WO2021174877A1 - 基于智能决策的目标检测模型的处理方法、及其相关设备 - Google Patents

基于智能决策的目标检测模型的处理方法、及其相关设备 Download PDF

Info

Publication number
WO2021174877A1
WO2021174877A1 PCT/CN2020/124439 CN2020124439W WO2021174877A1 WO 2021174877 A1 WO2021174877 A1 WO 2021174877A1 CN 2020124439 W CN2020124439 W CN 2020124439W WO 2021174877 A1 WO2021174877 A1 WO 2021174877A1
Authority
WO
WIPO (PCT)
Prior art keywords
target detection
detection model
model
model parameters
parameters
Prior art date
Application number
PCT/CN2020/124439
Other languages
English (en)
French (fr)
Inventor
王健宗
肖京
何安珣
Original Assignee
平安科技（深圳）有限公司
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 平安科技（深圳）有限公司 filed Critical 平安科技（深圳）有限公司
Publication of WO2021174877A1 publication Critical patent/WO2021174877A1/zh

Links

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
    • 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
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/07—Target detection

Definitions

  • This application relates to the field of artificial intelligence technology, and in particular to a method for processing a target detection model based on intelligent decision-making, and related equipment.
  • Target detection involves the detection model in intelligent decision-making.
  • the image can be input to the target detection model, and the target detection model can process the image and output the target object in the image.
  • the target detection model can process the image and output the target object in the image.
  • you can first input garbage pictures into the target detection model to identify the garbage, so as to guide people to recycle and classify the identified garbage.
  • the target detection model usually can only focus on a single target for detection, and when training the target detection model, because the amount of data in the local data set is usually limited, the target detection model is difficult to detect. With sufficient training, the detection accuracy of the target detection model is low.
  • the purpose of the embodiments of the present application is to propose a method, device, computer equipment, and storage medium for processing a target detection model based on intelligent decision-making, so as to solve the problem of low detection accuracy of the target detection model.
  • an embodiment of the present application provides a method for processing a target detection model based on intelligent decision-making, wherein the target detection model is a multi-target detection model, and the following technical solutions are adopted:
  • the updated relay multi-target detection model is used as the initial multi-target detection model for the next round of training, and iterative training is performed until the model converges, and the multi-target detection model is obtained.
  • an embodiment of the present application also provides a processing device for a target detection model based on intelligent decision-making, wherein the target detection model is a multi-target detection model, and the following technical solutions are adopted:
  • the acquisition module is used to acquire the local data set and the initial multi-target detection model
  • a model training module configured to train the initial multi-target detection model according to the local data set to obtain a relay multi-target detection model
  • a parameter calculation module for generating additional random numbers, and calculating composite model parameters according to the additional random numbers and the model parameters of the relay multi-target detection model
  • the parameter sending module is configured to send the composite model parameters to a central server to instruct the central server to generate global model parameters according to the composite model parameters of each node;
  • a model update module configured to receive the global model parameters from the central server to update the relay multi-target detection model
  • the iterative training module is used to use the updated relay multi-target detection model as the initial multi-target detection model for the next round of training for iterative training until the model converges to obtain the multi-target detection model.
  • an embodiment of the present application further provides a computer device, including a memory and a processor, the memory stores computer-readable instructions, and the processor implements the following steps when executing the computer-readable instructions:
  • the updated relay multi-target detection model is used as the initial multi-target detection model for the next round of training, and iterative training is performed until the model converges, and the multi-target detection model is obtained.
  • embodiments of the present application also provide a computer-readable storage medium, the computer-readable storage medium stores computer-readable instructions, and the computer-readable instructions implement the following steps when executed by a processor:
  • the updated relay multi-target detection model is used as the initial multi-target detection model for the next round of training, and iterative training is performed until the model converges, and the multi-target detection model is obtained.
  • the embodiments of the present application mainly have the following beneficial effects: firstly, the initial multi-target detection model is trained according to the local data set to obtain the relay multi-target detection model; additional random numbers are generated, and the additional random numbers are used for centering.
  • the model parameters of the multi-target detection model are calculated to encrypt the model parameters; the encrypted model parameters are sent to the central server in the alliance network, and the central server generates the global model parameters according to the composite model parameters of each node, the global model
  • the relay multi-target detection model is updated and iterative training is performed; the global model parameters are obtained on the basis of the local data set of each terminal, and the characteristics of multiple local data sets are integrated to protect the local
  • a larger-scale data set is used to train the multi-target detection model, which improves the detection accuracy of the multi-target detection model obtained after the training is completed.
  • Figure 1 is an exemplary system architecture diagram to which the present application can be applied;
  • FIG. 2 is a flowchart of an embodiment of a method for processing a target detection model based on intelligent decision-making according to the present application
  • FIG. 3 is a schematic structural diagram of an embodiment of an apparatus for processing a target detection model based on intelligent decision-making according to the present application
  • Fig. 4 is a schematic structural diagram of an embodiment of a computer device according to the present application.
  • the system architecture 100 may include terminals 101, 102, 103, a network 104, and a server 105.
  • the network 104 is used to provide a medium for communication links between the terminals 101, 102, 103 and the server 105.
  • the network 104 may include various connection types, such as wired, wireless communication links, or fiber optic cables, and so on.
  • the user can use the terminals 101, 102, and 103 to interact with the server 105 through the network 104 to receive or send messages and so on.
  • Various communication client applications such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social platform software, etc., can be installed on the terminals 101, 102, and 103.
  • the terminals 101, 102, and 103 may be various electronic devices with a display screen and supporting web browsing, including but not limited to smart phones, tablet computers, e-book readers, MP3 players (Moving Picture Experts Group Audio Layer III, moving images) Experts compress standard audio layer 3), MP4 (Moving Picture Experts Group Audio Layer IV, dynamic image experts compress standard audio layer 4) players, laptop portable computers and desktop computers, etc.
  • MP3 players Motion Picture Experts Group Audio Layer III, moving images
  • MP4 Moving Picture Experts Group Audio Layer IV, dynamic image experts compress standard audio layer 4
  • laptop portable computers and desktop computers etc.
  • the server 105 may be a server that provides various services, for example, a background server that provides support for pages displayed on the terminals 101, 102, and 103.
  • the processing method of the target detection model based on intelligent decision provided by the embodiments of the present application is generally executed by the terminal. Accordingly, the processing device of the target detection model based on intelligent decision is generally set in the terminal.
  • terminals, networks, and servers in FIG. 1 are merely illustrative. There can be any number of terminals, networks, and servers according to implementation needs.
  • the target detection model is a multi-target detection model
  • the processing method of the target detection model based on intelligent decision-making includes the following steps:
  • Step S201 Obtain a local data set and an initial multi-target detection model.
  • the electronic device (such as the terminal shown in FIG. 1) on which the method for processing the target detection model based on intelligent decision-making runs can communicate with other terminals or servers through a wired connection or a wireless connection.
  • the above-mentioned wireless connection methods can include, but are not limited to, 3G/4G connection, WiFi connection, Bluetooth connection, WiMAX connection, Zigbee connection, UWB (ultra wideband) connection, and other wireless connection methods currently known or developed in the future .
  • the local data set may be a data set stored in the terminal;
  • the initial multi-target detection model may be the initial multi-target detection model.
  • an application program for target detection may be installed in the terminal.
  • the terminal loads the local initial multi-target detection model and obtains the local data set stored in the terminal.
  • This application uses a multi-target detection model, and the multi-target detection model can detect multiple targets at one time, thereby improving the efficiency of target detection.
  • the multi-target detection model may be a RetinaNet network.
  • the multi-target detection model in this application is a lightweight model. In addition to being deployed in the server, it can also be deployed in various terminals.
  • the terminal holder can expand the local data set according to his own needs, for example, by taking photos or downloading from The Internet obtains the image expansion data set, which reduces the difficulty of expanding the local data set and enriches the data volume of the local data set.
  • the processing method of the target detection model based on intelligent decision can also be executed by the server, and the server loads the initial multi-target detection model and obtains the local data set. After the server training is completed, the target detection interface is provided, and the user can call the target detection interface on the terminal to perform target detection.
  • the above-mentioned local data set may also be stored in a node of a blockchain.
  • the blockchain referred to in this application is a new application mode of computer technology such as distributed data storage, point-to-point transmission, consensus mechanism, and encryption algorithm.
  • Blockchain essentially a decentralized database, is a series of data blocks associated with cryptographic methods. Each data block contains a batch of network transaction information for verification. The validity of the information (anti-counterfeiting) and the generation of the next block.
  • the blockchain can include the underlying platform of the blockchain, the platform product service layer, and the application service layer.
  • step S202 the initial multi-target detection model is trained according to the local data set to obtain the relay multi-target detection model.
  • the terminal trains the initial multi-target detection model according to the local data set, and stops training when the training stop condition is satisfied, to obtain the relay multi-target detection model.
  • the training stop condition may be that the number of iterations in the training reaches a preset value, or the prediction error obtained in the training is less than a preset error threshold.
  • the relay multi-target detection model obtained by the terminal may be convergent or non-convergent.
  • step S203 an additional random number is generated, and the composite model parameter is calculated according to the additional random number and the model parameters of the relay multi-target detection model.
  • the additional random number may be a parameter that implements an encryption function for the model parameters of the relay multi-target detection model.
  • the terminal performs federated learning
  • the terminal may be a node in the alliance network.
  • Each node in the alliance network generates additional random numbers, and the value of the additional random numbers obtained is zero after the addition.
  • the terminal extracts the model parameters of the relay multi-target detection model, calculates the additional random numbers and model parameters, and obtains the composite model parameters to protect the data privacy of the model parameters in the local relay multi-target detection model.
  • step S204 the composite model parameters are sent to the central server to instruct the central server to generate global model parameters according to the composite model parameters of each node.
  • the central server may be a server that plays a central control role in federated learning, and is used to instruct each node to perform federated learning.
  • the terminal sends the composite model parameters to the central server.
  • the central server can perform linear calculations on the composite model parameters of each model.
  • the central server averages the composite model parameters of each model to obtain the global model parameters.
  • Step S205 Receive global model parameters from the central server to update the relay multi-target detection model.
  • the central server sends the global model parameters to each node.
  • the terminal After the terminal receives the global model parameters, it updates the local relay multi-target detection model according to the global model parameters, specifically replacing the model parameters in the relay multi-target detection model with the global model parameters.
  • step S206 the updated relay multi-target detection model is used as the initial multi-target detection model for the next round of training, and iterative training is performed until the model converges to obtain a multi-target detection model.
  • the terminal uses the relay multi-target detection model as the initial multi-target detection model, and continues to train the obtained initial multi-target detection model according to the local data set, that is, iterative step S202 Go to step S206, until the model converges, the terminal stops training, and a multi-target detection model is obtained.
  • the condition for model convergence may be that the prediction error obtained in training is less than a preset error threshold.
  • the gradient information is transferred between the terminal and the central server.
  • the terminal performs calculations on the additional random number and the model gradient of the relay multi-target detection model to obtain the conforming model gradient, and sends the conforming model gradient to the central server.
  • the central server accumulates the gradients that conform to the model, then calculates the average value, and sends the average value as the global average gradient to each node to update the relay multi-target detection model.
  • each node when each node in the alliance network achieves model convergence, each node stops training, and the relay multi-target detection model when training is stopped is used as the multi-target detection model of each node.
  • the initial multi-target detection model is trained according to the local data set to obtain the relay multi-target detection model; additional random numbers are generated, and the additional random numbers are used to perform calculations on the model parameters of the relay multi-target detection model.
  • Model parameters are encrypted; the encrypted model parameters are sent to the central server in the alliance network.
  • the central server generates global model parameters based on the composite model parameters of each node.
  • the global model parameters are sent to the terminal to update the relay data.
  • Target detection model and iterative training global model parameters are obtained on the basis of the local data set of each terminal, and the characteristics of multiple local data sets are integrated. On the basis of protecting the local data set, a larger scale is achieved.
  • the multi-target detection model is trained on the data set of, which improves the detection accuracy of the multi-target detection model obtained after the training is completed.
  • step S201 it may further include: obtaining global model parameters from the central server; and constructing an initial multi-target detection model according to the global model parameters.
  • the terminal needs to be initialized before training.
  • the terminal receives global model parameters from the central server.
  • the global model parameters acquired at this time are initialized global model parameters, for example, they can be randomly generated by the central server.
  • the terminal replaces the model parameters stored in the local multi-target detection model with the obtained global model parameters, thereby obtaining the initial multi-target detection model.
  • the terminal constructs an initial multi-target detection model according to the global model parameters issued by the central server, thereby realizing model initialization.
  • step S202 may include: inputting the target image in the local data set into the initial multi-target detection model to obtain the target prediction result; determining the prediction error according to the target prediction result and the image label in the local data set;
  • the detection model performs parameter adjustment; the initial multi-target detection model after parameter adjustment is used as the initial multi-target detection model for the next round of training for iterative training until the number of iterations reaches a preset value, and the relay multi-target detection model is obtained.
  • the target image may be an image about the target object.
  • the terminal separately extracts the target image and the image label from the local data set, inputs the target image into the initial multi-target detection model, and obtains the target prediction result.
  • the terminal calculates the prediction error according to the target prediction result and the image label, and adjusts the model parameters in the initial multi-target detection model with the goal of reducing the prediction error.
  • the terminal performs iterative training on the initial multi-target detection model after parameter adjustments according to the local data set. Each time the terminal adjusts the model parameters, an iteration is implemented until the number of iterations reaches the preset value, and the terminal stops iterating to obtain the relay multi-target detection model. .
  • the terminal uses the Focal Loss function when calculating the prediction error, and the Focal Loss function is as follows:
  • y is the image label
  • ⁇ is the adjustment factor
  • the initial multi-target detection model is iteratively trained for a preset number of times according to the target images and image tags in the local data set to obtain the relay multi-target detection model.
  • the relay multi-target detection model is used for federated learning, which ensures The realization of federated learning.
  • step S203 may include: generating an additional random number; wherein, the additional random number generated by each node in the alliance network is added to a value of zero; and the generated additional random number is combined with the model parameters of the relay multi-target detection model. Linear operation to obtain compound model parameters.
  • the terminal generates an additional random number.
  • the terminal is a node in the alliance network, and each node in the alliance network generates an additional random number, and the additional random number of each node is added to a value of zero.
  • the terminal extracts the model parameters of the relay multi-target detection model, and performs linear operations on the model parameters and the generated additional random numbers to obtain composite model parameters.
  • the compound model parameters are different from the model parameters of the relay multi-target detection model, and can play a role in data confidentiality for the relay multi-target detection model.
  • the terminal adds and subtracts the generated additional random number and the model parameters of the relay multi-target detection model, and the generated composite model parameters are as follows:
  • ⁇ k is the model parameter of the relay multi-target detection model
  • It is a compound model parameter
  • Suv is an additional random number
  • K is the set of all nodes in the alliance network
  • u and v represent nodes in the alliance network, and they are all elements in K.
  • the additional random number is linearly calculated with the model parameters of the relay multi-target detection model to obtain the composite model parameters, thereby realizing the data encryption of the model parameters.
  • step S204 may include: communicating with the central server to determine the encryption key; encrypting the composite model parameters according to the encryption key to obtain the encrypted model parameters; sending the encrypted model parameters to the central server to instruct the central server to The encrypted model parameters of each node are decrypted, and the compound model parameters obtained after decryption are calculated to generate global model parameters.
  • the encryption key may be a key used to encrypt the parameters of the composite model.
  • the terminal can communicate with the central server in advance, based on the DH key exchange protocol/algorithm (Diffie-Hellman Key Exchange/Agreement Algorithm, Diffie-Hellman Key Exchange/Agreement Algorithm), which can enable both parties who need to communicate securely to pass this method Determine the shared key) and the central server to determine the encryption key.
  • DH key exchange protocol/algorithm Diffie-Hellman Key Exchange/Agreement Algorithm, Diffie-Hellman Key Exchange/Agreement Algorithm
  • the composite model parameters can be encrypted according to the encryption key to obtain the encrypted model parameters, and the encrypted model parameters are sent to the central server, thereby ensuring the privacy and security of data in communication with the central server.
  • the central server After obtaining the encrypted model parameters of each node, the central server decrypts the encrypted model parameters according to the decryption key to obtain the composite model parameters of each node.
  • the central server adds the composite model parameters of each node, so that the additional random numbers in each group of composite model parameters are reset to zero, namely:
  • ⁇ k is the model parameter of the relay multi-target detection model, It is a compound model parameter, and K is the set of all nodes in the alliance network.
  • the central server then performs a weighted linear operation on the composite model parameters to obtain the global model parameters.
  • the weights of each group of composite model parameters can be the same or different; when they are not the same, the central server can calculate the global model parameters based on FedAvg's secure aggregation algorithm, the formula is as follows:
  • ⁇ k is the model parameter of the relay multi-target detection model
  • ⁇ is the global model parameter
  • n k is the data volume of the k-th node
  • n is the total data volume of each node in the alliance network
  • K is all the data in the alliance network.
  • a collection of nodes, k represents the k-th node, and t represents the t-th update.
  • the compound model parameters are encrypted according to the encryption key to obtain the encrypted model parameters to further protect the data privacy in the federated learning; after the encrypted model parameters are sent to the central server, the central server decrypts the encrypted model parameters, and according to The composite model parameters obtained after decryption generate global model parameters, and the global model parameters are used to update the relay multi-target detection model in each node, which ensures the realization of federated learning.
  • step S206 it may further include: acquiring the image to be detected; inputting the image to be detected into the multi-target detection model to obtain the target object in the image to be detected; and displaying the detected target object.
  • the image to be detected may be an image used to input a multi-target detection model for target detection.
  • the user when applying the multi-target detection model, the user can operate the terminal to collect the image to be detected through the image acquisition device of the terminal, or select an image stored in the terminal as the image to be detected, and instruct the terminal to perform target detection on the image to be detected.
  • the terminal inputs the image to be detected into the multi-target detection model, processes the image to be detected through the multi-target detection model, recognizes the target object in the image to be detected, and displays the target object on the screen.
  • the multi-target detection model can add a detection frame and object description to the target object.
  • the object description is used to display the target object category.
  • the detection frame can also select different colors according to the target object category to display the target more clearly Information about the object.
  • the terminal may also display the introduction and related information of the target object of the category.
  • the terminal can display the garbage items in the image to be detected, and the user clicks on the detected garbage item, and the terminal can introduce the garbage category to which the garbage item belongs and provide the category Suggestions on the classification of garbage items so that users can better classify garbage.
  • the multi-target detection model for target detection on the image to be detected is obtained through federated learning, and a rich data set is used for training in federated learning, which improves the accuracy of target detection.
  • the detected target object may further include: when a triggered calibration instruction is received, displaying a calibration information input page; acquiring the image to be calibrated and the calibration instruction information input in the calibration information input page; A calibration data set is generated according to the image to be calibrated and the calibration instruction information; the multi-target detection model is calibrated and trained through the calibration data set.
  • the terminal provides a model calibration function.
  • the user can click the virtual calibration button on the display page to trigger the calibration instruction.
  • the terminal displays a calibration information input page, and instructs the user to input the image to be calibrated and the calibration instruction information in the calibration information input page.
  • the terminal can also directly use the image to be detected during target detection as the image to be calibrated.
  • the user can input text to describe the location, shape, color, size, name and other information of the target object, or directly circle the target object in the image to be calibrated by means of a recognition frame on the screen to obtain calibration instruction information.
  • the terminal generates a calibration data set according to the image to be calibrated and the calibration instruction information.
  • the calibration data set includes the image to be calibrated and the corresponding image label.
  • the calibration data set is used to calibrate the multi-target detection model to improve the detection of the multi-target detection model. accuracy.
  • the terminal can only perform calibration training on the multi-target detection model locally; it can also perform calibration training on the multi-target detection model immediately through federated learning; or after a preset time or after a preset number of calibration training locally , And then train the multi-target detection model through federated learning.
  • the calibration information input page is displayed, and the calibration data set is generated according to the image to be calibrated and the calibration instruction information input in the calibration information input page to perform calibration training on the multi-target detection model, which improves The accuracy of multi-target detection model detection.
  • a garbage classification application is installed in the terminal, and the user can take pictures of garbage through the terminal to expand the local data set.
  • the terminal obtains the global model parameters from the central server, and obtains the initialized multi-target detection model.
  • the initial multi-target detection model is trained for N (N is an integer greater than zero) rounds, and the relay multi-target detection model is obtained.
  • the terminal determines the key required for communication with the central server through the DH key exchange protocol.
  • the terminal generates additional random numbers, and adds the additional random numbers and the model parameters of the relay multi-target detection model to obtain composite model parameters.
  • the terminal uses the encryption key to encrypt the composite model parameters to obtain the encrypted model parameters, and sends the encrypted model parameters to the central server.
  • the central server decrypts the encrypted model parameters, obtains the composite model parameters of each node, and adds the composite model parameters to eliminate the influence of additional random numbers.
  • the central server can calculate global model parameters according to FedAvg's secure aggregation algorithm, and deliver the global model parameters to each node, so that each node can update the relay multi-target detection model according to the global model parameters.
  • the terminal performs iterative training on the updated relay multi-target detection model according to the local data set until the model converges to obtain the multi-target detection model.
  • the user can take pictures of junk items through the terminal to obtain the image to be detected.
  • the multi-target detection image can identify multiple junk items in the image to be detected at one time and display the identified junk items.
  • the garbage items can be marked with a detection frame and the category of the garbage item is displayed. For example, it is shown that the garbage item is recyclable or hazardous. Different types of garbage items can be distinguished by detection frames of different colors.
  • the user When the user believes that the target detection is accurate, that is, the garbage classification is correct, they can click on the detected garbage item, and the terminal will introduce the category of garbage items and display the garbage classification instructions, so that the user can better classify the garbage.
  • the user thinks that the target detection is inaccurate, that is, the garbage classification is incorrect he can click the virtual calibration button, upload the image that the user thinks the detection is inaccurate as the image to be calibrated on the calibration information input page, and enter the explanatory text as the calibration instruction information.
  • the image to be calibrated and the calibration instruction information perform calibration training on the multi-target detection model.
  • the target detection model processing method based on intelligent decision-making in this application involves neural networks, machine learning, predictive analysis, and computer vision in the field of artificial intelligence; in addition, it may also involve smart environmental protection in the field of smart cities.
  • the processes in the above-mentioned embodiment methods can be implemented by instructing relevant hardware through computer-readable instructions, which can be stored in a computer-readable storage medium.
  • the computer-readable instructions When executed, they may include the processes of the above-mentioned method embodiments.
  • the aforementioned storage medium may be a non-volatile storage medium such as a magnetic disk, an optical disc, a read-only memory (Read-Only Memory, ROM), or a random access memory (Random Access Memory, RAM), etc.
  • this application provides an embodiment of a device for processing a target detection model based on intelligent decision-making, where the target detection model is a multi-target detection model, and the device implements
  • the target detection model is a multi-target detection model
  • the device implements
  • the example corresponds to the method embodiment shown in FIG. 2, and the device can be specifically applied to various electronic devices.
  • the device 300 for processing a target detection model based on intelligent decision-making in this embodiment includes: an acquisition module 301, a model training module 302, a parameter calculation module 303, a parameter transmission module 304, a model update module 305, and iteration Training module 306, where:
  • the obtaining module 301 is used to obtain a local data set and an initial multi-target detection model.
  • the model training module 302 is used to train the initial multi-target detection model according to the local data set to obtain the relay multi-target detection model.
  • the parameter calculation module 303 is used to generate additional random numbers, and calculate the composite model parameters according to the additional random numbers and the model parameters of the relay multi-target detection model.
  • the parameter sending module 304 is configured to send the composite model parameters to the central server to instruct the central server to generate global model parameters according to the composite model parameters of each node.
  • the model update module 305 is configured to receive global model parameters from the central server to update the relay multi-target detection model.
  • the iterative training module 306 is configured to use the updated relay multi-target detection model as the initial multi-target detection model for the next round of training for iterative training until the model converges to obtain a multi-target detection model.
  • the initial multi-target detection model is trained according to the local data set to obtain the relay multi-target detection model; additional random numbers are generated, and the additional random numbers are used to perform calculations on the model parameters of the relay multi-target detection model.
  • Model parameters are encrypted; the encrypted model parameters are sent to the central server in the alliance network.
  • the central server generates global model parameters based on the composite model parameters of each node.
  • the global model parameters are sent to the terminal to update the relay data.
  • Target detection model and iterative training global model parameters are obtained on the basis of the local data set of each terminal, and the characteristics of multiple local data sets are integrated. On the basis of protecting the local data set, a larger scale is achieved.
  • the multi-target detection model is trained on the data set of, which improves the detection accuracy of the multi-target detection model obtained after the training is completed.
  • the processing device 300 of the target detection model based on intelligent decision further includes: a parameter acquisition module and a model construction module, wherein:
  • the parameter acquisition module is used to acquire global model parameters from the central server.
  • the model building module is used to build the initial multi-target detection model according to the global model parameters.
  • the terminal constructs an initial multi-target detection model according to the global model parameters issued by the central server, thereby realizing model initialization.
  • the model training module 302 includes: an image input submodule, an error determination submodule, a parameter adjustment submodule, and an iterative training submodule, where:
  • the image input sub-module is used to input the target image in the local data set into the initial multi-target detection model to obtain the target prediction result.
  • the error determination sub-module is used to determine the prediction error according to the target prediction result and the image label in the local data set.
  • the parameter adjustment sub-module is used to adjust the parameters of the initial multi-target detection model based on the prediction error.
  • the iterative training sub-module is used to use the adjusted initial multi-target detection model as the initial multi-target detection model for the next round of training for iterative training until the number of iterations reaches a preset value to obtain the relay multi-target detection model.
  • the initial multi-target detection model is iteratively trained for a preset number of times according to the target images and image tags in the local data set to obtain the relay multi-target detection model.
  • the relay multi-target detection model is used for federated learning, which ensures The realization of federated learning.
  • the parameter calculation module 303 includes: an additional generation sub-module and a parameter operation sub-module, wherein:
  • the additional generation sub-module is used to generate additional random numbers; among them, the additional random numbers generated by each node in the alliance network are added to a value of zero.
  • the parameter operation sub-module is used to perform linear operation on the generated additional random number and the model parameters of the relay multi-target detection model to obtain the composite model parameters.
  • the additional random number is linearly calculated with the model parameters of the relay multi-target detection model to obtain the composite model parameters, thereby realizing the data encryption of the model parameters.
  • the parameter sending module 304 includes: a key determination submodule, a parameter encryption submodule, and a parameter sending submodule, where:
  • the key determination sub-module is used to communicate with the central server to determine the encryption key.
  • the parameter encryption sub-module is used to encrypt the compound model parameters according to the encryption key to obtain the encrypted model parameters.
  • the parameter sending submodule is used to send the encrypted model parameters to the central server to instruct the central server to decrypt the encrypted model parameters of each node, and perform calculations based on the composite model parameters obtained after decryption to generate global model parameters.
  • the compound model parameters are encrypted according to the encryption key to obtain the encrypted model parameters to further protect the data privacy in the federated learning; after the encrypted model parameters are sent to the central server, the central server decrypts the encrypted model parameters, and according to The composite model parameters obtained after decryption generate global model parameters, and the global model parameters are used to update the relay multi-target detection model in each node, which ensures the realization of federated learning.
  • the processing device 300 of the target detection model based on intelligent decision further includes: an image acquisition module, an image input module, and a target display module, wherein:
  • the image acquisition module is used to acquire the image to be detected.
  • the image input module is used to input the image to be detected into the multi-target detection model to obtain the target object in the image to be detected.
  • the target display module is used to display the detected target object.
  • the multi-target detection model for target detection on the image to be detected is obtained through federated learning, and a rich data set is used for training in federated learning, which improves the accuracy of target detection.
  • the processing device 300 of the target detection model based on intelligent decision further includes: a page display module, a calibration acquisition module, a calibration generation module, and a calibration training module, wherein:
  • the page display module is used to display the calibration information input page when the triggered calibration instruction is received.
  • the calibration acquisition module is used to acquire the image to be calibrated and the calibration instruction information entered in the calibration information input page.
  • the calibration generating module is used to generate a calibration data set according to the image to be calibrated and the calibration instruction information.
  • the calibration training module is used to calibrate the multi-target detection model through the calibration data set.
  • the calibration information input page is displayed, and the calibration data set is generated according to the image to be calibrated and the calibration instruction information input in the calibration information input page to perform calibration training on the multi-target detection model, which improves The accuracy of multi-target detection model detection.
  • FIG. 4 is a block diagram of the basic structure of the computer device in this embodiment.
  • the computer device 4 includes a memory 41, a processor 42, and a network interface 43 that are connected to each other in communication via a system bus. It should be pointed out that the figure only shows the computer device 4 with components 41-43, but it should be understood that it is not required to implement all the shown components, and more or fewer components may be implemented instead. Among them, those skilled in the art can understand that the computer device here is a device that can automatically perform numerical calculation and/or information processing in accordance with pre-set or stored instructions.
  • Its hardware includes, but is not limited to, a microprocessor, a dedicated Integrated Circuit (Application Specific Integrated Circuit, ASIC), Programmable Gate Array (Field-Programmable Gate Array, FPGA), Digital Processor (Digital Signal Processor, DSP), embedded equipment, etc.
  • ASIC Application Specific Integrated Circuit
  • ASIC Application Specific Integrated Circuit
  • FPGA Field-Programmable Gate Array
  • DSP Digital Processor
  • the computer device may be a computing device such as a desktop computer, a notebook, a palmtop computer, and a cloud server.
  • the computer device can interact with the user through a keyboard, a mouse, a remote control, a touch panel, or a voice control device.
  • the memory 41 includes at least one type of computer-readable storage medium.
  • the computer-readable storage medium may be nonvolatile or volatile.
  • the computer-readable storage medium includes flash memory, hard disk, and multimedia card. , Card-type memory (for example, SD or DX memory, etc.), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), Programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc.
  • the memory 41 may be an internal storage unit of the computer device 4, such as a hard disk or memory of the computer device 4.
  • the memory 41 may also be an external storage device of the computer device 4, such as a plug-in hard disk equipped on the computer device 4, a smart memory card (Smart Media Card, SMC), and a secure digital (Secure Digital, SD) card, Flash Card, etc.
  • the memory 41 may also include both the internal storage unit of the computer device 4 and its external storage device.
  • the memory 41 is generally used to store an operating system and various application software installed in the computer device 4, such as computer readable instructions for a method for processing a target detection model based on intelligent decision-making.
  • the memory 41 can also be used to temporarily store various types of data that have been output or will be output.
  • the processor 42 may be a central processing unit (Central Processing Unit, CPU), a controller, a microcontroller, a microprocessor, or other data processing chips in some embodiments.
  • the processor 42 is generally used to control the overall operation of the computer device 4.
  • the processor 42 is configured to run computer-readable instructions or processed data stored in the memory 41, for example, run the computer-readable instructions of the method for processing a target detection model based on intelligent decision-making.
  • the network interface 43 may include a wireless network interface or a wired network interface, and the network interface 43 is generally used to establish a communication connection between the computer device 4 and other electronic devices.
  • the computer device provided in this embodiment can execute the aforementioned intelligent decision-based target detection model processing method.
  • the target detection model processing method based on intelligent decision herein may be the target detection model processing method based on intelligent decision in each of the above embodiments.
  • the initial multi-target detection model is trained according to the local data set to obtain the relay multi-target detection model; additional random numbers are generated, and the additional random numbers are used to perform calculations on the model parameters of the relay multi-target detection model.
  • Model parameters are encrypted; the encrypted model parameters are sent to the central server in the alliance network.
  • the central server generates global model parameters based on the composite model parameters of each node.
  • the global model parameters are sent to the terminal to update the relay data.
  • Target detection model and iterative training global model parameters are obtained on the basis of the local data set of each terminal, and the characteristics of multiple local data sets are integrated. On the basis of protecting the local data set, a larger scale is achieved.
  • the multi-target detection model is trained on the data set of, which improves the detection accuracy of the multi-target detection model obtained after the training is completed.
  • the present application also provides another implementation manner, that is, a computer-readable storage medium is provided with computer-readable instructions stored thereon, and the computer-readable instructions can be executed by at least one processor to The at least one processor is caused to execute the steps of the target detection model processing method based on intelligent decision as described above.
  • the initial multi-target detection model is trained according to the local data set to obtain the relay multi-target detection model; additional random numbers are generated, and the additional random numbers are used to perform calculations on the model parameters of the relay multi-target detection model.
  • Model parameters are encrypted; the encrypted model parameters are sent to the central server in the alliance network.
  • the central server generates global model parameters based on the composite model parameters of each node.
  • the global model parameters are sent to the terminal to update the relay data.
  • Target detection model and iterative training global model parameters are obtained on the basis of the local data set of each terminal, and the characteristics of multiple local data sets are integrated. On the basis of protecting the local data set, a larger scale is achieved.
  • the multi-target detection model is trained on the data set of, which improves the detection accuracy of the multi-target detection model obtained after the training is completed.
  • the technical solution of this application essentially or the part that contributes to the existing technology can be embodied in the form of a software product, and the computer software product is stored in a storage medium (such as ROM/RAM, magnetic disk, The optical disc) includes several instructions to make a terminal (which may be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) execute the methods described in the various embodiments of the present application.
  • a terminal which may be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Bioethics (AREA)
  • General Health & Medical Sciences (AREA)
  • Computer Hardware Design (AREA)
  • Computer Security & Cryptography (AREA)
  • Software Systems (AREA)
  • Image Analysis (AREA)

Abstract

一种基于智能决策的目标检测模型处理方法、装置、计算机设备及存储介质，所述方法包括：根据获取的本地数据集对初始多目标检测模型进行训练，得到中继多目标检测模型（S202）；根据生成的附加随机数和中继多目标检测模型的模型参数计算复合模型参数（S203）；将复合模型参数发送至中央服务器，以指示中央服务器根据各节点的复合模型参数生成全局模型参数（S204）；接收全局模型参数，以更新中继多目标检测模型（S205）；将更新后的中继多目标检测模型作为下轮训练的初始多目标检测模型进行迭代训练，直至模型收敛，得到多目标检测模型（S206）。此外，还涉及区块链技术，本地数据集可存储于区块链中。提高了目标检测的准确率。

Description

基于智能决策的目标检测模型的处理方法、及其相关设备
本申请要求于2020年09月18日提交中国专利局、申请号为202010990029.4，发明名称为“基于智能决策的目标检测模型的处理方法、及其相关设备”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及人工智能技术领域，尤其涉及一种基于智能决策的目标检测模型的处理方法、及其相关设备。
背景技术
随着人工智能的发展，目标检测在生活及生产中的应用越来越广泛。目标检测涉及智能决策中的检测模型，可以将图像输入目标检测模型，由目标检测模型对图像进行处理并输出图像中的目标对象。例如，在垃圾分类应用中，可以先将垃圾图片输入目标检测模型以对垃圾进行识别，从而指导人们对识别到的垃圾进行回收分类。
发明人意识到，传统的目标检测技术中，目标检测模型在检测时通常只能聚焦于单一目标进行检测，而且在训练目标检测模型时，因为本地数据集的数据量通常有限，目标检测模型难以得到充分的训练，使得目标检测模型的检测准确率较低。
发明内容
本申请实施例的目的在于提出一种基于智能决策的目标检测模型的处理方法、装置、计算机设备及存储介质，以解决目标检测模型的检测准确率较低的问题。
为了解决上述技术问题，本申请实施例提供一种基于智能决策的目标检测模型的处理方法，其中，所述目标检测模型为多目标检测模型，采用了如下所述的技术方案：
获取本地数据集以及初始多目标检测模型；
根据所述本地数据集对所述初始多目标检测模型进行训练，得到中继多目标检测模型；
生成附加随机数，并根据所述附加随机数以及所述中继多目标检测模型的模型参数计算复合模型参数；
将所述复合模型参数发送至中央服务器，以指示所述中央服务器根据各节点的复合模型参数生成全局模型参数；
从所述中央服务器接收所述全局模型参数，以更新所述中继多目标检测模型；
将更新后的中继多目标检测模型作为下轮训练的初始多目标检测模型进行迭代训练，直至模型收敛，得到所述多目标检测模型。
为了解决上述技术问题，本申请实施例还提供一种基于智能决策的目标检测模型的处理装置，其中，所述目标检测模型为多目标检测模型，采用了如下所述的技术方案：
获取模块，用于获取本地数据集以及初始多目标检测模型；
模型训练模块，用于根据所述本地数据集对所述初始多目标检测模型进行训练，得到中继多目标检测模型；
参数计算模块，用于生成附加随机数，并根据所述附加随机数以及所述中继多目标检测模型的模型参数计算复合模型参数；
参数发送模块，用于将所述复合模型参数发送至中央服务器，以指示所述中央服务器根据各节点的复合模型参数生成全局模型参数；
模型更新模块，用于从所述中央服务器接收所述全局模型参数，以更新所述中继多目标检测模型；
迭代训练模块，用于将更新后的中继多目标检测模型作为下轮训练的初始多目标检测 模型进行迭代训练，直至模型收敛，得到所述多目标检测模型。
为了解决上述技术问题，本申请实施例还提供一种计算机设备，包括存储器和处理器，所述存储器中存储有计算机可读指令，所述处理器执行所述计算机可读指令时实现如下步骤：
获取本地数据集以及初始多目标检测模型；
根据所述本地数据集对所述初始多目标检测模型进行训练，得到中继多目标检测模型；
生成附加随机数，并根据所述附加随机数以及所述中继多目标检测模型的模型参数计算复合模型参数；
将所述复合模型参数发送至中央服务器，以指示所述中央服务器根据各节点的复合模型参数生成全局模型参数；
从所述中央服务器接收所述全局模型参数，以更新所述中继多目标检测模型；
将更新后的中继多目标检测模型作为下轮训练的初始多目标检测模型进行迭代训练，直至模型收敛，得到所述多目标检测模型。
为了解决上述技术问题，本申请实施例还提供一种计算机可读存储介质，所述计算机可读存储介质存储有计算机可读指令，所述计算机可读指令被处理器执行时实现如下步骤：
获取本地数据集以及初始多目标检测模型；
根据所述本地数据集对所述初始多目标检测模型进行训练，得到中继多目标检测模型；
生成附加随机数，并根据所述附加随机数以及所述中继多目标检测模型的模型参数计算复合模型参数；
将所述复合模型参数发送至中央服务器，以指示所述中央服务器根据各节点的复合模型参数生成全局模型参数；
从所述中央服务器接收所述全局模型参数，以更新所述中继多目标检测模型；
将更新后的中继多目标检测模型作为下轮训练的初始多目标检测模型进行迭代训练，直至模型收敛，得到所述多目标检测模型。
与现有技术相比，本申请实施例主要有以下有益效果：先根据本地数据集对初始多目标检测模型进行训练得到中继多目标检测模型；生成附加随机数，附加随机数用于对中继多目标检测模型的模型参数进行运算从而对模型参数进行加密；加密后得到的符合模型参数被发送至联盟网络中的中央服务器，中央服务器根据各节点的复合模型参数生成全局模型参数，全局模型参数被下发至终端后以更新中继多目标检测模型并进行迭代训练；全局模型参数是在各终端的本地数据集的基础上得到的，融合了多个本地数据集的特征，在保护本地数据集的基础上，实现了以更大规模的数据集对多目标检测模型进行训练，提高了训练完毕后得到的多目标检测模型检测的准确率。
附图说明
为了更清楚地说明本申请中的方案，下面将对本申请实施例描述中所需要使用的附图作一个简单介绍，显而易见地，下面描述中的附图是本申请的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。
图1是本申请可以应用于其中的示例性系统架构图；
图2是根据本申请的基于智能决策的目标检测模型处理方法的一个实施例的流程图；
图3是根据本申请的基于智能决策的目标检测模型处理装置的一个实施例的结构示意图；
图4是根据本申请的计算机设备的一个实施例的结构示意图。
具体实施方式
除非另有定义，本文所使用的所有的技术和科学术语与属于本申请的技术领域的技术人员通常理解的含义相同；本文中在申请的说明书中所使用的术语只是为了描述具体的实施例的目的，不是旨在于限制本申请；本申请的说明书和权利要求书及上述附图说明中的术语“包括”和“具有”以及它们的任何变形，意图在于覆盖不排他的包含。本申请的说明书和权利要求书或上述附图中的术语“第一”、“第二”等是用于区别不同对象，而不是用于描述特定顺序。
在本文中提及“实施例”意味着，结合实施例描述的特定特征、结构或特性可以包含在本申请的至少一个实施例中。在说明书中的各个位置出现该短语并不一定均是指相同的实施例，也不是与其它实施例互斥的独立的或备选的实施例。本领域技术人员显式地和隐式地理解的是，本文所描述的实施例可以与其它实施例相结合。
为了使本技术领域的人员更好地理解本申请方案，下面将结合附图，对本申请实施例中的技术方案进行清楚、完整地描述。
如图1所示，系统架构100可以包括终端101、102、103，网络104和服务器105。网络104用以在终端101、102、103和服务器105之间提供通信链路的介质。网络104可以包括各种连接类型，例如有线、无线通信链路或者光纤电缆等等。
用户可以使用终端101、102、103通过网络104与服务器105交互，以接收或发送消息等。终端101、102、103上可以安装有各种通讯客户端应用，例如网页浏览器应用、购物类应用、搜索类应用、即时通信工具、邮箱客户端、社交平台软件等。
终端101、102、103可以是具有显示屏并且支持网页浏览的各种电子设备，包括但不限于智能手机、平板电脑、电子书阅读器、MP3播放器(Moving Picture Experts Group Audio Layer III，动态影像专家压缩标准音频层面3)、MP4(Moving Picture Experts Group Audio Layer IV，动态影像专家压缩标准音频层面4)播放器、膝上型便携计算机和台式计算机等等。
服务器105可以是提供各种服务的服务器，例如对终端101、102、103上显示的页面提供支持的后台服务器。
需要说明的是，本申请实施例所提供的基于智能决策的目标检测模型的处理方法一般由终端执行，相应地，基于智能决策的目标检测模型的处理装置一般设置于终端中。
应该理解，图1中的终端、网络和服务器的数目仅仅是示意性的。根据实现需要，可以具有任意数目的终端、网络和服务器。
继续参考图2，示出了根据本申请的基于智能决策的目标检测模型的处理方法的一个实施例的流程图。所述的目标检测模型为多目标检测模型，所述的基于智能决策的目标检测模型的处理方法，包括以下步骤：
步骤S201，获取本地数据集以及初始多目标检测模型。
在本实施例中，基于智能决策的目标检测模型的处理方法运行于其上的电子设备(例如图1所示的终端)可以通过有线连接方式或者无线连接方式与其他终端或服务器进行通信。需要指出的是，上述无线连接方式可以包括但不限于3G/4G连接、WiFi连接、蓝牙连接、WiMAX连接、Zigbee连接、UWB(ultra wideband)连接、以及其他现在已知或将来开发的无线连接方式。
其中，本地数据集可以是存储在终端中的数据集；初始多目标检测模型可以是初始的多目标检测模型。
具体地，终端中可以安装有用于目标检测的应用程序，当应用程序启动后，终端加载位于本地的初始多目标检测模型，并获取存储在终端中的本地数据集。
本申请使用多目标检测模型，多目标检测模型可以一次性检测出多个目标，从而提高目标检测的效率。在一个实施例中，多目标检测模型可以是RetinaNet网络。
本申请中的多目标检测模型为轻量级的模型，除了部署在服务器中，还可以部署在各 种终端中，终端的持有者可以根据自身需要扩充本地数据集，例如可以通过拍照或者从互联网获取图像扩充数据集，降低了本地数据集的扩充难度，丰富了本地数据集的数据量。
在一个实施例中，基于智能决策的目标检测模型的处理方法还可以由服务器执行，服务器加载初始多目标检测模型并获取本地数据集。服务器训练完成后提供目标检测接口，用户可以在终端调用目标检测接口进行目标检测。
需要强调的是，为进一步保证上述本地数据集的私密和安全性，上述本地数据集还可以存储于一区块链的节点中。
本申请所指区块链是分布式数据存储、点对点传输、共识机制、加密算法等计算机技术的新型应用模式。区块链(Blockchain)，本质上是一个去中心化的数据库，是一串使用密码学方法相关联产生的数据块，每一个数据块中包含了一批次网络交易的信息，用于验证其信息的有效性(防伪)和生成下一个区块。区块链可以包括区块链底层平台、平台产品服务层以及应用服务层等。
步骤S202，根据本地数据集对初始多目标检测模型进行训练，得到中继多目标检测模型。
具体地，终端根据本地数据集对初始多目标检测模型进行训练，并在满足训练停止条件时停止训练，得到中继多目标检测模型。训练停止条件可以是训练中的迭代次数达到预设数值，或者训练中得到的预测误差小于预设的误差阈值。
终端得到的中继多目标检测模型可以是收敛的，也可以是不收敛的。
步骤S203，生成附加随机数，并根据附加随机数以及中继多目标检测模型的模型参数计算复合模型参数。
其中，附加随机数可以是对中继多目标检测模型的模型参数实现加密功能的参数。
具体地，终端进行联邦学习，终端可以是联盟网络中的一个节点。联盟网络中的每个节点均生成附加随机数，且得到的附加随机数相加后数值为零。终端提取中继多目标检测模型的模型参数，将附加随机数与模型参数进行计算，得到复合模型参数，以保护本地的中继多目标检测模型中模型参数的数据隐私。
步骤S204，将复合模型参数发送至中央服务器，以指示中央服务器根据各节点的复合模型参数生成全局模型参数。
其中，中央服务器可以是联邦学习中起到中央控制作用的服务器，用于指示各节点进行联邦学习。
具体地，终端将复合模型参数发送至中央服务器。中央服务器接收到各节点的复合模型参数后，可以将各模型的复合模型参数进行线性运算，在一个实施例中，中央服务器对各模型的复合模型参数求平均值，得到全局模型参数。
步骤S205，从中央服务器接收全局模型参数，以更新中继多目标检测模型。
具体地，中央服务器将全局模型参数发送至各节点。当终端接收到全局模型参数后，根据全局模型参数更新本地的中继多目标检测模型，具体为将中继多目标检测模型中的模型参数替换为全局模型参数。
步骤S206，将更新后的中继多目标检测模型作为下轮训练的初始多目标检测模型进行迭代训练，直至模型收敛，得到多目标检测模型。
具体地，终端对中继多目标检测模型完成更新后，将中继多目标检测模型作为初始多目标检测模型，并继续根据本地数据集对得到的初始多目标检测模型进行训练，即迭代步骤S202至步骤S206，直至模型收敛，终端停止训练，得到多目标检测模型。其中，模型收敛的条件可以是训练中得到的预测误差小于预设的误差阈值。
在一个实施例中，终端和中央服务器间传递的是梯度信息。终端将附加随机数和中继多目标检测模型的模型梯度进行运算得到符合模型梯度，将符合模型梯度发送至中央服务器。中央服务器对符合模型梯度进行累加，然后求平均值，将平均值作为全局平均梯度下发至各节点，以更新中继多目标检测模型。
在一个实施例中，当联盟网络中各节点均实现模型收敛时，各节点停止训练，将停止训练时的中继多目标检测模型作为各节点的多目标检测模型。
本实施例中，先根据本地数据集对初始多目标检测模型进行训练得到中继多目标检测模型；生成附加随机数，附加随机数用于对中继多目标检测模型的模型参数进行运算从而对模型参数进行加密；加密后得到的符合模型参数被发送至联盟网络中的中央服务器，中央服务器根据各节点的复合模型参数生成全局模型参数，全局模型参数被下发至终端后以更新中继多目标检测模型并进行迭代训练；全局模型参数是在各终端的本地数据集的基础上得到的，融合了多个本地数据集的特征，在保护本地数据集的基础上，实现了以更大规模的数据集对多目标检测模型进行训练，提高了训练完毕后得到的多目标检测模型检测的准确率。
进一步的，上述步骤S201之前还可以包括：从中央服务器获取全局模型参数；根据全局模型参数构建初始多目标检测模型。
具体地，终端在进行训练之前，需要先进行初始化。在初始化时，终端从中央服务器接收全局模型参数，此时获取的全局模型参数是初始化的全局模型参数，比如，可以由中央服务器随机生成。终端将存储在本地的多目标检测模型中的模型参数替换为得到的全局模型参数，从而得到初始多目标检测模型。
本实施例中，终端根据中央服务器下发的全局模型参数构建初始多目标检测模型，从而实现模型初始化。
进一步的，上述步骤S202可以包括：将本地数据集中的目标图像输入初始多目标检测模型，得到目标预测结果；根据目标预测结果以及本地数据集中的图像标签确定预测误差；基于预测误差对初始多目标检测模型进行参数调整；将参数调整后的初始多目标检测模型作为下轮训练的初始多目标检测模型进行迭代训练，直至迭代次数达到预设数值，得到中继多目标检测模型。
其中，目标图像可以是关于目标对象的图像。
具体地，终端从本地数据集中分别提取目标图像以及图像标签，将目标图像输入初始多目标检测模型，得到目标预测结果。终端根据目标预测结果以及图像标签计算预测误差，以减小预测误差为目标调整初始多目标检测模型中的模型参数。
终端根据本地数据集对参数调整后的初始多目标检测模型进行迭代训练，终端每调整一次模型参数，便实现一次迭代，直至迭代次数达到预设数值，终端停止迭代，得到中继多目标检测模型。
在一个实施例中，终端计算预测误差时采用Focal Loss损失函数，Focal Loss损失函数如下：
Figure PCTCN2020124439-appb-000001
其中，y为图像标签，γ为调整因子。
本实施例中，根据本地数据集中的目标图像以及图像标签对初始多目标检测模型进行预设次数的迭代训练，得到中继多目标检测模型，中继多目标检测模型用于联邦学习，保证了联邦学习的实现。
进一步的，上述步骤S203可以包括：生成附加随机数；其中，联盟网络中各节点生成的附加随机数相加后值为零；将生成的附加随机数与中继多目标检测模型的模型参数进行线性运算，得到复合模型参数。
具体地，终端生成附加随机数。终端为联盟网络中的一个节点，联盟网络中的各节点均生成附加随机数，且各节点的附加随机数相加后数值为零。终端提取中继多目标检测模型的模型参数，将模型参数与生成的附加随机数做线性运算，得到复合模型参数。复合模型参数不同于中继多目标检测模型的模型参数，可以对中继多目标检测模型起到数据保密的作用。
在一个实施例中，终端将生成的附加随机数与中继多目标检测模型的模型参数进行加减运算，生成的复合模型参数如下：
Figure PCTCN2020124439-appb-000002
其中，ω k为中继多目标检测模型的模型参数，
Figure PCTCN2020124439-appb-000003
为复合模型参数，S uv为附加随机数，K为联盟网络中所有节点的集合，u和v表示联盟网络中的节点，都是K中的元素。
本实施例中，生成附加随机数后，将附加随机数与中继多目标检测模型的模型参数进行线性运算，得到复合模型参数，从而实现对模型参数的数据加密。
进一步的，上述步骤S204可以包括：与中央服务器进行通信以确定加密密钥；根据加密密钥对复合模型参数进行加密，得到加密模型参数；将加密模型参数发送至中央服务器，以指示中央服务器对各节点的加密模型参数进行解密，并根据解密后得到的复合模型参数进行运算，生成全局模型参数。
其中，加密密钥可以是用于对复合模型参数进行加密的密钥。
具体地，终端可以预先与中央服务器进行通信，基于DH密钥交换协议/算法(Diffie-Hellman密钥交换协议/算法，Diffie-Hellman Key Exchange/Agreement Algorithm，可以使得需要安全通信的双方通过该方法确定共享密钥)与中央服务器确定加密密钥。
终端将复合模型参数发送至中央服务器时，可以根据加密密钥再对复合模型参数进行加密得到加密模型参数，将加密模型参数发送至中央服务器，从而确保与中央服务器通信中的数据隐私安全。
中央服务器得到各节点的加密模型参数后，根据解密密钥对加密模型参数进行解密，得到各节点的复合模型参数。中央服务器将各节点的复合模型参数进行加法运算，使得各组复合模型参数中的附加随机数归零，即：
Figure PCTCN2020124439-appb-000004
其中，ω k为中继多目标检测模型的模型参数，
Figure PCTCN2020124439-appb-000005
为复合模型参数，K为联盟网络中所有节点的集合。
中央服务器再对复合模型参数进行加权的线性运算，得到全局模型参数。
计算全局模型参数时，各组复合模型参数的权重可以相同，也可以不相同；当不相同时，中央服务器可以基于FedAvg的安全聚合算法计算全局模型参数，公式如下：
Figure PCTCN2020124439-appb-000006
其中，ω k为中继多目标检测模型的模型参数，ω为全局模型参数，n k为第k个节点的数据量，n为联盟网络中各节点的总数据量，K为联盟网络中所有节点的集合，k表示第k个节点，t表示第t次更新。
本实施例中，根据加密密钥对复合模型参数进行加密得到加密模型参数，以进一步保护联邦学习中的数据隐私；加密模型参数发送至中央服务器后，中央服务器对加密模型参数进行解密，并根据解密后得到的复合模型参数生成全局模型参数，全局模型参数用于更新各节点中的中继多目标检测模型，保证了联邦学习的实现。
进一步的，上述步骤S206之后还可以包括：获取待检测图像；将待检测图像输入多目标检测模型，得到待检测图像中的目标对象；将检测到的目标对象进行展示。
其中，待检测图像可以是用于输入多目标检测模型进行目标检测的图像。
具体地，在应用多目标检测模型时，用户可以操作终端，通过终端的图像采集装置采集待检测图像，或者选择终端中已经存储的图像作为待检测图像，并指示终端对待检测图像进行目标检测。
终端将待检测图像输入多目标检测模型，通过多目标检测模型对待检测图像进行处理，识别到待检测图像中的目标对象，并通过屏幕展示目标对象。展示目标对象时，多目 标检测模型可以给目标对象添加检测框以及对象说明，对象说明用于显示目标对象的类别，检测框还可以依据目标对象的类别选取不同的颜色，以便更清晰地显示目标对象的信息。
在一个实施例中，当用户点击展示的目标对象时，终端还可以展示该类别目标对象的介绍以及相关信息。例如，当多目标检测模型应用于垃圾分类的应用时，终端可以展示待检测图像中的垃圾物品，用户点击检测到的垃圾物品，终端可以对垃圾物品所属的垃圾类别进行介绍，并提供该类别垃圾物品的分类建议，以便用户更好地进行垃圾分类。
本实施例中，对待检测图像进行目标检测的多目标检测模型通过联邦学习得到，并在联邦学习中使用了丰富的数据集进行训练，提高了目标检测的准确性。
进一步的，上述将检测到的目标对象进行展示之后，还可以包括：当接收到触发的校准指令时，展示校准信息输入页面；获取在校准信息输入页面中输入的待校准图像以及校准指示信息；根据待校准图像以及校准指示信息生成校准数据集；通过校准数据集对多目标检测模型进行校准训练。
具体地，终端提供模型校准功能，当用户认为多目标检测模型的检测结果不准确时，可以点击展示页面中的虚拟校准按钮，触发校准指令。终端接收到校准指令后，展示校准信息输入页面，指示用户在校准信息输入页面中输入待校准图像以及校准指示信息。终端还可以直接将目标检测时的待检测图像作为待校准图像。用户可以输入文字，描述目标对象的位置、形状、颜色、大小、名称等信息，或者直接通过屏幕在待校准图像中以识别框的方式圈出目标对象，得到校准指示信息。
终端根据待校准图像以及校准指示信息生成校准数据集，校准数据集中包括了待校准图像以及对应的图像标签，校准数据集用于对多目标检测模型进行校准训练，以提升多目标检测模型的检测准确性。
终端可以仅在本地对多目标检测模型进行校准训练；也可以即刻通过联邦学习的方式对多目标检测模型进行校准训练；也可以在预设时间后，或在本地进行预设次数的校准训练后，再通过联邦学习的方式对多目标检测模型进行训练。
本实施例中，当接收到校准指令时，展示校准信息输入页面，根据校准信息输入页面中输入的待校准图像以及校准指示信息生成校准数据集，以对多目标检测模型进行校准训练，提升了多目标检测模型检测的准确性。
以下通过一个具体的实施例对本申请的基于智能决策的目标检测模型的处理方法进行介绍。具体地，终端中安装有垃圾分类的应用，用户可以通过终端拍摄垃圾图片以扩充本地数据集。终端从中央服务器获取全局模型参数，得到初始化多目标检测模型。根据本地数据集对初始化多目标检测模型训练N(N为大于零的整数)个回合，得到中继多目标检测模型。
终端通过DH密钥交换协议与中央服务器确定通信所需的密钥。终端生成附加随机数，对附加随机数以及中继多目标检测模型的模型参数相加得到复合模型参数。终端使用加密密钥对复合模型参数进行加密得到加密模型参数，将加密模型参数发送到中央服务器。
中央服务器对加密模型参数解密，得到各节点的复合模型参数，将复合模型参数相加从而消除附加随机数的影响。中央服务器可以根据FedAvg的安全聚合算法计算全局模型参数，将全局模型参数下发至各节点，以使各节点根据全局模型参数更新中继多目标检测模型。
终端根据本地数据集对更新后的中继多目标检测模型进行迭代训练，直至模型收敛，得到多目标检测模型。
在应用时，用户可以通过终端对垃圾物品进行拍照，得到待检测图像，多目标检测图像可以一次性识别待检测图像中的多个垃圾物品，并展示识别到的垃圾物品。展示垃圾物品时，垃圾物品可以用检测框标出，并显示垃圾物品的类别，例如显示该垃圾物品是可回收物或有害垃圾，不同种类的垃圾物品可以使用不同颜色的检测框加以区分。
当用户认为目标检测准确，即垃圾分类正确时，可以点击检测到的垃圾物品，终端对该类别垃圾物品进行介绍，并显示垃圾分类说明，以便用户更好地进行垃圾分类。
当用户认为目标检测不准确，即垃圾分类有误时，可以点击虚拟校准按钮，在校准信息输入页面上传用户认为检测不准确的图像作为待校准图像，并输入说明文字作为校准指示信息，终端根据待校准图像和校准指示信息对多目标检测模型进行校准训练。
本申请中基于智能决策的目标检测模型处理方法涉及人工智能领域中的神经网络、机器学习、预测分析和计算机视觉；此外，还可以涉及智慧城市领域中的智慧环保。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程，是可以通过计算机可读指令来指令相关的硬件来完成，该计算机可读指令可存储于一计算机可读取存储介质中，该计算机可读指令在执行时，可包括如上述各方法的实施例的流程。其中，前述的存储介质可为磁碟、光盘、只读存储记忆体(Read-Only Memory，ROM)等非易失性存储介质，或随机存储记忆体(Random Access Memory，RAM)等。
应该理解的是，虽然附图的流程图中的各个步骤按照箭头的指示依次显示，但是这些步骤并不是必然按照箭头指示的顺序依次执行。除非本文中有明确的说明，这些步骤的执行并没有严格的顺序限制，其可以以其他的顺序执行。而且，附图的流程图中的至少一部分步骤可以包括多个子步骤或者多个阶段，这些子步骤或者阶段并不必然是在同一时刻执行完成，而是可以在不同的时刻执行，其执行顺序也不必然是依次进行，而是可以与其他步骤或者其他步骤的子步骤或者阶段的至少一部分轮流或者交替地执行。
进一步参考图3，作为对上述图2所示方法的实现，本申请提供了一种基于智能决策的目标检测模型处理装置的一个实施例，其中，目标检测模型为多目标检测模型，该装置实施例与图2所示的方法实施例相对应，该装置具体可以应用于各种电子设备中。
如图3所示，本实施例所述的基于智能决策的目标检测模型的处理装置300包括：获取模块301、模型训练模块302、参数计算模块303、参数发送模块304、模型更新模块305以及迭代训练模块306，其中：
获取模块301，用于获取本地数据集以及初始多目标检测模型。
模型训练模块302，用于根据本地数据集对初始多目标检测模型进行训练，得到中继多目标检测模型。
参数计算模块303，用于生成附加随机数，并根据附加随机数以及中继多目标检测模型的模型参数计算复合模型参数。
参数发送模块304，用于将复合模型参数发送至中央服务器，以指示中央服务器根据各节点的复合模型参数生成全局模型参数。
模型更新模块305，用于从中央服务器接收全局模型参数，以更新中继多目标检测模型。
迭代训练模块306，用于将更新后的中继多目标检测模型作为下轮训练的初始多目标检测模型进行迭代训练，直至模型收敛，得到多目标检测模型。
本实施例中，先根据本地数据集对初始多目标检测模型进行训练得到中继多目标检测模型；生成附加随机数，附加随机数用于对中继多目标检测模型的模型参数进行运算从而对模型参数进行加密；加密后得到的符合模型参数被发送至联盟网络中的中央服务器，中央服务器根据各节点的复合模型参数生成全局模型参数，全局模型参数被下发至终端后以更新中继多目标检测模型并进行迭代训练；全局模型参数是在各终端的本地数据集的基础上得到的，融合了多个本地数据集的特征，在保护本地数据集的基础上，实现了以更大规模的数据集对多目标检测模型进行训练，提高了训练完毕后得到的多目标检测模型检测的准确率。
在本实施例的一些可选的实现方式中，基于智能决策的目标检测模型的处理装置300还包括：参数获取模块以及模型构建模块，其中：
参数获取模块，用于从中央服务器获取全局模型参数。
模型构建模块，用于根据全局模型参数构建初始多目标检测模型。
本实施例中，终端根据中央服务器下发的全局模型参数构建初始多目标检测模型，从而实现模型初始化。
在本实施例的一些可选的实现方式中，模型训练模块302包括：图像输入子模块、误差确定子模块、参数调整子模块、迭代训练子模块，其中：
图像输入子模块，用于将本地数据集中的目标图像输入初始多目标检测模型，得到目标预测结果。
误差确定子模块，用于根据目标预测结果以及本地数据集中的图像标签确定预测误差。
参数调整子模块，用于基于预测误差对初始多目标检测模型进行参数调整。
迭代训练子模块，用于将参数调整后的初始多目标检测模型作为下轮训练的初始多目标检测模型进行迭代训练，直至迭代次数达到预设数值，得到中继多目标检测模型。
本实施例中，根据本地数据集中的目标图像以及图像标签对初始多目标检测模型进行预设次数的迭代训练，得到中继多目标检测模型，中继多目标检测模型用于联邦学习，保证了联邦学习的实现。
在本实施例的一些可选的实现方式中，参数计算模块303包括：附加生成子模块以及参数运算子模块，其中：
附加生成子模块，用于生成附加随机数；其中，联盟网络中各节点生成的附加随机数相加后值为零。
参数运算子模块，用于将生成的附加随机数与中继多目标检测模型的模型参数进行线性运算，得到复合模型参数。
本实施例中，生成附加随机数后，将附加随机数与中继多目标检测模型的模型参数进行线性运算，得到复合模型参数，从而实现对模型参数的数据加密。
在本实施例的一些可选的实现方式中，参数发送模块304包括：密钥确定子模块、参数加密子模块以及参数发送子模块，其中：
密钥确定子模块，用于与中央服务器进行通信以确定加密密钥。
参数加密子模块，用于根据加密密钥对复合模型参数进行加密，得到加密模型参数。
参数发送子模块，用于将加密模型参数发送至中央服务器，以指示中央服务器对各节点的加密模型参数进行解密，并根据解密后得到的复合模型参数进行运算，生成全局模型参数。
本实施例中，根据加密密钥对复合模型参数进行加密得到加密模型参数，以进一步保护联邦学习中的数据隐私；加密模型参数发送至中央服务器后，中央服务器对加密模型参数进行解密，并根据解密后得到的复合模型参数生成全局模型参数，全局模型参数用于更新各节点中的中继多目标检测模型，保证了联邦学习的实现。
在本实施例的一些可选的实现方式中，基于智能决策的目标检测模型的处理装置300还包括：图像获取模块、图像输入模块以及目标展示模块，其中：
图像获取模块，用于获取待检测图像。
图像输入模块，用于将待检测图像输入多目标检测模型，得到待检测图像中的目标对象。
目标展示模块，用于将检测到的目标对象进行展示。
本实施例中，对待检测图像进行目标检测的多目标检测模型通过联邦学习得到，并在联邦学习中使用了丰富的数据集进行训练，提高了目标检测的准确性。
在本实施例的一些可选的实现方式中，基于智能决策的目标检测模型的处理装置300还包括：页面展示模块、校准获取模块、校准生成模块以及校准训练模块，其中：
页面展示模块，用于当接收到触发的校准指令时，展示校准信息输入页面。
校准获取模块，用于获取在校准信息输入页面中输入的待校准图像以及校准指示信 息。
校准生成模块，用于根据待校准图像以及校准指示信息生成校准数据集。
校准训练模块，用于通过校准数据集对多目标检测模型进行校准训练。
本实施例中，当接收到校准指令时，展示校准信息输入页面，根据校准信息输入页面中输入的待校准图像以及校准指示信息生成校准数据集，以对多目标检测模型进行校准训练，提升了多目标检测模型检测的准确性。
为解决上述技术问题，本申请实施例还提供计算机设备。具体请参阅图4，图4为本实施例计算机设备基本结构框图。
所述计算机设备4包括通过系统总线相互通信连接存储器41、处理器42、网络接口43。需要指出的是，图中仅示出了具有组件41-43的计算机设备4，但是应理解的是，并不要求实施所有示出的组件，可以替代的实施更多或者更少的组件。其中，本技术领域技术人员可以理解，这里的计算机设备是一种能够按照事先设定或存储的指令，自动进行数值计算和/或信息处理的设备，其硬件包括但不限于微处理器、专用集成电路(Application Specific Integrated Circuit，ASIC)、可编程门阵列(Field－Programmable Gate Array，FPGA)、数字处理器(Digital Signal Processor，DSP)、嵌入式设备等。
所述计算机设备可以是桌上型计算机、笔记本、掌上电脑及云端服务器等计算设备。所述计算机设备可以与用户通过键盘、鼠标、遥控器、触摸板或声控设备等方式进行人机交互。
所述存储器41至少包括一种类型的计算机可读存储介质，所述计算机可读存储介质可以是非易失性，也可以是易失性，所述计算机可读存储介质包括闪存、硬盘、多媒体卡、卡型存储器(例如，SD或DX存储器等)、随机访问存储器(RAM)、静态随机访问存储器(SRAM)、只读存储器(ROM)、电可擦除可编程只读存储器(EEPROM)、可编程只读存储器(PROM)、磁性存储器、磁盘、光盘等。在一些实施例中，所述存储器41可以是所述计算机设备4的内部存储单元，例如该计算机设备4的硬盘或内存。在另一些实施例中，所述存储器41也可以是所述计算机设备4的外部存储设备，例如该计算机设备4上配备的插接式硬盘，智能存储卡(Smart Media Card,SMC)，安全数字(Secure Digital,SD)卡，闪存卡(Flash Card)等。当然，所述存储器41还可以既包括所述计算机设备4的内部存储单元也包括其外部存储设备。本实施例中，所述存储器41通常用于存储安装于所述计算机设备4的操作系统和各类应用软件，例如基于智能决策的目标检测模型处理方法的计算机可读指令等。此外，所述存储器41还可以用于暂时地存储已经输出或者将要输出的各类数据。
所述处理器42在一些实施例中可以是中央处理器(Central Processing Unit，CPU)、控制器、微控制器、微处理器、或其他数据处理芯片。该处理器42通常用于控制所述计算机设备4的总体操作。本实施例中，所述处理器42用于运行所述存储器41中存储的计算机可读指令或者处理数据，例如运行所述基于智能决策的目标检测模型处理方法的计算机可读指令。
所述网络接口43可包括无线网络接口或有线网络接口，该网络接口43通常用于在所述计算机设备4与其他电子设备之间建立通信连接。
本实施例中提供的计算机设备可以执行上述基于智能决策的目标检测模型处理方法。此处基于智能决策的目标检测模型处理方法可以是上述各个实施例的基于智能决策的目标检测模型处理方法。
本实施例中，先根据本地数据集对初始多目标检测模型进行训练得到中继多目标检测模型；生成附加随机数，附加随机数用于对中继多目标检测模型的模型参数进行运算从而对模型参数进行加密；加密后得到的符合模型参数被发送至联盟网络中的中央服务器，中央服务器根据各节点的复合模型参数生成全局模型参数，全局模型参数被下发至终端后以更新中继多目标检测模型并进行迭代训练；全局模型参数是在各终端的本地数据集的基础 上得到的，融合了多个本地数据集的特征，在保护本地数据集的基础上，实现了以更大规模的数据集对多目标检测模型进行训练，提高了训练完毕后得到的多目标检测模型检测的准确率。
本申请还提供了另一种实施方式，即提供一种计算机可读存储介质，所述计算机可读存储介质存储有计算机可读指令，所述计算机可读指令可被至少一个处理器执行，以使所述至少一个处理器执行如上述的基于智能决策的目标检测模型处理方法的步骤。
本实施例中，先根据本地数据集对初始多目标检测模型进行训练得到中继多目标检测模型；生成附加随机数，附加随机数用于对中继多目标检测模型的模型参数进行运算从而对模型参数进行加密；加密后得到的符合模型参数被发送至联盟网络中的中央服务器，中央服务器根据各节点的复合模型参数生成全局模型参数，全局模型参数被下发至终端后以更新中继多目标检测模型并进行迭代训练；全局模型参数是在各终端的本地数据集的基础上得到的，融合了多个本地数据集的特征，在保护本地数据集的基础上，实现了以更大规模的数据集对多目标检测模型进行训练，提高了训练完毕后得到的多目标检测模型检测的准确率。
通过以上的实施方式的描述，本领域的技术人员可以清楚地了解到上述实施例方法可借助软件加必需的通用硬件平台的方式来实现，当然也可以通过硬件，但很多情况下前者是更佳的实施方式。基于这样的理解，本申请的技术方案本质上或者说对现有技术做出贡献的部分可以以软件产品的形式体现出来，该计算机软件产品存储在一个存储介质(如ROM/RAM、磁碟、光盘)中，包括若干指令用以使得一台终端(可以是手机，计算机，服务器，空调器，或者网络设备等)执行本申请各个实施例所述的方法。
显然，以上所描述的实施例仅仅是本申请一部分实施例，而不是全部的实施例，附图中给出了本申请的较佳实施例，但并不限制本申请的专利范围。本申请可以以许多不同的形式来实现，相反地，提供这些实施例的目的是使对本申请的公开内容的理解更加透彻全面。尽管参照前述实施例对本申请进行了详细的说明，对于本领域的技术人员来而言，其依然可以对前述各具体实施方式所记载的技术方案进行修改，或者对其中部分技术特征进行等效替换。凡是利用本申请说明书及附图内容所做的等效结构，直接或间接运用在其他相关的技术领域，均同理在本申请专利保护范围之内。

Claims (20)

  1. 一种基于智能决策的目标检测模型的处理方法，其中，所述目标检测模型为多目标检测模型，所述方法包括：
    获取本地数据集以及初始多目标检测模型；
    根据所述本地数据集对所述初始多目标检测模型进行训练，得到中继多目标检测模型；
    生成附加随机数，并根据所述附加随机数以及所述中继多目标检测模型的模型参数计算复合模型参数；
    将所述复合模型参数发送至中央服务器，以指示所述中央服务器根据各节点的复合模型参数生成全局模型参数；
    从所述中央服务器接收所述全局模型参数，以更新所述中继多目标检测模型；
    将更新后的中继多目标检测模型作为下轮训练的初始多目标检测模型进行迭代训练，直至模型收敛，得到所述多目标检测模型。
  2. 根据权利要求1所述的基于智能决策的目标检测模型的处理方法，其中，所述获取本地数据集以及初始多目标检测模型的步骤之前，所述方法还包括：
    从中央服务器获取全局模型参数；
    根据所述全局模型参数构建初始多目标检测模型。
  3. 根据权利要求1所述的基于智能决策的目标检测模型的处理方法，其中，所述根据所述本地数据集对所述初始多目标检测模型进行训练，得到中继多目标检测模型的步骤包括：
    将所述本地数据集中的目标图像输入所述初始多目标检测模型，得到目标预测结果；
    根据所述目标预测结果以及所述本地数据集中的图像标签确定预测误差；
    基于所述预测误差对所述初始多目标检测模型进行参数调整；
    将参数调整后的初始多目标检测模型作为下轮训练的初始多目标检测模型进行迭代训练，直至迭代次数达到预设数值，得到中继多目标检测模型。
  4. 根据权利要求1所述的基于智能决策的目标检测模型的处理方法，其中，所述生成附加随机数，并根据所述附加随机数以及所述中继多目标检测模型的模型参数计算复合模型参数的步骤包括：
    生成附加随机数；其中，联盟网络中各节点生成的附加随机数相加后值为零；
    将生成的附加随机数与所述中继多目标检测模型的模型参数进行线性运算，得到复合模型参数。
  5. 根据权利要求1所述的基于智能决策的目标检测模型的处理方法，其中，所述将所述复合模型参数发送至中央服务器，以指示所述中央服务器根据各节点的复合模型参数生成全局模型参数的步骤包括：
    与中央服务器进行通信以确定加密密钥；
    根据所述加密密钥对所述复合模型参数进行加密，得到加密模型参数；
    将所述加密模型参数发送至所述中央服务器，以指示所述中央服务器对所述各节点的加密模型参数进行解密，并根据解密后得到的复合模型参数进行运算，生成全局模型参数。
  6. 根据权利要求1所述的基于智能决策的目标检测模型的处理方法，其中，在所述将更新后的中继多目标检测模型作为下轮训练的初始多目标检测模型进行迭代训练，直至模型收敛，得到多目标检测模型的步骤之后，所述方法还包括：
    获取待检测图像；
    将所述待检测图像输入所述多目标检测模型，得到所述待检测图像中的目标对象；
    将检测到的目标对象进行展示。
  7. 根据权利要求6所述的基于智能决策的目标检测模型的处理方法，其中，所述将检测到的目标对象进行展示的步骤之后，所述方法还包括：
    当接收到触发的校准指令时，展示校准信息输入页面；
    获取在所述校准信息输入页面中输入的待校准图像以及校准指示信息；
    根据所述待校准图像以及所述校准指示信息生成校准数据集；
    通过所述校准数据集对所述多目标检测模型进行校准训练。
  8. 一种基于智能决策的目标检测模型的处理装置，其中，所述目标检测模型为多目标检测模型，所述装置包括：
    获取模块，用于获取本地数据集以及初始多目标检测模型；
    模型训练模块，用于根据所述本地数据集对所述初始多目标检测模型进行训练，得到中继多目标检测模型；
    参数计算模块，用于生成附加随机数，并根据所述附加随机数以及所述中继多目标检测模型的模型参数计算复合模型参数；
    参数发送模块，用于将所述复合模型参数发送至中央服务器，以指示所述中央服务器根据各节点的复合模型参数生成全局模型参数；
    模型更新模块，用于从所述中央服务器接收所述全局模型参数，以更新所述中继多目标检测模型；
    迭代训练模块，用于将更新后的中继多目标检测模型作为下轮训练的初始多目标检测模型进行迭代训练，直至模型收敛，得到所述多目标检测模型。
  9. 一种计算机设备，包括存储器和处理器，所述存储器中存储有计算机可读指令，所述处理器执行所述计算机可读指令时实现如下步骤：
    获取本地数据集以及初始多目标检测模型；
    根据所述本地数据集对所述初始多目标检测模型进行训练，得到中继多目标检测模型；
    生成附加随机数，并根据所述附加随机数以及所述中继多目标检测模型的模型参数计算复合模型参数；
    将所述复合模型参数发送至中央服务器，以指示所述中央服务器根据各节点的复合模型参数生成全局模型参数；
    从所述中央服务器接收所述全局模型参数，以更新所述中继多目标检测模型；
    将更新后的中继多目标检测模型作为下轮训练的初始多目标检测模型进行迭代训练，直至模型收敛，得到所述多目标检测模型。
  10. 根据权利要求9所述的计算机设备，其中，所述根据所述本地数据集对所述初始多目标检测模型进行训练，得到中继多目标检测模型的步骤包括：
    将所述本地数据集中的目标图像输入所述初始多目标检测模型，得到目标预测结果；
    根据所述目标预测结果以及所述本地数据集中的图像标签确定预测误差；
    基于所述预测误差对所述初始多目标检测模型进行参数调整；
    将参数调整后的初始多目标检测模型作为下轮训练的初始多目标检测模型进行迭代训练，直至迭代次数达到预设数值，得到中继多目标检测模型。
  11. 根据权利要求9所述的计算机设备，其中，所述生成附加随机数，并根据所述附加随机数以及所述中继多目标检测模型的模型参数计算复合模型参数的步骤包括：
    生成附加随机数；其中，联盟网络中各节点生成的附加随机数相加后值为零；
    将生成的附加随机数与所述中继多目标检测模型的模型参数进行线性运算，得到复合模型参数。
  12. 根据权利要求9所述的计算机设备，其中，所述将所述复合模型参数发送至中央服务器，以指示所述中央服务器根据各节点的复合模型参数生成全局模型参数的步骤包括：
    与中央服务器进行通信以确定加密密钥；
    根据所述加密密钥对所述复合模型参数进行加密，得到加密模型参数；
    将所述加密模型参数发送至所述中央服务器，以指示所述中央服务器对所述各节点的 加密模型参数进行解密，并根据解密后得到的复合模型参数进行运算，生成全局模型参数。
  13. 根据权利要求9所述的计算机设备，其中，在所述将更新后的中继多目标检测模型作为下轮训练的初始多目标检测模型进行迭代训练，直至模型收敛，得到多目标检测模型的步骤之后，所述方法还包括：
    获取待检测图像；
    将所述待检测图像输入所述多目标检测模型，得到所述待检测图像中的目标对象；
    将检测到的目标对象进行展示。
  14. 根据权利要求13所述的计算机设备，其中，所述将检测到的目标对象进行展示的步骤之后，所述方法还包括：
    当接收到触发的校准指令时，展示校准信息输入页面；
    获取在所述校准信息输入页面中输入的待校准图像以及校准指示信息；
    根据所述待校准图像以及所述校准指示信息生成校准数据集；
    通过所述校准数据集对所述多目标检测模型进行校准训练。
  15. 一种计算机可读存储介质，所述计算机可读存储介质上存储有计算机可读指令；其中，所述计算机可读指令被处理器执行时实现如下步骤：
    获取本地数据集以及初始多目标检测模型；
    根据所述本地数据集对所述初始多目标检测模型进行训练，得到中继多目标检测模型；
    生成附加随机数，并根据所述附加随机数以及所述中继多目标检测模型的模型参数计算复合模型参数；
    将所述复合模型参数发送至中央服务器，以指示所述中央服务器根据各节点的复合模型参数生成全局模型参数；
    从所述中央服务器接收所述全局模型参数，以更新所述中继多目标检测模型；
    将更新后的中继多目标检测模型作为下轮训练的初始多目标检测模型进行迭代训练，直至模型收敛，得到所述多目标检测模型。
  16. 根据权利要求15所述的一种计算机可读存储介质，其中，所述根据所述本地数据集对所述初始多目标检测模型进行训练，得到中继多目标检测模型的步骤包括：
    将所述本地数据集中的目标图像输入所述初始多目标检测模型，得到目标预测结果；
    根据所述目标预测结果以及所述本地数据集中的图像标签确定预测误差；
    基于所述预测误差对所述初始多目标检测模型进行参数调整；
    将参数调整后的初始多目标检测模型作为下轮训练的初始多目标检测模型进行迭代训练，直至迭代次数达到预设数值，得到中继多目标检测模型。
  17. 根据权利要求15所述的一种计算机可读存储介质，其中，所述生成附加随机数，并根据所述附加随机数以及所述中继多目标检测模型的模型参数计算复合模型参数的步骤包括：
    生成附加随机数；其中，联盟网络中各节点生成的附加随机数相加后值为零；
    将生成的附加随机数与所述中继多目标检测模型的模型参数进行线性运算，得到复合模型参数。
  18. 根据权利要求15所述的一种计算机可读存储介质，其中，所述将所述复合模型参数发送至中央服务器，以指示所述中央服务器根据各节点的复合模型参数生成全局模型参数的步骤包括：
    与中央服务器进行通信以确定加密密钥；
    根据所述加密密钥对所述复合模型参数进行加密，得到加密模型参数；
    将所述加密模型参数发送至所述中央服务器，以指示所述中央服务器对所述各节点的加密模型参数进行解密，并根据解密后得到的复合模型参数进行运算，生成全局模型参数。
  19. 根据权利要求15所述的一种计算机可读存储介质，其中，在所述将更新后的中继 多目标检测模型作为下轮训练的初始多目标检测模型进行迭代训练，直至模型收敛，得到多目标检测模型的步骤之后，所述方法还包括：
    获取待检测图像；
    将所述待检测图像输入所述多目标检测模型，得到所述待检测图像中的目标对象；
    将检测到的目标对象进行展示。
  20. 根据权利要求19所述的一种计算机可读存储介质，其中，所述将检测到的目标对象进行展示的步骤之后，所述方法还包括：
    当接收到触发的校准指令时，展示校准信息输入页面；
    获取在所述校准信息输入页面中输入的待校准图像以及校准指示信息；
    根据所述待校准图像以及所述校准指示信息生成校准数据集；
    通过所述校准数据集对所述多目标检测模型进行校准训练。
PCT/CN2020/124439 2020-09-18 2020-10-28 基于智能决策的目标检测模型的处理方法、及其相关设备 WO2021174877A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010990029.4A CN112116008A (zh) 2020-09-18 2020-09-18 基于智能决策的目标检测模型的处理方法、及其相关设备
CN202010990029.4 2020-09-18

Publications (1)

Publication Number Publication Date
WO2021174877A1 true WO2021174877A1 (zh) 2021-09-10

Family

ID=73800792

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/124439 WO2021174877A1 (zh) 2020-09-18 2020-10-28 基于智能决策的目标检测模型的处理方法、及其相关设备

Country Status (2)

Country Link
CN (1) CN112116008A (zh)
WO (1) WO2021174877A1 (zh)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115085921A (zh) * 2022-07-22 2022-09-20 北京智芯微电子科技有限公司 模型训练方法、节点检测方法、装置、设备及介质
CN115225344A (zh) * 2022-06-28 2022-10-21 广东润联信息技术有限公司 网络信息安全处理方法、装置、计算机设备及存储介质
WO2023098511A1 (zh) * 2021-12-03 2023-06-08 华为技术有限公司 通信方法及装置、存储介质、程序产品
CN117395083A (zh) * 2023-12-11 2024-01-12 东信和平科技股份有限公司 基于联邦学习的数据保护方法及系统
WO2024114640A1 (zh) * 2022-11-30 2024-06-06 北京航天数据股份有限公司 一种基于用户画像的用户服务系统、方法及电子设备

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114358316A (zh) * 2022-01-14 2022-04-15 中国人民解放军总医院 联邦学习系统及其大规模影像训练方法和装置
CN117671385A (zh) * 2023-12-20 2024-03-08 北京斯年智驾科技有限公司 一种目标识别模型的训练方法、系统、装置及存储介质

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190005657A1 (en) * 2017-06-30 2019-01-03 Baidu Online Network Technology (Beijing) Co., Ltd . Multiple targets-tracking method and apparatus, device and storage medium
CN110572253A (zh) * 2019-09-16 2019-12-13 济南大学 一种联邦学习训练数据隐私性增强方法及系统
CN111126303A (zh) * 2019-12-25 2020-05-08 北京工业大学 一种面向智能停车的多车位检测方法
CN111462174A (zh) * 2020-03-06 2020-07-28 北京百度网讯科技有限公司 多目标跟踪方法、装置以及电子设备

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP5860556B1 (ja) * 2015-02-06 2016-02-16 日本電信電話株式会社 不整合検知方法、不整合検知システム、不整合検知装置、およびプログラム
CN110704860A (zh) * 2019-11-18 2020-01-17 深圳前海微众银行股份有限公司 提升安全性的纵向联邦学习方法、设备、系统及存储介质
CN111611610B (zh) * 2020-04-12 2023-05-30 西安电子科技大学 联邦学习信息处理方法、系统、存储介质、程序、终端

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190005657A1 (en) * 2017-06-30 2019-01-03 Baidu Online Network Technology (Beijing) Co., Ltd . Multiple targets-tracking method and apparatus, device and storage medium
CN110572253A (zh) * 2019-09-16 2019-12-13 济南大学 一种联邦学习训练数据隐私性增强方法及系统
CN111126303A (zh) * 2019-12-25 2020-05-08 北京工业大学 一种面向智能停车的多车位检测方法
CN111462174A (zh) * 2020-03-06 2020-07-28 北京百度网讯科技有限公司 多目标跟踪方法、装置以及电子设备

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2023098511A1 (zh) * 2021-12-03 2023-06-08 华为技术有限公司 通信方法及装置、存储介质、程序产品
CN115225344A (zh) * 2022-06-28 2022-10-21 广东润联信息技术有限公司 网络信息安全处理方法、装置、计算机设备及存储介质
CN115225344B (zh) * 2022-06-28 2024-05-14 华润智算科技(广东)有限公司 网络信息安全处理方法、装置、计算机设备及存储介质
CN115085921A (zh) * 2022-07-22 2022-09-20 北京智芯微电子科技有限公司 模型训练方法、节点检测方法、装置、设备及介质
CN115085921B (zh) * 2022-07-22 2023-01-20 北京智芯微电子科技有限公司 模型训练方法、节点检测方法、装置、设备及介质
WO2024114640A1 (zh) * 2022-11-30 2024-06-06 北京航天数据股份有限公司 一种基于用户画像的用户服务系统、方法及电子设备
CN117395083A (zh) * 2023-12-11 2024-01-12 东信和平科技股份有限公司 基于联邦学习的数据保护方法及系统
CN117395083B (zh) * 2023-12-11 2024-03-19 东信和平科技股份有限公司 基于联邦学习的数据保护方法及系统

Also Published As

Publication number Publication date
CN112116008A (zh) 2020-12-22

Similar Documents

Publication Publication Date Title
WO2021174877A1 (zh) 基于智能决策的目标检测模型的处理方法、及其相关设备
WO2021120676A1 (zh) 联邦学习网络下的模型训练方法及其相关设备
WO2021120677A1 (zh) 一种仓储模型训练方法、装置、计算机设备及存储介质
WO2021155713A1 (zh) 基于权重嫁接的模型融合的人脸识别方法及相关设备
US10664613B2 (en) Virtualization input component
CN108776787A (zh) 图像处理方法及装置、电子设备、存储介质
WO2021143267A1 (zh) 基于图像检测的细粒度分类模型处理方法、及其相关设备
WO2021151336A1 (zh) 基于注意力机制的道路图像目标检测方法及相关设备
US10311288B1 (en) Determining identity of a person in a digital image
US20150199541A1 (en) Method and system for secured communication of personal information
CN111400504B (zh) 企业关键人的识别方法和装置
CN113159327A (zh) 基于联邦学习系统的模型训练方法、装置、电子设备
CN112529101B (zh) 分类模型的训练方法、装置、电子设备及存储介质
WO2023071105A1 (zh) 一种特征变量的分析方法、装置、计算机设备及存储介质
CN111325352A (zh) 基于纵向联邦学习的模型更新方法、装置、设备及介质
US10742627B2 (en) System and method for dynamic network data validation
US20210065353A1 (en) Automated collectible card grading system
CN114186256A (zh) 神经网络模型的训练方法、装置、设备和存储介质
CN111651749A (zh) 基于密码找回账号的方法、装置、计算机设备及存储介质
WO2022001233A1 (zh) 基于层次化迁移学习的预标注方法及其相关设备
CN107818323A (zh) 用于处理图像的方法和装置
US20190324606A1 (en) Online training of segmentation model via interactions with interactive computing environment
CN112418442A (zh) 联邦迁移学习的数据处理方法、装置、设备及存储介质
CN112381236A (zh) 联邦迁移学习的数据处理方法、装置、设备及存储介质
CN111641517A (zh) 同质网络的社群划分方法、装置、计算机设备和存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20923054

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20923054

Country of ref document: EP

Kind code of ref document: A1