WO2021155713A1 - Procédé de reconnaissance faciale à base de fusion de modèle de greffage de poids, et dispositif y relatif - Google Patents

Procédé de reconnaissance faciale à base de fusion de modèle de greffage de poids, et dispositif y relatif Download PDF

Info

Publication number
WO2021155713A1
WO2021155713A1 PCT/CN2020/135621 CN2020135621W WO2021155713A1 WO 2021155713 A1 WO2021155713 A1 WO 2021155713A1 CN 2020135621 W CN2020135621 W CN 2020135621W WO 2021155713 A1 WO2021155713 A1 WO 2021155713A1
Authority
WO
WIPO (PCT)
Prior art keywords
model
fusion
face recognition
weight
training
Prior art date
Application number
PCT/CN2020/135621
Other languages
English (en)
Chinese (zh)
Inventor
胡魁
戴磊
张国辉
宋晨
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 WO2021155713A1 publication Critical patent/WO2021155713A1/fr

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G06V40/168—Feature extraction; Face representation
    • 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
    • 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
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G06V40/172—Classification, e.g. identification
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Definitions

  • This application relates to the field of artificial intelligence technology, and in particular to a face recognition method and related equipment based on model fusion based on weight grafting.
  • the method to improve the accuracy of face recognition is that different models of face recognition have different recognition features, and different models are fused to obtain a fusion model of face recognition.
  • the inventor realizes that the existing model fusion method is to combine, arrange, or vote the outputs of multiple models to obtain the final output. This will improve the accuracy of the model as a whole, but it also introduces the need for multiple independent fronts.
  • the problem of orientation affects the execution efficiency of the model.
  • the Deep Convolution Neural Network contains weights that have little or no effect on the output, and these weights have minimal positive feedback in the entire model.
  • the existing method is to cut out these invalid weights through some pruning strategies to remove the influence of the weights on the final result.
  • the pruning method will delete some weights, feature layers, etc. The advantage of this is that after the weights are deleted, the number of parameters of the model becomes less and the execution speed becomes faster, but the accuracy of the model will be reduced to a certain extent.
  • the purpose of the embodiments of the present application is to propose a face recognition method and related equipment based on model fusion based on weight grafting, so as to solve the problems of low execution efficiency and decreased recognition accuracy of the fusion model generated in related technologies.
  • an embodiment of the present application provides a face recognition method based on model fusion based on weight grafting, and adopts the following technical solutions:
  • N is a positive integer
  • an embodiment of the present application also provides a face recognition device based on model fusion of weight grafting, which adopts the following technical solutions:
  • the training module is used to train N face recognition models with the same structure, where N is a positive integer;
  • the fine-tuning module is used to select one of the face recognition models as the model to be fused and make fine-tuning;
  • the training module is also used to train all face recognition models for a training cycle after fine-tuning
  • the calculation module is used to calculate the fusion coefficient of each parameter in each model after training.
  • the update module is configured to determine the fusion weight of each parameter according to the fusion coefficient, and update the parameters of the model to be fused based on the fusion weight of all parameters to obtain a fusion model;
  • the recognition module is used to perform face recognition on the received face image through the fusion model.
  • the embodiments of the present application also provide a computer device, which adopts the following technical solutions:
  • the computer device includes a memory and a processor, and computer-readable instructions are stored in the memory.
  • the processor executes the computer-readable instructions, the steps of the face recognition method based on weight grafting and model fusion as described below are implemented :
  • N is a positive integer
  • the embodiments of the present application also provide a computer-readable storage medium, which adopts the following technical solutions:
  • the computer-readable storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by a processor, the steps of the face recognition method based on weight grafting and model fusion as described below are realized:
  • N is a positive integer
  • This application trains N face recognition models of the same structure separately, selects one of the face recognition models as the model to be fused, and fine-tunes them. After fine-tuning, all face recognition models are trained for one training cycle, and each face recognition model is calculated after training.
  • the fusion coefficient of each parameter in the model, the fusion weight of each parameter is determined according to the fusion coefficient, the parameters of the model to be fused based on the fusion weight of all parameters are updated to obtain the fusion model, and finally the received face image is analyzed through the fusion model Perform face recognition; this application uses weight grafting to fuse multiple face recognition models, and weights and merges the weights corresponding to each parameter in each model to obtain the fusion weight of the fusion model, so that the information content is higher
  • the weight has a greater impact on the fusion weight, and the weight with a small amount of information has less impact on the fusion weight, which improves the recognition accuracy of the fusion model while improving the execution efficiency of the model.
  • 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 face recognition method based on model fusion based on weight grafting according to the present application;
  • FIG. 3 is a flowchart of a specific implementation of step S203 in FIG. 2;
  • FIG. 4 is a schematic structural diagram of an embodiment of a face recognition device based on model fusion based on weight grafting according to the present application;
  • Fig. 5 is a schematic structural diagram of an embodiment of a computer device according to the present application.
  • the system architecture 100 may include terminal devices 101, 102, 103, a network 104, and a server 105.
  • the network 104 is used to provide a medium for communication links between the terminal devices 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 terminal devices 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, and social platform software, can be installed on the terminal devices 101, 102, and 103.
  • the terminal devices 101, 102, 103 may be various electronic devices with display screens and support for web browsing, including but not limited to smart phones, tablet computers, e-book readers, MP3 players (Moving Picture Experts Group Audio Layer III, dynamic Video 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 Moving Picture Experts Group Audio Layer III, dynamic Video experts compress standard audio layer 3
  • 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 terminal devices 101, 102, and 103.
  • the face recognition method based on model fusion based on weight grafting provided by the embodiments of the present application is generally executed by a server/terminal device. Accordingly, the face recognition device based on model fusion based on weight grafting is generally set in the server/terminal. In the device.
  • terminal devices, networks, and servers in FIG. 1 are merely illustrative. There can be any number of terminal devices, networks, and servers according to implementation needs.
  • FIG. 2 there is shown a flowchart of an embodiment of a face recognition method based on model fusion based on weight grafting according to the present application.
  • the face recognition method based on model fusion based on weight grafting includes the following steps:
  • Step S201 Train N face recognition models with the same structure respectively.
  • model structure of the trained N face recognition models is the same, where N is a positive integer, and the number of parameters of each model is also the same, but the weight value distribution of each parameter in different models Not the same, the hyperparameters and initialization of each model are also different.
  • a face recognition model Before training, a face recognition model must be created. At the beginning of the model creation, the weight value of each parameter in different models is set, and at the same time, the hyperparameters and initialization method of each model are set.
  • the face recognition model includes: a feature extraction layer, a fully connected layer, and a loss function layer;
  • the feature extraction layer is used to extract the face features of the input face sample data
  • the fully connected layer includes a first feature output layer, a second feature output layer, and at least one attribute output layer.
  • the first feature output layer and the second feature output layer are used to output facial features, and the attribute output layer is used to output the corresponding face. Attribute classification results, the output of the feature extraction layer is connected to the input of the second feature output layer and the input of the attribute output layer after passing through the first feature output layer;
  • the loss function layer includes a feature extraction loss function layer and at least one attribute loss function layer.
  • the output of the second feature output layer is connected to the input of the feature extraction loss function layer, and the output of the attribute output layer is connected to the corresponding attribute loss function layer. Input phase connection.
  • the loss function of the loss function layer is used to measure the difference between the predicted value and the true value obtained by constructing the model, that is, to measure the quality of a prediction of the model.
  • the face features are synthesized and attributed, and the feature representation and classification results are output;
  • the feature extraction loss value and the attribute loss value of the loss function layer are calculated based on the loss function, and the feature extraction loss value and the attribute loss value are weighted and summed to obtain the target loss value, and the target loss value is used to determine the face recognition Whether the model training is over.
  • the steps of separately training N face recognition models with the same structure include:
  • the feature extraction layer is used to extract the face features of the input face sample data
  • the face features are synthesized and attributed, and the feature representation and classification results are output;
  • the target loss value it is judged whether the training of the face recognition model is finished.
  • the first feature output layer synthesizes the facial features extracted by the feature extraction layer to output the feature representation of the face sample data
  • the attribute output layer compares the facial features output by the first feature output layer to the face sample data.
  • the attributes are classified, and the classification results are output.
  • the loss functions of feature extraction and attribute analysis are trained at the same time.
  • the second feature output layer inputs the image features output by the first feature output layer to the feature extraction loss function layer, and each attribute output layer outputs the image features to the corresponding attributes.
  • Loss function layer Since there are multiple loss functions during training, and each loss function has a corresponding weight, multiple loss functions affect each other, which is beneficial to training better model parameters. It is judged whether the target loss value is less than the preset threshold value, and if it is, the model training is completed.
  • the face recognition model in this embodiment may be a convolutional neural network model
  • the feature extraction layer includes an input layer, a convolutional layer, and a pooling layer.
  • step S202 one of the face recognition models is selected as the model to be fused and fine-tuned. After fine-tuning, all face recognition models are trained for a training period.
  • N face recognition models with the same structure are trained, denoted by m 0 , m 1 , m 2 ,..., m N-1 , and one of the face recognition models is selected as the model to be fused,
  • the face recognition model m 0 can be selected as the model to be fused and finetuned (finetune). Specifically, the parameters of most convolutional layers close to the input of the model to be fused are frozen, and only the remaining convolutional layers and The parameters of the fully connected layer. After fine-tuning the fusion model, all the models are trained for a complete training cycle.
  • Step S203 Calculate the fusion coefficient of each parameter in each model after training.
  • the weight entropy of each parameter in each face recognition model is calculated, and then the fusion parameter of the corresponding parameter is calculated according to the weight entropy.
  • step S301 the weight value corresponding to each parameter in each model is divided into different value ranges according to the size
  • Step S302 calculating the probability of the weight value corresponding to each value range
  • Step S303 Calculate the weight entropy corresponding to each parameter in each model according to the probability.
  • step S301 the weight value of each parameter in different face recognition models is divided into n different value ranges according to the value of the
  • the model has ten parameters x 0 , x 1 , ......, x 9 , the weight value corresponding to each parameter in model 1 is 0.1, 0.05, 0.2, 0.15, 0.25, 0.075, 0.055, 0.02, 0.07, 0.03, model
  • the weight value corresponding to each parameter in 2 is 0.01, 0.36, 0.13, 0.035, 0.075, 0.09, 0.1, 0.08, 0.07, 0.05.
  • the above weight value is equally divided into 5 different value ranges, the first value range to the first value range
  • the five value ranges are 0.01 ⁇ 0.08, 0.08 ⁇ 0.15, 0.15 ⁇ 0.22, 0.22 ⁇ 0.29, 0.29 ⁇ 0.36 respectively.
  • step S302 the weight value probability corresponding to each value range is calculated, and the probability is the ratio of the number of parameters contained in each value range to the total number of parameters. Still taking the above example to illustrate, the two models have a total of 20 parameters, and the weight values of 12 parameters are within the range of the first value range. Therefore, the probability of the first value range is By analogy, the probability of the second range is The probability of the third range is The probability of the fourth range is The probability of the fifth range is
  • step S303 after calculating the weight value probability p k corresponding to each value range range, the weight entropy of each parameter is calculated according to the obtained probability, and the weight entropy is calculated using the following formula:
  • n is the number of value ranges
  • p k represents the probability of the weight value distributed in the k-th value range. It should be appreciated that the smaller the weight corresponding to H (w i), indicates a lower weight of the weight change, the less the corresponding information.
  • the fusion coefficient of each parameter in the different face recognition models is further determined.
  • the calculation formula of the fusion coefficient is as follows:
  • a and c are hyperparameters, Is the fusion coefficient of the parameter i in the face recognition model m j.
  • Step S204 Determine the fusion weight of each parameter according to the fusion coefficient, and update the model parameters of the fusion model to obtain the fusion model.
  • the N face recognition models are fused into the face recognition model m 0 , and the model parameters after the fusion are calculated. Specifically, the weight entropy of the same target parameter in all models is weighted and merged by the fusion coefficient, To obtain the fusion weight of the target parameter in the fusion model, use the following calculation formula:
  • I the weight corresponding to the parameter i in the fusion model.
  • multiple face recognition models are fused by weight grafting, and weights are combined according to the corresponding weight entropy of each parameter in different models, so that a higher weight entropy corresponds to a higher information weight pair fusion
  • the weight has a greater impact, and the weight with a low entropy corresponding to a small amount of information has less impact on the fusion weight, avoiding the practice of cutting off the weight with a small impact factor in related technologies, improving the recognition accuracy of the fusion model, and ensuring The robustness of the fusion model is improved.
  • the preset learning rate is further used to fine-tune the fusion model, and the weight of each layer of the face recognition model is adjusted.
  • Fine-tuning is to modify a part of the network structure to the model you need, using the convolutional neural network VGG16 as an example to illustrate.
  • the structure of the VGG16 network is a convolutional layer and a fully connected layer.
  • the convolutional layer is divided into 5 parts, a total of 13 layers, and 3 layers fc6, fc7, and fc8 are fully connected layers. If the structure of VGG16 is to be used in a new data set, the fc8 layer must be removed first. The reason is that the input of the fc8 layer is the feature of fc7, and the output is the probability of 1000 classes. These 1000 classes correspond to the 1000 classes in the ImageNet model.
  • the number of categories is generally not 1000, so the structure of the fc8 layer is not applicable at this time.
  • the fc8 layer must be removed, and the fully connected layer that meets the number of categories of the data set is used as the new fc8.
  • the data set is of 5 types, so the output of the new fc8 should also be of 5 types.
  • some parameters of the fusion model can be modified as needed.
  • learning rate controls the learning progress of the model, and a proper learning rate can make the model converge in a proper time.
  • a smaller learning rate is used to fine-tune the fusion model to obtain the required model, which solves the problem of multi-model fusion training due to data isolation.
  • step S204 it further includes determining whether the fusion model has converged, and if so, using the updated model parameters as the final parameters of the fusion model; if not, proceed to the next training cycle Training until the fusion model converges.
  • the way to determine whether the fusion model is in a convergent state can be to determine whether the number of model updates has reached the preset number of times. If it reaches the preset number of times, the fusion model is determined to be in a convergent state; it can also be judged whether the training duration is greater than the preset duration. If the duration is greater than the preset time, it is determined that the fusion model is in a convergent state. Among them, the preset times and the preset duration can be set as required. After the fusion model converges, the face image to be recognized is input into the fusion model for face image recognition.
  • the training of the next training cycle continues to repeat steps S202 to S204. Specifically, the fusion model is to be fine-tuned again. After fine-tuning, all face recognition models are trained for a training cycle, and the face recognition models in each face recognition model after the training are calculated. For the fusion coefficient of each parameter, the fusion weight of each parameter is determined according to the fusion coefficient, and the parameters of the model to be fused are updated based on the fusion weight of all parameters to obtain the fusion model.
  • the aforementioned facial image information to be recognized can 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.
  • This application trains N face recognition models of the same structure separately, selects one of the face recognition models as the model to be fused, and fine-tunes them. After fine-tuning, all face recognition models are trained for one training cycle, and each face recognition model is calculated after training.
  • the fusion coefficient of each parameter in the model, the fusion weight of each parameter is determined according to the fusion coefficient, the parameters of the model to be fused based on the fusion weight of all parameters are updated to obtain the fusion model, and finally the received face image is analyzed through the fusion model Perform face recognition; this application uses weight grafting to fuse multiple face recognition models, and weights and merges the weights corresponding to each parameter in each model to obtain the fusion weight of the fusion model, so that the information content is higher
  • the weight has a greater impact on the fusion weight, and the weight with a small amount of information has a small impact on the fusion weight, which improves the recognition accuracy of the fusion model while improving the execution efficiency of the model.
  • This application can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld devices or portable devices, tablet devices, multi-processor systems, microprocessor-based systems, set-top boxes, programmable consumer electronic devices, network PCs, small computers, large computers, including Distributed computing environment for any of the above systems or equipment, etc.
  • This application may be described in the general context of computer-executable instructions executed by a computer, such as a program module.
  • program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types.
  • This application can also be practiced in distributed computing environments. In these distributed computing environments, tasks are performed by remote processing devices connected through a communication network.
  • program modules can be located in local and remote computer storage media including storage devices.
  • This application can be applied to the field of identity recognition of smart security, so as to promote the construction of smart cities.
  • 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.
  • the present application provides an embodiment of a face recognition device based on model fusion based on weight grafting, and this device embodiment is implemented with the method shown in FIG.
  • the device can be specifically applied to various electronic devices.
  • the face recognition device based on model fusion based on weight grafting in this embodiment includes: a training module 401, a fine-tuning module 402, a calculation module 403, an update module 404, and a recognition module 405. in:
  • the training module 401 is used to train N face recognition models with the same structure respectively, where N is a positive integer;
  • the fine-tuning module 402 is used to select one of the face recognition models m 0 as the model to be fused and perform fine-tuning;
  • the training module 401 is also used to train all face recognition models for a training cycle after fine-tuning;
  • the calculation module 403 is used to calculate the fusion coefficient of each parameter in each model after training
  • the update module 404 is configured to determine the fusion weight of each parameter according to the fusion coefficient, and update the parameters of the model to be fused based on the fusion weight of all parameters to obtain a fusion model;
  • the recognition module 405 is configured to perform face recognition on the received face image through the fusion model.
  • the calculation module 403 includes a weight entropy calculation submodule and a fusion coefficient calculation submodule.
  • the weight entropy calculation submodule is used to calculate the weight entropy of each parameter in the face recognition model after training; fusion;
  • the coefficient calculation sub-module is used to calculate the fusion coefficient of the corresponding parameter according to the weight entropy.
  • the weight entropy calculation sub-module includes a division unit and a calculation unit.
  • the division unit is used to calculate the weight value corresponding to each parameter in each face recognition model after training according to the size, etc. Divided into different value ranges; the calculation unit is used to calculate the weight value probability corresponding to each value range, and calculate the weight entropy corresponding to each parameter in each face recognition model according to the probability.
  • the fusion coefficient calculation sub-module is further configured to combine the weight entropy of the same target parameter in all models through the fusion coefficient to obtain the target parameter in the fusion model. Fusion weight.
  • the fine-tuning module 402 is further configured to update the parameters of the model to be fused based on the fusion weight of all parameters, so as to obtain the fusion model after the step of using the preset learning rate to The fusion model is fine-tuned.
  • the face recognition device based on model fusion based on weight grafting further includes a judgment module for judging whether the fusion model has converged; if so, the updated model parameters are used as the result. The final parameters of the fusion model; if not, continue training in the next training cycle until the fusion model converges.
  • the face recognition device based on model fusion based on weight grafting further includes a creation module, which is used to construct before the step of separately training N face recognition models of the same structure N face recognition models with the same structure, the face recognition model including a feature extraction layer, a fully connected layer, and a loss function layer;
  • training module 401 is specifically used for:
  • the feature extraction loss value and the attribute loss value of the loss function layer are calculated based on the loss function, and the feature extraction loss value and the attribute loss value are weighted and summed to obtain the target loss value, and the target loss value is used to judge the person Whether the face recognition model has been trained.
  • the above-mentioned face recognition device based on model fusion based on weight grafting separately trains N face recognition models of the same structure, selects one of the face recognition models as the model to be fused, and fine-tunes all face recognition models after fine-tuning Perform training for one training cycle, calculate the fusion coefficient of each parameter in each model after training, determine the fusion weight of each parameter according to the fusion coefficient, and update the parameters of the model to be fused based on the fusion weight of all parameters to Obtain the fusion model, and finally perform face recognition on the received face image through the fusion model;
  • this application uses weight grafting to fuse multiple face recognition models, and perform the process according to the weight corresponding to each parameter in each model
  • Weighted merging obtains the fusion weight of the fusion model, so that the weight with higher information content has a greater impact on the fusion weight, and the weight with less information has less impact on the fusion weight, which improves the recognition accuracy of the fusion model and improves the execution efficiency of the model.
  • FIG. 5 is a block diagram of the basic structure of the computer device in this embodiment.
  • the computer device 5 includes a memory 51, a processor 52, and a network interface 53 that communicate with each other through a system bus. It should be pointed out that only the computer device 5 with components 51-53 is shown in the figure, but it should be understood that it is not required to implement all the illustrated 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 GateArray, FPGA), Digital Processor (Digital Signal Processor, DSP), embedded equipment, etc.
  • ASIC Application Specific Integrated Circuit
  • ASIC Application Specific Integrated Circuit
  • FPGA Field-Programmable GateArray
  • 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 51 stores computer-readable instructions, and the processor 52 implements the following steps when executing the computer-readable instructions:
  • N is a positive integer
  • the memory 51 includes at least one type of readable storage medium.
  • the readable storage medium includes flash memory, hard disk, multimedia card, card-type memory (for example, SD or DX memory, etc.), random access memory (RAM), and static memory.
  • the memory 51 may be an internal storage unit of the computer device 5, such as a hard disk or a memory of the computer device 5.
  • the memory 51 may also be an external storage device of the computer device 5, for example, a plug-in hard disk equipped on the computer device 5, a smart memory card (Smart Media Card, SMC), and a secure digital (Secure Digital, SD) card, Flash Card, etc.
  • the memory 51 may also include both the internal storage unit of the computer device 5 and its external storage device.
  • the memory 51 is generally used to store an operating system and various application software installed in the computer device 5, such as computer-readable instructions of a face recognition method based on weight grafting and model fusion.
  • the memory 51 can also be used to temporarily store various types of data that have been output or will be output.
  • the processor 52 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 52 is generally used to control the overall operation of the computer device 5.
  • the processor 52 is configured to run computer-readable instructions or processed data stored in the memory 51, for example, computer-readable instructions to run the face recognition method based on weight grafting and model fusion.
  • the network interface 53 may include a wireless network interface or a wired network interface, and the network interface 53 is generally used to establish a communication connection between the computer device 5 and other electronic devices.
  • the steps of the face recognition method based on the model fusion based on weight grafting are implemented as in the above embodiment, and multiple face recognition models are fused by the weight grafting method, and according to The weights corresponding to each parameter in each model are weighted and merged to obtain the fusion weight of the fusion model, so that the weight with a higher amount of information has a greater influence on the fusion weight, and the weight with a smaller amount of information has less influence on the fusion weight, which improves the fusion model While improving the accuracy of recognition, the execution efficiency of the model is improved.
  • the computer-readable storage medium may be non-volatile or volatile.
  • the computer-readable storage medium stores computer-readable instructions, and the computer-readable instructions can be executed by at least one processor, so that the at least one processor executes face recognition based on model fusion based on weight grafting as described above
  • the steps of the method improve the recognition accuracy of the fusion model while improving the execution efficiency of the model.
  • the steps of implementing the aforementioned method for face recognition based on weight grafting and model fusion specifically include:
  • N is a positive integer
  • 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 enable a terminal device (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the method described in each embodiment of the present application.
  • a terminal device which can 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)
  • Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Oral & Maxillofacial Surgery (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computational Linguistics (AREA)
  • Human Computer Interaction (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Multimedia (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Image Analysis (AREA)

Abstract

L'invention concerne un procédé et un appareil de reconnaissance faciale à base de fusion de modèle de greffage de poids, un dispositif informatique et un support d'enregistrement. Le procédé comprend les étapes suivantes : entraînement respectif de N modèles de reconnaissance faciale de structure identique (S201) ; sélection d'un modèle de reconnaissance faciale pour servir de modèle à fusionner, réalisation d'un réglage fin et entraînement de tous les modèles de reconnaissance faciale pendant une période d'apprentissage après le réglage fin (S202) ; calculer un coefficient de fusion de chaque paramètre dans les modèles au terme de l'apprentissage (S203) ; détermination d'un poids de fusion de chaque paramètre en fonction du coefficient de fusion, et mise à jour des paramètres du modèle à fusionner sur la base des poids de fusion de tous les paramètres, de manière à obtenir un modèle de fusion (S204) ; réalisation d'une reconnaissance faciale sur une image faciale reçue au moyen du modèle de fusion (S205). Le procédé a recours à un moyen de greffage de poids pour réaliser une fusion sur une pluralité de modèles de reconnaissance faciale, et augmente l'efficacité d'exécution de modèle tout en améliorant la précision de reconnaissance de modèle de fusion.
PCT/CN2020/135621 2020-09-08 2020-12-11 Procédé de reconnaissance faciale à base de fusion de modèle de greffage de poids, et dispositif y relatif WO2021155713A1 (fr)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010937265.X 2020-09-08
CN202010937265.XA CN112101172B (zh) 2020-09-08 2020-09-08 基于权重嫁接的模型融合的人脸识别方法及相关设备

Publications (1)

Publication Number Publication Date
WO2021155713A1 true WO2021155713A1 (fr) 2021-08-12

Family

ID=73752522

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/135621 WO2021155713A1 (fr) 2020-09-08 2020-12-11 Procédé de reconnaissance faciale à base de fusion de modèle de greffage de poids, et dispositif y relatif

Country Status (2)

Country Link
CN (1) CN112101172B (fr)
WO (1) WO2021155713A1 (fr)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113642491A (zh) * 2021-08-20 2021-11-12 北京百度网讯科技有限公司 人脸融合方法、人脸融合模型的训练方法及装置
CN113657486A (zh) * 2021-08-16 2021-11-16 浙江新再灵科技股份有限公司 基于电梯图片数据的多标签多属性分类模型建立方法
CN113688789A (zh) * 2021-09-17 2021-11-23 华中师范大学 一种基于深度学习的在线学习投入度识别方法及系统
CN115564043A (zh) * 2022-10-18 2023-01-03 上海计算机软件技术开发中心 一种图像分类模型剪枝方法、装置、电子设备及存储介质
CN116824659A (zh) * 2023-05-08 2023-09-29 国网黑龙江省电力有限公司佳木斯供电公司 一种基于特征提取的智能机柜人脸识别方法

Families Citing this family (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112862669B (zh) * 2021-02-02 2024-02-09 百果园技术(新加坡)有限公司 图像生成模型的训练方法、生成方法、装置及设备
CN112598085B (zh) * 2021-03-02 2021-07-30 聚时科技(江苏)有限公司 支持混合模型融合的Model Ensemble方法
CN113435588B (zh) * 2021-08-26 2022-01-04 之江实验室 基于深度卷积神经网络bn层尺度系数的卷积核嫁接方法
CN113706390A (zh) * 2021-10-29 2021-11-26 苏州浪潮智能科技有限公司 图像转换模型训练方法和图像转换方法、设备及介质
WO2023121563A2 (fr) * 2021-12-24 2023-06-29 Grabtaxi Holdings Pte. Ltd. Procédé et système de consultation et d'identification de visage de précision à l'aide d'ensembles multicouches
CN115761850B (zh) * 2022-11-16 2024-03-22 智慧眼科技股份有限公司 人脸识别模型训练方法、人脸识别方法、装置及存储介质
CN117388374B (zh) * 2023-12-13 2024-02-20 南京建正建设工程质量检测有限责任公司 一种建筑用混凝土的强度检测方法及系统

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170132528A1 (en) * 2015-11-06 2017-05-11 Microsoft Technology Licensing, Llc Joint model training
CN109472360A (zh) * 2018-10-30 2019-03-15 北京地平线机器人技术研发有限公司 神经网络的更新方法、更新装置和电子设备
CN110598840A (zh) * 2018-06-13 2019-12-20 富士通株式会社 知识迁移方法、信息处理设备以及存储介质
CN110705600A (zh) * 2019-09-06 2020-01-17 西安交通大学 一种基于互相关熵的多深度学习模型融合方法、终端设备及可读存储介质

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107045618B (zh) * 2016-02-05 2020-07-03 北京陌上花科技有限公司 一种人脸表情识别方法及装置
CN108197532B (zh) * 2017-12-18 2019-08-16 深圳励飞科技有限公司 人脸识别的方法、装置及计算机装置
CN108446650A (zh) * 2018-03-27 2018-08-24 百度在线网络技术（北京）有限公司 用于识别人脸的方法和装置
CN109472211A (zh) * 2018-10-16 2019-03-15 深圳爱莫科技有限公司 人脸识别方法及装置
CN109934197B (zh) * 2019-03-21 2023-07-07 深圳力维智联技术有限公司 人脸识别模型的训练方法、装置和计算机可读存储介质
CN110084216B (zh) * 2019-05-06 2021-11-09 苏州科达科技股份有限公司 人脸识别模型训练和人脸识别方法、系统、设备及介质

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170132528A1 (en) * 2015-11-06 2017-05-11 Microsoft Technology Licensing, Llc Joint model training
CN110598840A (zh) * 2018-06-13 2019-12-20 富士通株式会社 知识迁移方法、信息处理设备以及存储介质
CN109472360A (zh) * 2018-10-30 2019-03-15 北京地平线机器人技术研发有限公司 神经网络的更新方法、更新装置和电子设备
CN110705600A (zh) * 2019-09-06 2020-01-17 西安交通大学 一种基于互相关熵的多深度学习模型融合方法、终端设备及可读存储介质

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113657486A (zh) * 2021-08-16 2021-11-16 浙江新再灵科技股份有限公司 基于电梯图片数据的多标签多属性分类模型建立方法
CN113657486B (zh) * 2021-08-16 2023-11-07 浙江新再灵科技股份有限公司 基于电梯图片数据的多标签多属性分类模型建立方法
CN113642491A (zh) * 2021-08-20 2021-11-12 北京百度网讯科技有限公司 人脸融合方法、人脸融合模型的训练方法及装置
CN113688789A (zh) * 2021-09-17 2021-11-23 华中师范大学 一种基于深度学习的在线学习投入度识别方法及系统
CN113688789B (zh) * 2021-09-17 2023-11-10 华中师范大学 一种基于深度学习的在线学习投入度识别方法及系统
CN115564043A (zh) * 2022-10-18 2023-01-03 上海计算机软件技术开发中心 一种图像分类模型剪枝方法、装置、电子设备及存储介质
CN115564043B (zh) * 2022-10-18 2023-10-27 上海计算机软件技术开发中心 一种图像分类模型剪枝方法、装置、电子设备及存储介质
CN116824659A (zh) * 2023-05-08 2023-09-29 国网黑龙江省电力有限公司佳木斯供电公司 一种基于特征提取的智能机柜人脸识别方法

Also Published As

Publication number Publication date
CN112101172A (zh) 2020-12-18
CN112101172B (zh) 2023-09-08

Similar Documents

Publication Publication Date Title
WO2021155713A1 (fr) Procédé de reconnaissance faciale à base de fusion de modèle de greffage de poids, et dispositif y relatif
WO2021120677A1 (fr) Procédé et appareil d'entraînement de modèle d'entreposage, dispositif informatique et support de stockage
WO2022095352A1 (fr) Procédé et appareil d'identification d'utilisateur anormal basés sur une décision intelligente, et dispositif informatique
CN113435583B (zh) 基于联邦学习的对抗生成网络模型训练方法及其相关设备
CN112035549B (zh) 数据挖掘方法、装置、计算机设备及存储介质
WO2021143267A1 (fr) Procédé de traitement de modèle de classification à grain fin basé sur la détection d'image, et dispositifs associés
CN112766649B (zh) 基于多评分卡融合的目标对象评价方法及其相关设备
WO2021174877A1 (fr) Procédé de traitement pour modèle de détection de cible basé sur les décisions intelligentes, et dispositif associé
WO2022110640A1 (fr) Procédé et appareil d'optimisation de modèle, dispositif informatique et support de stockage
WO2022156084A1 (fr) Procédé pour prédire le comportement d'un objet cible sur la base d'un visage et d'un texte interactif, et dispositif associé
CN112633973A (zh) 一种商品推荐方法及其相关设备
KR20220154062A (ko) 내용 추천 및 정렬 모델 훈련 방법, 장치, 기기, 저장 매체 및 컴퓨터 프로그램
CN112036483B (zh) 基于AutoML的对象预测分类方法、装置、计算机设备及存储介质
CN112785005A (zh) 多目标任务的辅助决策方法、装置、计算机设备及介质
CN112634158A (zh) 人脸图像恢复方法、装置、计算机设备及存储介质
CN113420212A (zh) 基于深度特征学习的推荐方法、装置、设备及存储介质
WO2022001233A1 (fr) Procédé de pré-étiquetage basé sur un apprentissage par transfert hiérarchique et dispositif associé
CN113360683A (zh) 训练跨模态检索模型的方法以及跨模态检索方法和装置
CN112100491A (zh) 基于用户数据的信息推荐方法、装置、设备及存储介质
CN116777646A (zh) 基于人工智能的风险识别方法、装置、设备及存储介质
CN115099875A (zh) 基于决策树模型的数据分类方法及相关设备
WO2022142032A1 (fr) Procédé et appareil de vérification de signature manuscrite, dispositif informatique et support de stockage
CN115203391A (zh) 一种信息检索方法、装置、计算机设备及存储介质
CN114881033A (zh) 一种文本摘要生成方法、装置、计算机设备及存储介质
CN113936677A (zh) 音色转换方法、装置、计算机设备及存储介质

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: 20917752

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: 20917752

Country of ref document: EP

Kind code of ref document: A1