is carried out until the global model on the cloud server reaches the expected performance.
In this embodiment, the updated global model is sent back to each end device again, the local model on the end device is updated, the local model uploaded to the cloud server every time is guaranteed to be heavy, and the prediction accuracy is improved.
Referring to fig. 2, an urban prevention and control image detection system based on asynchronous federal learning includes a cloud server and end devices;
the cloud server is used for initializing a global model; carrying out global model training on the local model by adopting an asynchronous federal calculation strategy to obtain an updated global model;
the end equipment is used for carrying out scene labeling on the urban prevention and control image data, dividing the urban prevention and control image data into a training set and a testing set, and screening out a data set which is distributed close to a sample of the testing set from the training set to serve as a to-be-trained set; acquiring the global model from a cloud server, and carrying out local training on the set to be trained to obtain a local model; and uploading the local model to a cloud server after homomorphic encryption.
In an embodiment, the end device is further configured to locally train the labeled city prevention and control image data using a Sparse R-CNN model.
In an embodiment, the cloud server is configured to train the local model by using a Sparse R-CNN model to obtain an updated global model.
In an embodiment, the cloud server is configured to count the number of local models participating in the global model training, and determine whether each local model participates in past training in a specified time window;
if not, the local model is not added with the global model training;
if the local model is judged to be the global model, corresponding weight is given to the local model, and global model training is added.
In one embodiment, the cloud server is configured to determine whether each local model has been in past training within a specified time window using the following formula:
Figure 477813DEST_PATH_IMAGE046
；
wherein s is the number of rounds of the current global model update,
Figure 814116DEST_PATH_IMAGE048
for the number of rounds a local model has taken part in training,
Figure 274048DEST_PATH_IMAGE050
and the time window is a super parameter and is used for setting the local model time window of how many versions can be accommodated in the global model training.
In an embodiment, the cloud server is configured to assign a corresponding weight to the local model by using the following formula:
Figure 317090DEST_PATH_IMAGE052
；
and calculating the influence weight of each local model on the global model in the s-th round by adopting the following formula:
Figure 328908DEST_PATH_IMAGE054
；
wherein the content of the first and second substances,
Figure 762295DEST_PATH_IMAGE056
for the amount of data in a local model that the same round participates in updating,
Figure 822655DEST_PATH_IMAGE058
the total amount of data for all local models participating in the update for the same round,
Figure 48100DEST_PATH_IMAGE060
is the weight value of the local model for the s-th round,
Figure 575027DEST_PATH_IMAGE062
the s wheelThe weight value of the global model of (2),
Figure 26868DEST_PATH_IMAGE064
for each set of local model codes for the s-th round,
Figure 687657DEST_PATH_IMAGE066
in order to obtain the learning rate of the learning,
Figure 843307DEST_PATH_IMAGE068
to determine the hyper-parameters that influence the decay rate of the weights.
While the invention has been described with reference to specific embodiments, the invention is not limited thereto, and various equivalent modifications and substitutions can be easily made by those skilled in the art within the technical scope of the invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims (8)

1. A city prevention and control image detection method based on asynchronous federal learning is characterized by comprising the following steps:
s101, a cloud server initializes a global model;
s102, each end device carries out scene labeling on city prevention and control image data, divides the city prevention and control image data into a training set and a testing set, and screens out a data set which is distributed close to a sample of the testing set from the training set as a to-be-trained set;
s103, acquiring the global model from a cloud server, and performing local training on the to-be-trained set to obtain a local model;
s104, each end device uploads the local model to a cloud server after homomorphic encryption;
s105, the cloud server performs global model training on the local model by adopting an asynchronous federated computing strategy to obtain an updated global model;
s106, turning to the step S102 until the global model on the cloud server reaches the expected performance;
in step S102, the screening out a data set from the training set, which is distributed close to the test set sample, as a to-be-trained set includes:
calculating a data distribution of the test set
Figure DEST_PATH_IMAGE002
；
Calculating the feature distribution KL divergence of the training set and the test set by adopting the following formula:
Figure DEST_PATH_IMAGE004
；
judging whether the training set is close to the sample distribution of the test set or not according to the KL divergence;
if yes, sampling is not needed to be carried out on the training set;
if not, sampling the training set to obtain the set to be trained;
wherein the content of the first and second substances,
Figure DEST_PATH_IMAGE006
representing the normalized data characteristic corresponding to the partially sampled data in the kth end device,
Figure DEST_PATH_IMAGE008
a data distribution representing the training set is determined,
Figure DEST_PATH_IMAGE010
representing normalized data characteristics collected by an end device;
wherein, the sampling the training set to obtain the to-be-trained set includes:
data distribution from the training set
Figure DEST_PATH_IMAGE012
Randomly sampling image data features
Figure DEST_PATH_IMAGE014
；
According to the suggested distribution
Figure DEST_PATH_IMAGE016
Compliance
Figure DEST_PATH_IMAGE018
Sampling to obtain image data features
Figure DEST_PATH_IMAGE020
；
The first intermediate value is calculated using the following formula
Figure DEST_PATH_IMAGE022
：
Figure DEST_PATH_IMAGE024
；
From
Figure DEST_PATH_IMAGE026
Sampling in the uniform distribution to obtain a second intermediate value u;
judging whether the first intermediate value is greater than or equal to a second intermediate value;
if so, accepting the image data feature
Figure 436829DEST_PATH_IMAGE020
Into the set to be trained, i.e.
Figure DEST_PATH_IMAGE028
；
If not, receiving image data characteristics
Figure 21525DEST_PATH_IMAGE014
Into the set to be trained, i.e.
Figure DEST_PATH_IMAGE030
；
Repeating the above operations to characterize all image data
Figure 34481DEST_PATH_IMAGE014
Screening is carried out, and obedience data distribution is obtained after t rounds of iteration
Figure 589090DEST_PATH_IMAGE002
The set of image data features of;
carrying out reverse iteration on the image data feature set to obtain the to-be-trained set;
wherein the content of the first and second substances,
Figure DEST_PATH_IMAGE032
to represent
Figure 144836DEST_PATH_IMAGE020
The distribution of the data of (a) is,
Figure DEST_PATH_IMAGE034
to represent
Figure 163083DEST_PATH_IMAGE014
The data distribution of (2).
2. The urban defense and control image detection method based on asynchronous federal learning according to claim 1, wherein the step of obtaining the global model from a cloud server and performing local training on the set to be trained to obtain a local model comprises the steps of:
and each end device adopts a spark R-CNN model to carry out local training on the marked set to be trained.
3. The urban prevention and control image detection method based on asynchronous federal learning according to claim 1, wherein the step of training the local model by adopting an asynchronous federal calculation strategy to obtain an updated global model comprises the following steps:
the cloud server counts the number of local models participating in the global model training, and judges whether each local model participates in past training in a specified time window;
if not, the local model is not added with the global model training;
if the local model is judged to be the global model, corresponding weight is given to the local model, and global model training is added.
4. The city prevention and control image detection method based on asynchronous federal learning as claimed in claim 3, wherein said determining whether each local model has been involved in past training in a specified time window comprises:
determining whether each local model has engaged in past training in a specified time window using the following formula:
Figure DEST_PATH_IMAGE036
；
wherein s is the number of rounds of the current global model update,
Figure DEST_PATH_IMAGE038
for the number of rounds a local model has taken part in training,
Figure DEST_PATH_IMAGE040
and the time window is a super parameter and is used for setting the local model time window of how many versions can be accommodated in the global model training.
5. The method for detecting the urban prevention and control images based on the asynchronous federated learning as recited in claim 3, wherein if the result of the determination is yes, the local model is given corresponding weight, and a global model training is added, comprising:
corresponding weight is given to the local model by adopting the following formula:
Figure DEST_PATH_IMAGE042
；
and calculating the influence weight of each local model on the global model in the s-th round by adopting the following formula:
Figure DEST_PATH_IMAGE044
；
wherein the content of the first and second substances,
Figure DEST_PATH_IMAGE046
for the amount of data in a local model that the same round participates in updating,
Figure DEST_PATH_IMAGE048
the total amount of data for all local models participating in the update for the same round,
Figure DEST_PATH_IMAGE050
is the weight value of the local model for the s-th round,
Figure DEST_PATH_IMAGE052
the weight values of the global model for round s,
Figure DEST_PATH_IMAGE054
for each set of local model codes for the s-th round,
Figure DEST_PATH_IMAGE056
in order to obtain a learning rate,
Figure DEST_PATH_IMAGE058
to determine the hyperparameters that influence the decay rate of the weights.
6. An urban prevention and control image detection system based on asynchronous federal learning is characterized by comprising: the system comprises a cloud server and end devices;
the cloud server is used for initializing a global model; carrying out global model training on the local model by adopting an asynchronous federal calculation strategy to obtain an updated global model;
the end equipment is used for carrying out scene labeling on the urban prevention and control image data, dividing the urban prevention and control image data into a training set and a testing set, and screening out a data set which is distributed close to a sample of the testing set from the training set to serve as a to-be-trained set; acquiring the global model from a cloud server, and carrying out local training on the set to be trained to obtain a local model; uploading the local model to a cloud server after homomorphic encryption;
wherein the end device is further configured to:
calculating a data distribution of the test set
Figure 365132DEST_PATH_IMAGE002
；
Calculating the feature distribution KL divergence of the training set and the test set by adopting the following formula:
Figure 457853DEST_PATH_IMAGE004
；
judging whether the training set is close to the sample distribution of the test set or not according to the KL divergence;
if yes, sampling is not needed to be carried out on the training set;
if not, sampling the training set to obtain the set to be trained;
wherein the content of the first and second substances,
Figure 461581DEST_PATH_IMAGE006
representing the normalized data characteristic corresponding to the partially sampled data in the kth end device,
Figure 981555DEST_PATH_IMAGE008
a data distribution representing the training set is determined,
Figure 437944DEST_PATH_IMAGE010
normalization to represent an end device acquisitionTransforming the data characteristics;
wherein the end device is further configured to:
data distribution from the training set
Figure 334356DEST_PATH_IMAGE012
Randomly sampling image data features
Figure 67957DEST_PATH_IMAGE014
；
According to the suggested distribution
Figure DEST_PATH_IMAGE059
Compliance
Figure DEST_PATH_IMAGE060
Sampling to obtain image data features
Figure 227674DEST_PATH_IMAGE020
；
The first intermediate value is calculated using the following formula
Figure 436939DEST_PATH_IMAGE022
：
Figure DEST_PATH_IMAGE061
；
From
Figure 298495DEST_PATH_IMAGE026
Sampling in the uniform distribution to obtain a second intermediate value u;
judging whether the first intermediate value is greater than or equal to a second intermediate value;
if so, accepting the image data feature
Figure 11236DEST_PATH_IMAGE020
Into the set to be trained, i.e.
Figure 138592DEST_PATH_IMAGE028
；
If not, receiving image data characteristics
Figure 710518DEST_PATH_IMAGE014
Into the set to be trained, i.e.
Figure 807787DEST_PATH_IMAGE030
；
Repeating the above operations to characterize all image data
Figure 515980DEST_PATH_IMAGE014
Screening is carried out, and obedience data distribution is obtained after t rounds of iteration
Figure 673292DEST_PATH_IMAGE002
The set of image data features of;
carrying out reverse iteration on the image data feature set to obtain the to-be-trained set;
wherein the content of the first and second substances,
Figure DEST_PATH_IMAGE062
represent
Figure 732515DEST_PATH_IMAGE020
The distribution of the data of (a) is,
Figure DEST_PATH_IMAGE063
represent
Figure 40000DEST_PATH_IMAGE014
The data distribution of (2).
7. The urban defense and control image detection system based on asynchronous federal learning of claim 6, wherein: the end device is further used for locally training the marked urban prevention and control image data by adopting a Sparse R-CNN model.
8. The urban defense and control image detection system based on asynchronous federal learning of claim 6, wherein:
the cloud server is used for counting the number of local models participating in the global model training and judging whether each local model participates in past training in a specified time window;
if not, the local model is not added with the global model training;
if the local model is judged to be the global model, corresponding weight is given to the local model, and global model training is added.
CN202111633287.8A 2021-12-29 2021-12-29 City prevention and control image detection method and system based on asynchronous federal learning Active CN113989627B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111633287.8A CN113989627B (en) 2021-12-29 2021-12-29 City prevention and control image detection method and system based on asynchronous federal learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111633287.8A CN113989627B (en) 2021-12-29 2021-12-29 City prevention and control image detection method and system based on asynchronous federal learning

Publications (2)

Publication Number Publication Date
CN113989627A CN113989627A (en) 2022-01-28
CN113989627B true CN113989627B (en) 2022-05-27

Family

ID=79734845

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111633287.8A Active CN113989627B (en) 2021-12-29 2021-12-29 City prevention and control image detection method and system based on asynchronous federal learning

Country Status (1)

Country Link
CN (1) CN113989627B (en)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114694015B (en) * 2022-06-02 2022-08-30 深圳市万物云科技有限公司 General framework-based multi-task federal learning scene recognition method and related components
CN115082903B (en) * 2022-08-24 2022-11-11 深圳市万物云科技有限公司 Non-motor vehicle illegal parking identification method and device, computer equipment and storage medium
EP4391472A1 (en) * 2022-12-22 2024-06-26 Ntt Docomo, Inc. Method for training a machine learning model

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111368886A (en) * 2020-02-25 2020-07-03 华南理工大学 Sample screening-based label-free vehicle picture classification method
CN112232528A (en) * 2020-12-15 2021-01-15 之江实验室 Method and device for training federated learning model and federated learning system

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11461593B2 (en) * 2019-11-26 2022-10-04 International Business Machines Corporation Federated learning of clients
CN111062215B (en) * 2019-12-10 2024-02-13 金蝶软件（中国）有限公司 Named entity recognition method and device based on semi-supervised learning training
CN111708640A (en) * 2020-06-23 2020-09-25 苏州联电能源发展有限公司 Edge calculation-oriented federal learning method and system
CN112532451B (en) * 2020-11-30 2022-04-26 安徽工业大学 Layered federal learning method and device based on asynchronous communication, terminal equipment and storage medium
CN112686370A (en) * 2020-12-25 2021-04-20 深圳前海微众银行股份有限公司 Network structure search method, device, equipment, storage medium and program product

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111368886A (en) * 2020-02-25 2020-07-03 华南理工大学 Sample screening-based label-free vehicle picture classification method
CN112232528A (en) * 2020-12-15 2021-01-15 之江实验室 Method and device for training federated learning model and federated learning system

Also Published As

Publication number Publication date
CN113989627A (en) 2022-01-28

Similar Documents

Publication Publication Date Title
CN113989627B (en) City prevention and control image detection method and system based on asynchronous federal learning
CN109753928B (en) Method and device for identifying illegal buildings
CN110610197B (en) Method and device for mining difficult sample and training model and electronic equipment
CN111210399B (en) Imaging quality evaluation method, device and equipment
CN113537172B (en) Crowd density determination method, device, equipment and storage medium
CN109325429A (en) A kind of method, apparatus, storage medium and the terminal of linked character data
CN110969215A (en) Clustering method and device, storage medium and electronic device
CN109740479A (en) A kind of vehicle recognition methods, device, equipment and readable storage medium storing program for executing again
CN112711757B (en) Data security centralized management and control method and system based on big data platform
CN110852164A (en) YOLOv 3-based method and system for automatically detecting illegal building
CN113255590A (en) Defect detection model training method, defect detection method, device and system
CN114360030A (en) Face recognition method based on convolutional neural network
CN111680733A (en) Component detection method, sample library establishment method, device, platform, system and medium
CN110186923B (en) River pollution monitoring method and device and storage medium
Gueguen et al. Mapping human settlements and population at country scale from VHR images
Khudov et al. The Choice of Quality Indicator for the Image Segmentation Evaluation
Bush et al. Deep Neural Networks for visual bridge inspections and defect visualisation in Civil Engineering
Khosravi et al. Vehicle speed and dimensions estimation using on-road cameras by identifying popular vehicles
CN116597193A (en) Tunnel intelligent construction-oriented tunnel face unmanned geological logging method and system
CN111680175B (en) Face database construction method, computer equipment and computer readable storage medium
CN111222370A (en) Case studying and judging method, system and device
CN112990350B (en) Target detection network training method and target detection network-based coal and gangue identification method
CN114218383A (en) Method, device and application for judging repeated events
CN114488337A (en) High-altitude parabolic detection method and device
Thakur et al. Evidence of long range dependence and self-similarity in urban traffic systems

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