WO2021180243A1 - 基于机器学习的图像信息识别的优化方法及装置 - Google Patents

基于机器学习的图像信息识别的优化方法及装置 Download PDF

Info

Publication number
WO2021180243A1
WO2021180243A1 PCT/CN2021/083875 CN2021083875W WO2021180243A1 WO 2021180243 A1 WO2021180243 A1 WO 2021180243A1 CN 2021083875 W CN2021083875 W CN 2021083875W WO 2021180243 A1 WO2021180243 A1 WO 2021180243A1
Authority
WO
WIPO (PCT)
Prior art keywords
training
deep learning
sample set
learning model
training sample
Prior art date
Application number
PCT/CN2021/083875
Other languages
English (en)
French (fr)
Inventor
张国辉
姜禹
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 WO2021180243A1 publication Critical patent/WO2021180243A1/zh

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/22—Matching criteria, e.g. proximity measures
    • 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
    • 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 data processing technology, and in particular to an optimization method and device for image information recognition based on machine learning.
  • machine learning has become the basis of data processing, especially the use of deep learning models to process image information.
  • the inventor realizes that before the existing deep learning model is used to process image information, specific training data will be selected to train the deep learning model during the training process of the deep learning model, for example, as a training task for a training task.
  • the data set trains the deep learning model, and when the loss function loss, which is a basis for judging whether the deep learning model training is completed, does not change, it indicates that the deep learning model has completed the training, which can be characterized as the deep learning model falling into a local optimal solution, or
  • the first derivative of the deep learning model is equal to zero
  • the deep learning model cannot continue to be optimized, but at this time for the deep learning model, the optimal training has not been completed, and the deep learning model at this time is used for image information Recognition will affect the accuracy of image information recognition, resulting in low image information recognition efficiency, thereby affecting the recognition effect of image information based on machine learning.
  • the present application provides an optimization method and device for image information recognition based on machine learning.
  • the main purpose is to solve the problem that the existing deep learning model in the local optimal or saddle point recognizes the image information, which affects the recognition of image information. Accuracy leads to low image information recognition efficiency, which affects the recognition effect of image information based on machine learning.
  • an optimization method for image information recognition based on machine learning including: obtaining a main task image training sample set and at least one branch task image training sample set, the main task image training sample set The branch task image training sample set is matched; based on the main task image training sample set and the branch task image training sample set, the preset deep learning model is subjected to the training process of switching between the main task training and the branch task training, so that According to the model accuracy and loss value, it is determined that the preset deep learning model is in a non-local optimal and/or non-saddle-point state to complete the training process; based on the preset deep learning model that has completed the training, the image information to be recognized is recognized, Obtain the recognition result of the image information.
  • an optimization device for image information recognition based on machine learning including: an acquisition module for acquiring a main task image training sample set and at least one branch task image training sample set, the main task The image training sample set matches the branch task image training sample set; the training module is used to perform main task training on the preset deep learning model based on the main task image training sample set and the branch task image training sample set.
  • a storage medium in which at least one executable instruction is stored, and the executable instruction causes a processor to execute the following method: acquiring a main task image training sample set and at least one A branch task image training sample set, the main task image training sample set matches the branch task image training sample set; based on the main task image training sample set, the branch task image training sample set to preset deep learning
  • the model performs the training process of switching between the main task training and the branch task training, so that the preset deep learning model is determined to be in a non-local optimal and/or non-saddle point state according to the model accuracy and loss value to complete the training process; based on completion
  • the trained preset deep learning model performs recognition processing on the image information to be recognized, and obtains the recognition result of the image information.
  • a computer device including: a processor, a memory, a communication interface, and a communication bus.
  • the processor, the memory, and the communication interface complete mutual communication through the communication bus.
  • Communication; the memory is used to store at least one executable instruction, the executable instruction causes the processor to execute the following method: obtain a main task image training sample set and at least one branch task image training sample set, the main task image
  • the training sample set matches the branch task image training sample set; based on the main task image training sample set and the branch task image training sample set to perform the main task training and branch task training of the preset deep learning model.
  • the preset deep learning model is determined to be in a non-local optimal and/or non-saddle-point state according to the model accuracy and loss value to complete the training process; based on the preset deep learning model that has completed the training, the image to be recognized
  • the information is subjected to recognition processing, and the recognition result of the image information is obtained.
  • This application helps to avoid the model training of the deep learning model from falling into the local optimum or saddle point, which makes the model accuracy poor, greatly improves the optimization effect of the deep learning model, and effectively solves the problem of the deep learning model jumping out of the optimum Therefore, when the image information is recognized, the need for high-precision recognition based on machine learning is realized, and the recognition efficiency of image information is improved.
  • Fig. 1 shows a flowchart of a method for optimizing image information recognition based on machine learning provided by an embodiment of the present application.
  • Fig. 2 shows a flowchart of another method for optimizing image information recognition based on machine learning provided by an embodiment of the present application.
  • Fig. 3 shows a block diagram of a device for optimizing image information recognition based on machine learning provided by an embodiment of the present application.
  • Fig. 4 shows a block diagram of another device for optimizing image information recognition based on machine learning provided by an embodiment of the present application.
  • Fig. 5 shows a schematic structural diagram of a computer device provided by an embodiment of the present application.
  • the technical solution of the present application may involve the field of artificial intelligence and/or big data technology, for example, may specifically involve neural network technology to realize image recognition.
  • the data involved in this application such as training samples and/or recognition results, can be stored in a database, or can be stored in a blockchain, which is not limited in this application.
  • the embodiment of the present application provides an optimization method for image information recognition based on machine learning. As shown in FIG. 1, the method includes the following steps.
  • the main task training is used to characterize the main task image training sample set as the training that must be performed on the preset deep learning model
  • the branch task training is used to characterize the preset deep learning model with the branch task image training sample set as an optional Training performed
  • the main task image training sample set matches the branch task image training sample set, specifically, for image information recognition
  • the main task image training sample set is determined according to the recognition characteristics of the image information
  • at least one The branch task image training sample set so as to realize the main task training as the main training, and the branch task training as the auxiliary training.
  • one main task image training sample set can correspond to multiple branch task image training samples, and the corresponding relationship is bound according to different recognition features.
  • the recognition features are face age, face color, and gender, then they can be obtained as the main task training
  • the main task image training sample set of is the face age
  • the branch task image training sample set is the image information set of the face color and gender, which is not specifically limited in the embodiment of the application.
  • the model executes the branch task training processing of the branch task image training samples, so that the preset deep learning model is determined to be in a non-local optimal and/or non-saddle point state according to the model accuracy and loss value to complete the training processing.
  • the preset deep learning model trained on the branch task jumps out of the local optimal and/or saddle point state
  • the main task training of the main task image training sample set is re-executed to complete the complete training process of the preset deep learning model .
  • branch task training can be multiple, when switching branch task training, branch task training can be performed on the branch task image training sample set in a preset order.
  • preset deep learning model in the embodiment of the present application is a model for recognizing image information, and may be a neural network model, a support vector machine model, etc., and is not specifically limited.
  • the preset deep learning model in order that after the training of the preset deep learning model is completed, the preset deep learning model is in a non-local optimum, and/or the image recognition accuracy in the saddle point state reaches the preset accuracy threshold, therefore, Using this preset deep learning model to perform recognition processing on the image information to be recognized greatly improves the recognition accuracy of the image information through the machine learning model trained in the main task and branch tasks, and improves the optimization effect of the model training.
  • the embodiment of the present application provides an optimization method for image information recognition based on machine learning.
  • the embodiment of the present application obtains a main task image training sample set and at least one branch task image training sample set, the main task image training sample set matches the branch task image training sample set;
  • the main task image training sample set and the branch task image training sample set perform the training process of switching between the main task training and the branch task training on a preset deep learning model, so as to determine according to the model accuracy and loss value
  • the preset deep learning model is in a non-local optimal and/or non-saddle-point state to complete the training process; based on the preset deep learning model that has completed the training, the image information to be identified is identified to obtain the identification of the image information
  • it avoids the model training of the deep learning model from falling into the local optimum or saddle point, which makes the model accuracy poor, greatly improves the optimization effect of the deep learning model, and effectively solves the problem of the deep learning model jumping out of the optimum.
  • the image information is
  • the embodiment of the present application provides another method for optimizing image information recognition based on machine learning. As shown in FIG. 2, the method includes the following steps.
  • the identification features include at least one of five sense features, gender features, age features, facial features, text features, and numeric features in the image information.
  • main task training is required to The specific key and necessary feature training tasks for classification and recognition in image information, for example, in face recognition, for the main task training model for recognizing the age of the face, the branch task is trained as at least one pair of image information related to the main task Perform training tasks for non-critical features of classification and recognition.
  • the branch task training can include the training tasks of face color and gender recognition, or increase the eyebrow density and the aspect ratio of the nose bridge during the face recognition model training. Branch task training. Therefore, for different main task training and branch task training, corresponding training samples are constructed.
  • the image information uses face recognition as the main recognition target. Therefore, it is preferable that the image information recognized or marked as facial features is used as the main task image training sample set, and the recognition Or image information marked as gender features, age features, expression features, and text features are used as the branch task image training sample set.
  • image training sample sets for branch tasks there can be one or multiple image training sample sets for branch tasks, and image training sample sets for branch tasks can be used or not used when training for the main task, for example, branch task image sample sets Including a, b, c, the corresponding main task image training sample set is S, when the main task image training sample set is used to perform the main task training processing, the preset deep learning model falls into the local optimum, then the main task training is stopped, The branch task image sample set a performs branch task training processing on the preset deep learning model that has performed the main task training.
  • the main task image training sample set is reused
  • the branch task image sample sets b and c are no longer executed, which is not specifically limited in the embodiment of the present application.
  • the preset deep learning model that executes the main task training processing is in a local optimal and/or saddle point state, perform the main task image training sample set matching the main task image training sample set.
  • the preset deep learning model of the task training processing executes branch task training processing.
  • the local optimal and saddle point state is used to indicate that the training of the preset deep learning model is in a stagnant state, and it is impossible to continue to obtain the matching training target through the training model. Therefore, it is necessary to use the branch task image sample set pair at this time.
  • This preset deep learning model performs branch task training processing. That is, when the main task training falls into the local optimum, the weight of the preset deep learning model is updated through the branch task training, thereby jumping out of the local optimum.
  • one training sample set of image training samples for branch tasks can be randomly selected, and after a certain number of iterations of branch task training, the main task training is executed, which is not specifically limited in the embodiment of this application. .
  • the training process of the main task training and the branch task training is only different from the training sample set, and the training steps of the preset deep learning model are exactly the same, and the embodiment of the application does not specifically limit it.
  • step 203 it further includes: performing recursive calculations based on the model accuracy and loss value of the preset deep learning model that performs the main task training process, When the model accuracy and loss value remain unchanged, it is determined that the preset deep learning model that performs the main task training processing is in a local optimal and/or saddle point state, and the preset deep learning model is a neural network model.
  • the model accuracy and loss value of the preset deep learning model that performs the main task training processing are calculated by recursive calculation.
  • the model accuracy and loss value are not Change, it means that the preset deep learning model is in a local optimal, saddle point state.
  • the preset deep learning model is a neural network model.
  • the loss value loss or model accuracy acc is unchanged, it means that the first derivative of the neural network model is zero, and the gradient descent cannot be continued, that is, the neural network model cannot continue to be optimized and is at the local optimum. , There is a saddle point.
  • the loss function is used to evaluate the degree of difference between the predicted value of the model and the true value, and the loss value is calculated.
  • the loss function is also the objective function optimized in the neural network.
  • the process of neural network training or optimization is the process of minimizing the loss function.
  • the smaller the loss function the closer the predicted value of the neural network model is to the true value, and the accuracy is also The better.
  • the embodiment of the application only calculates the loss value or model accuracy of the neural network model to determine whether to enter the branch task, in the embodiment of the application, the number of layers of the neural network is not specifically limited, and it can be one layer. It can also be multi-layered, for example, for a layer of neural network, a perceptron model with several inputs and one output.
  • the loss function may be a square loss function, a logarithmic loss function, a cross-entropy loss function, and other different forms of loss functions to calculate the loss value of the neural network, which is not specifically limited in the embodiment of the application.
  • the judgment basis for configuring the jump out of the local optimum and saddle point state is to meet the preset loss threshold or reach the preset training
  • the time and number of iterations are determined to be out of the local optimal and saddle point state. Therefore, the branch task training is stopped, and the main task training process is performed again on the neural network model that has undergone branch task training through the main task image training sample set.
  • the neural network model converges slowly and requires a large number of neurons to fit similar features, so that the number of iterations should be as few as possible in the initial stage of branch task training. , You can also set the number of iterations as the stop condition, for example, the branch task will continue the training of the main task after 10 iterations.
  • the preset training time and preset loss threshold can be used to make judgments.
  • the neural network at this time is determined
  • the model is to complete the training process.
  • the embodiment of the present application It also includes: when the main task training process is re-executed, and the preset deep learning model is in a local optimal and/or saddle point state, then reapplying the preset depth based on the branch task image training sample set again
  • the learning model executes the step of branch task training processing; when the preset deep learning model that executes the branch task training processing jumps out of the local optimum and the saddle point state, it executes again to perform the alignment based on the main task image training sample set
  • the preset deep learning model re-executes the steps of the main task training process until the model accuracy of the preset deep learning model matches the preset accuracy threshold, and it is determined that the training of the preset deep learning model is completed.
  • the branch task image training sample set is used again to re-execute the branch task training process.
  • the main task training is performed again, and this cycle until the main task training performed on the preset deep learning model If the model accuracy matches the preset accuracy threshold, it is determined to complete the training of the preset deep learning model.
  • the selection of the branch task image sample set for re-executing the branch task training can be performed sequentially according to a preset order.
  • the branch task image sample set includes a, b, c, and the corresponding main task image training sample set is S.
  • the preset deep learning model falls into the local optimum. Then stop the main task training, and execute the branch task training process on the preset deep learning model that has performed the main task training through the branch task image sample set a.
  • the preset deep learning model for branch task training jumps out of the local optimum, it is reused
  • the main task image training sample set is trained.
  • the execution of the main task training is stopped, and the branch task image sample set b is used to execute the branch again on the preset deep learning model that re-executes the main task training.
  • Task training when using the preset loss threshold, or the number of iterations, and the training time to determine to jump out of the local optimal and saddle point state, the main task training is re-executed, and this cycle is repeated until the preset deep learning model that executes the main task training completes the training.
  • the neural network model for the neural network model that has been trained, it is applicable but not limited to the classification and recognition of facial image information and card image information.
  • the 18-version Hong Kong ID card can be judged in the card authentication system Whether the middle hologram exists, you can also increase whether the image in the corresponding area is bright, the color vividness, and whether the background color meets the normal background recognition.
  • the method further includes: analyzing the recognition result of the image information, and performing analysis on the main task based on the parsed recognition feature.
  • the image training sample set and the branch task image training sample set are updated.
  • the recognition result of image information includes the classification of the recognition characteristics of the image information. Therefore, as a supplement and optimization to the image training set, the main task image training sample set and branch are analyzed based on the recognition characteristics in the recognition result.
  • the task image training sample set is updated.
  • the update includes the deletion, merging, and replacement of different recognition features of the same image information, and the specific update method is determined according to the number of image information in the image training set, which is not specifically limited in the embodiment of the present application.
  • the embodiments of the present application provide another optimization method for image information recognition based on machine learning.
  • the embodiment of the present application obtains a main task image training sample set and at least one branch task image training sample set, the main task image training sample set matches the branch task image training sample set;
  • the main task image training sample set and the branch task image training sample set perform the training process of switching between the main task training and the branch task training on a preset deep learning model, so as to determine according to the model accuracy and loss value
  • the preset deep learning model is in a non-local optimal and/or non-saddle-point state to complete the training process; based on the preset deep learning model that has completed the training, the image information to be identified is identified to obtain the identification of the image information
  • it avoids the model training of the deep learning model from falling into the local optimum or saddle point, which makes the model accuracy poor, greatly improves the optimization effect of the deep learning model, and effectively solves the problem of the deep learning model jumping out of the optimum.
  • the image information to be identified is
  • an embodiment of the present application provides an optimization device for image information recognition based on machine learning.
  • the device includes: an acquisition module 31 for acquiring A main task image training sample set and at least one branch task image training sample set, where the main task image training sample set matches the branch task image training sample set; the training module 32 is configured to train samples based on the main task image Set, the branch task image training sample set performs the training process of switching between the main task training and the branch task training on a preset deep learning model, so that the preset deep learning model is determined according to model accuracy and loss value The training process is completed in a non-local optimal and/or non-saddle point state; the recognition module 33 is configured to perform recognition processing on the image information to be recognized based on the preset deep learning model that has completed the training to obtain the recognition result of the image information .
  • the embodiment of the present application provides an optimization device for image information recognition based on machine learning.
  • the embodiment of the present application obtains a main task image training sample set and at least one branch task image training sample set, the main task image training sample set matches the branch task image training sample set;
  • the main task image training sample set and the branch task image training sample set perform the training process of switching between the main task training and the branch task training on a preset deep learning model, so as to determine according to the model accuracy and loss value
  • the preset deep learning model is in a non-local optimal and/or non-saddle-point state to complete the training process; based on the preset deep learning model that has completed the training, the image information to be identified is identified to obtain the identification of the image information
  • it avoids the model training of the deep learning model from falling into the local optimum or saddle point, which makes the model accuracy poor, greatly improves the optimization effect of the deep learning model, and effectively solves the problem of the deep learning model jumping out of the optimum.
  • the image information is
  • an embodiment of the present application provides another device for optimizing image information recognition based on machine learning.
  • the device includes: an acquisition module 41 for Obtain a main task image training sample set and at least one branch task image training sample set, where the main task image training sample set matches the branch task image training sample set; the training module 42 is configured to train based on the main task image
  • the sample set and the branch task image training sample set perform the training process of switching between the main task training and the branch task training on a preset deep learning model, so that the preset deep learning is determined according to the model accuracy and the loss value
  • the training process is completed when the model is in a non-local optimal and/or non-saddle point state;
  • the recognition module 43 is used to perform recognition processing on the image information to be recognized based on the preset deep learning model that has completed the training to obtain the recognition of the image information result.
  • the training module 42 includes: a first training unit 4201, configured to perform the main task training process when the preset deep learning model is in a local optimum and/or a saddle point state, according to the main task
  • the branch task image training sample set matched by the image training sample set performs branch task training processing on the preset deep learning model that has performed the main task training processing
  • the second training unit 4202 is used when performing the branch task training processing
  • the preset deep learning model jumps out of the local optimum and the saddle point state, and then re-executes the main task training process on the preset deep learning model according to the main task image training sample set
  • the first determining unit 4203 is configured to When the model accuracy of the preset deep learning model after re-executing the main task training process matches the preset accuracy threshold, it is determined that the training of the preset deep learning model is completed.
  • the first training unit 4201 is further configured to, when the main task training process is re-executed, and the preset deep learning model is in a local optimal and/or saddle point state, then based on the branch task again
  • the image training sample set re-executes the step of branch task training processing on the preset deep learning model
  • the second training unit 4202 is also used for when the preset deep learning model performing branch task training processing jumps out of the local
  • the step of re-executing the main task training process on the preset deep learning model according to the main task image training sample set is performed again until the model accuracy of the preset deep learning model matches the preset depth learning model.
  • the accuracy threshold is set to determine the completion of the training of the preset deep learning model.
  • the first training unit 4201 is specifically configured to, if the branch task image training sample set is multiple, based on the identification feature matched by the branch task image training sample set, and the identification feature corresponds to the image information mapping The relationship determines the order of performing branch task training processing, and executes multiple branch task training processing on the preset deep learning model that has performed the main task training processing in the order.
  • the training module further includes: a second determining unit 4204, configured to perform recursive calculations based on the model accuracy and loss value of the preset deep learning model that performs the main task training process, when the model accuracy, loss When the value does not change, it is determined that the preset deep learning model that performs the main task training processing is in a local optimal and/or saddle point state, and the preset deep learning model is a neural network model.
  • a second determining unit 4204 configured to perform recursive calculations based on the model accuracy and loss value of the preset deep learning model that performs the main task training process, when the model accuracy, loss When the value does not change, it is determined that the preset deep learning model that performs the main task training processing is in a local optimal and/or saddle point state, and the preset deep learning model is a neural network model.
  • the device further includes: a construction module 44 for constructing a main task image training sample set and at least one branch task image training sample set based on the recognition characteristics of each image information in the image training set to be trained.
  • Recognition features include at least one of five sense features, gender features, age features, facial features, text features, and numeric features in the image information.
  • the device further includes: an update module 45, configured to parse and obtain the recognition result of the image information, and perform processing on the main task image training sample set and the branch task image training sample set based on the parsed recognition feature renew.
  • an update module 45 configured to parse and obtain the recognition result of the image information, and perform processing on the main task image training sample set and the branch task image training sample set based on the parsed recognition feature renew.
  • the embodiment of the present application provides another optimization device for image information recognition based on machine learning.
  • the embodiment of the present application obtains a main task image training sample set and at least one branch task image training sample set, the main task image training sample set matches the branch task image training sample set;
  • the main task image training sample set and the branch task image training sample set perform the training process of switching between the main task training and the branch task training on a preset deep learning model, so as to determine according to the model accuracy and loss value
  • the preset deep learning model is in a non-local optimal and/or non-saddle-point state to complete the training process; based on the preset deep learning model that has completed the training, the image information to be identified is identified to obtain the identification of the image information
  • it avoids the model training of the deep learning model from falling into the local optimum or saddle point, which makes the model accuracy poor, greatly improves the optimization effect of the deep learning model, and effectively solves the problem of the deep learning model jumping out of the optimum.
  • the image information is
  • a storage medium stores at least one executable instruction, and the computer executable instruction can execute the optimization method for image information recognition based on machine learning in any of the foregoing method embodiments.
  • the storage medium involved in this application may be a computer-readable storage medium, and the storage medium, such as a computer-readable storage medium, may be non-volatile or volatile.
  • FIG. 5 shows a schematic structural diagram of a computer device according to an embodiment of the present application, and the specific embodiment of the present application does not limit the specific implementation of the computer device.
  • the computer device may include: a processor 502, a communication interface (Communications Interface) 504, memory (memory) 506, and communication bus 508.
  • a processor 502 may include: a processor 502, a communication interface (Communications Interface) 504, memory (memory) 506, and communication bus 508.
  • the processor 502, the communication interface 504, and the memory 506 communicate with each other through the communication bus 508.
  • the communication interface 504 is used to communicate with other devices, such as network elements such as clients or other servers.
  • the processor 502 is configured to execute the program 510, and specifically can execute the relevant steps in the embodiment of the above-mentioned optimization method for image information recognition based on machine learning.
  • the program 510 may include program code, and the program code includes a computer operation instruction.
  • the processor 502 may be a central processing unit CPU, or a specific integrated circuit ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of the present application.
  • the one or more processors included in the computer device may be the same type of processor, such as one or more CPUs, or different types of processors, such as one or more CPUs and one or more ASICs.
  • the memory 506 is used to store the program 510.
  • the memory 506 may include a high-speed RAM memory, and may also include a non-volatile memory (non-volatile memory), for example, at least one magnetic disk memory.
  • the program 510 may specifically be used to cause the processor 502 to perform the following operations: obtain a main task image training sample set and at least one branch task image training sample set, where the main task image training sample set matches the branch task image training sample set Based on the main task image training sample set, the branch task image training sample set to perform the main task training and branch task training of the preset deep learning model to perform the training process; based on the completion of the training preset deep learning model, treat The recognized image information is subjected to recognition processing, and the recognition result of the image information is obtained.
  • modules or steps of this application can be implemented by a general computing device, and they can be concentrated on a single computing device or distributed in a network composed of multiple computing devices.
  • they can be implemented with program codes executable by a computing device, so that they can be stored in a storage device for execution by the computing device, and in some cases, they can be executed in a different order than here.

Landscapes

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

Abstract

一种基于机器学习的图像信息识别的优化方法及装置，涉及数据处理技术领域，主要目的在于解决现有处于局部最优或鞍点的深度学习模型对图片信息进行识别，影响对图像信息识别的准确性，导致图像信息识别效率较低，影响基于机器学习的图像信息的识别效果的问题。包括：获取主任务图像训练样本集以及至少一个分支任务图像训练样本集（S101）；基于所述主任务图像训练样本集、所述分支任务图像训练样本集对预设深度学习模型执行主任务训练及分支任务训练相切换的训练处理（S102）；基于完成训练的预设深度学习模型，对待识别的图像信息进行识别处理，得到所述图像信息的识别结果（S103）。

Description

基于机器学习的图像信息识别的优化方法及装置
本申请要求于2020年11月2日提交中国专利局、申请号为202011201765.3，发明名称为“基于机器学习的图像信息识别的优化方法及装置”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
本申请涉及一种数据处理技术领域，特别是涉及一种基于机器学习的图像信息识别的优化方法及装置。
 背景技术
随着机器学习的快速发展，机器学习已经成为对数据处理的基础，尤其是利用深度学习模型来对图像信息进行处理。
目前，发明人意识到，现有的利用深度学习模型对图片信息进行处理前，对于深度学习模型的训练过程会选取特定的训练数据对深度学习模型进行训练，例如，作为一个训练任务的一个训练数据集对深度学习模型进行训练，且作为深度学习模型训练是否完成的一个判断依据的损失函数loss在不变化时，说明深度学习模型完成训练，可以表征为深度学习模型陷入局部最优解，或者处于鞍点，即深度学习模型的一阶导数等于零，深度学习模型无法在继续优化，但是此时对于深度学习模型来说，并没有完成最优的训练，利用此时的深度学习模型进行图像信息的识别，会影响对图像信息识别的准确性，导致图像信息识别效率较低，从而影响基于机器学习的图像信息的识别效果。
 技术问题
有鉴于此，本申请提供一种基于机器学习的图像信息识别的优化方法及装置，主要目的在于解决现有处于局部最优或鞍点的深度学习模型对图片信息进行识别，影响对图像信息识别的准确性，导致图像信息识别效率较低，从而影响基于机器学习的图像信息的识别效果的问题。
 技术解决方案
依据本申请一个方面，提供了一种基于机器学习的图像信息识别的优化方法，包括：获取主任务图像训练样本集以及至少一个分支任务图像训练样本集，所述主任务图像训练样本集与所述分支任务图像训练样本集相匹配；基于所述主任务图像训练样本集、所述分支任务图像训练样本集对预设深度学习模型执行主任务训练及分支任务训练相切换的训练处理，以使根据模型精度、损失值确定所述预设深度学习模型处于非局部最优、和/或非鞍点状态下完成训练处理；基于完成训练的预设深度学习模型，对待识别的图像信息进行识别处理，得到所述图像信息的识别结果。
依据本申请另一个方面，提供了一种基于机器学习的图像信息识别的优化装置，包括：获取模块，用于获取主任务图像训练样本集以及至少一个分支任务图像训练样本集，所述主任务图像训练样本集与所述分支任务图像训练样本集相匹配；训练模块，用于基于所述主任务图像训练样本集、所述分支任务图像训练样本集对预设深度学习模型执行主任务训练及分支任务训练相切换的训练处理，以使根据模型精度、损失值确定所述预设深度学习模型处于非局部最优、和/或非鞍点状态下完成训练处理；识别模块，用于基于完成训练的预设深度学习模型，对待识别的图像信息进行识别处理，得到所述图像信息的识别结果。
根据本申请的又一方面，提供了一种存储介质，所述存储介质中存储有至少一可执行指令，所述可执行指令使处理器执行以下方法：获取主任务图像训练样本集以及至少一个分支任务图像训练样本集，所述主任务图像训练样本集与所述分支任务图像训练样本集相匹配；基于所述主任务图像训练样本集、所述分支任务图像训练样本集对预设深度学习模型执行主任务训练及分支任务训练相切换的训练处理，以使根据模型精度、损失值确定所述预设深度学习模型处于非局部最优、和/或非鞍点状态下完成训练处理；基于完成训练的预设深度学习模型，对待识别的图像信息进行识别处理，得到所述图像信息的识别结果。
根据本申请的再一方面，提供了一种计算机设备，包括：处理器、存储器、通信接口和通信总线，所述处理器、所述存储器和所述通信接口通过所述通信总线完成相互间的通信；所述存储器用于存放至少一可执行指令，所述可执行指令使所述处理器执行以下方法：获取主任务图像训练样本集以及至少一个分支任务图像训练样本集，所述主任务图像训练样本集与所述分支任务图像训练样本集相匹配；基于所述主任务图像训练样本集、所述分支任务图像训练样本集对预设深度学习模型执行主任务训练及分支任务训练相切换的训练处理，以使根据模型精度、损失值确定所述预设深度学习模型处于非局部最优、和/或非鞍点状态下完成训练处理；基于完成训练的预设深度学习模型，对待识别的图像信息进行识别处理，得到所述图像信息的识别结果。
 有益效果
本申请有助于避免因对深度学习模型进行模型训练陷入局部最优或鞍点使得模型精度较差，大大提高了对深度学习模型的优化效果，并有效的解决了深度学习模型跳出不接最优的问题，从而实现了对图像信息进行识别时，基于机器学习的识别高精度的需求，提高图像信息的识别效率。
 附图说明
图1示出了本申请实施例提供的一种基于机器学习的图像信息识别的优化方法流程图。
图2示出了本申请实施例提供的另一种基于机器学习的图像信息识别的优化方法流程图。
图3示出了本申请实施例提供的一种基于机器学习的图像信息识别的优化装置组成框图。
图4示出了本申请实施例提供的另一种基于机器学习的图像信息识别的优化装置组成框图。
图5示出了本申请实施例提供的一种计算机设备的结构示意图。
 本发明的实施方式
下面将参照附图更详细地描述本公开的示例性实施例。虽然附图中显示了本公开的示例性实施例，然而应当理解，可以以各种形式实现本公开而不应被这里阐述的实施例所限制。相反，提供这些实施例是为了能够更透彻地理解本公开，并且能够将本公开的范围完整的传达给本领域的技术人员。
本申请的技术方案可涉及人工智能和/或大数据技术领域，如可具体涉及神经网络技术，以实现图像识别。可选的，本申请涉及的数据如训练样本和/或识别结果等可存储于数据库中，或者可以存储于区块链中，本申请不做限定。
本申请实施例提供了一种基于机器学习的图像信息识别的优化方法，如图1所示，该方法包括以下步骤。
101、获取主任务图像训练样本集以及至少一个分支任务图像训练样本集。
其中，主任务训练用于表征以主任务图像训练样本集为必须对预设深度学习模型执行的训练，分支任务训练用于表征以分支任务图像训练样本集为可选的对预设深度学习模型执行的训练，所述主任务图像训练样本集与所述分支任务图像训练样本集相匹配，具体的，对于图像信息的识别，根据图像信息的识别特征确定主任务图像训练样本集，以及至少一个分支任务图像训练样本集，从而实现以主任务训练为主要训练，以分支任务训练为辅助训练。另外，一个主任务图像训练样本集可以对应多个分支任务图像训练样本，根据不同的识别特征进行绑定对应关系，例如，识别特征为人脸年龄、人脸颜色、性别，则获取作为主任务训练的主任务图像训练样本集为人脸年龄，则分支任务图像训练样本集为人脸颜色、性别的图像信息集合，本申请实施例不做具体限定。
102、基于所述主任务图像训练样本集、所述分支任务图像训练样本集对预设深度学习模型执行主任务训练及分支任务训练相切换的训练处理。
其中，首先通过主任务图像训练样本对预设深度学习模型执行训练处理，然后当预设深度学习模型处于局部最优、和/或鞍点状态下切换为对已执行主任务训练的预设深度学习模型执行分支任务图像训练样本的分支任务训练处理，以使根据模型精度、损失值确定所述预设深度学习模型处于非局部最优、和/或非鞍点状态下完成训练处理。具体的，当基于分支任务训练的预设深度学习模型跳出局部最优、和/或鞍点状态，则重新执行主任务图像训练样本集的主任务训练，以完成预设深度学习模型的完整训练过程。
需要说明的是，由于分支任务训练可以为多个，因此，在切换分支任务训练时，可以按照预设的顺序对分支任务图像训练样本集进行分支任务训练。另外，本申请实施例中的预设深度学习模型为针对图像信息进行识别的模型，可以为神经网络模型、支持向量机模型等，不做具体限定。
103、基于完成训练的预设深度学习模型，对待识别的图像信息进行识别处理，得到所述图像信息的识别结果。
对于本申请实施例，为了当完成了对预设深度学习模型的训练后，此预设深度学习模型为处于非局部最优、和/或鞍点状态下图像识别精度达到预设精度阈值，因此，利用此预设深度学习模型对待识别的图像信息进行识别处理，大大提高了通过主任务、分支任务训练的机器学习模型对图像信息进行识别的识别精度，提高了模型训练的优化效果。
本申请实施例提供了一种基于机器学习的图像信息识别的优化方法。与现有技术相比，本申请实施例通过 获取主任务图像训练样本集以及至少一个分支任务图像训练样本集，所述主任务图像训练样本集与所述分支任务图像训练样本集相匹配；基于所述主任务图像训练样本集、所述分支任务图像训练样本集对预设深度学习模型执行所述主任务训练及所述分支任务训练相切换的训练处理，以使根据模型精度、损失值确定所述预设深度学习模型处于非局部最优、和/或非鞍点状态下完成训练处理；基于完成训练的预设深度学习模型，对待识别的图像信息进行识别处理，得到所述图像信息的识别结果，避免因对深度学习模型进行模型训练陷入局部最优或鞍点使得模型精度较差，大大提高了对深度学习模型的优化效果，并有效的解决了深度学习模型跳出不接最优的问题，从而实现了对图像信息进行识别时，基于机器学习的识别高精度的需求，提高图像信息的识别效率。
本申请实施例提供了另一种基于机器学习的图像信息识别的优化方法，如图2所示，该方法包括以下步骤。
201、基于待训练的图像训练集中各图像信息的识别特征，构建一个主任务图像训练样本集，以及至少一个分支任务图像训练样本集。
对于本申请实施例，为了对图像信息的准确识别，并实现对预设深度学习模型的训练优化，基于各图像信息的识别特征，构建适用于预设深度学习模型的一个主任务图像训练样本集，以及至少一个分支任务图像训练样本集。其中，所述识别特征至少包括图像信息中五官特征、性别特征、年龄特征、表情特征、文字特征、数字特征之一，并且，对于主任务训练及分支任务训练的区分，主任务训练为需要对图像信息中进行分类识别的具体关键且必须特征的训练任务，例如人脸识别中，对于识别人脸年龄为主任务训练模型，分支任务训练为与主任务具有相关性的至少一个对图像信息中进行分类识别的非关键特征的训练任务。例如，若主任务训练为人脸年龄的训练任务，分支任务训练可以包括人脸颜色、性别识别的训练任务，或者在人脸识别模型训练时增加眉毛浓淡程度、鼻梁呈像的长宽比等的分支任务训练。因此，对于不同的主任务训练、分支任务训练，构建对应的训练样本。
具体的，本申请实施例中的一个实施场景中，图像信息以人脸识别为主要识别目标，因此，优选为识别出或者标记为五官特征的图像信息作为主任务图像训练样本集，将识别出或者标记为性别特征、年龄特征、表情特征、文字特征的图像信息作为分支任务图像训练样本集。
需要说明的是，对于分支任务图像训练样本集可以为一个、也可以为多个，对于分支任务图像训练样本集在针对主任务训练时可以使用，也可以不使用，例如，分支任务图像样本集包括a、b、c，对应的主任务图像训练样本集为S，当利用主任务图像训练样本集执行主任务训练处理后，预设深度学习模型陷入局部最优，则停止主任务训练，通过分支任务图像样本集a对已执行主任务训练的预设深度学习模型执行分支任务训练处理，当进行分支任务训练的预设深度学习模型跳出局部最优后，重新利用主任务图像训练样本集进行训练，当达到训练目标后，不再执行分支任务图像样本集b、c，本申请实施例不做具体限定。
202、获取主任务图像训练样本集以及至少一个分支任务图像训练样本集。
203、当执行主任务训练处理的所述预设深度学习模型处于局部最优、和/或鞍点状态，则根据与所述主任务图像训练样本集匹配的分支任务图像训练样本集对已执行主任务训练处理的所述预设深度学习模型执行分支任务训练处理。
对于本申请实施例，所述局部最优、鞍点状态用于表征预设深度学习模型的训练处于停滞状态，无法继续通过训练模型得到匹配训练目标，因此，此时需要利用分支任务图像样本集对此预设深度学习模型执行分支任务训练处理。即当主任务训练陷入局部最优，通过分支任务训练使得预设深度学习模型的权重进行更新，从而跳出局部最优。对于分支任务训练的选取和执行顺序，可以按照随机选取一个分支任务图像训练样本集训练一次，进行一定次数的分支任务训练的迭代后，进行主任务训练的执行，本申请实施例不做具体限定。
需要说明的是，主任务训练与分支任务训练的训练过程仅仅区别与训练样本集的不同，对于预设深度学习模型的训练步骤，完全相同，本申请实施例不做具体限定。
进一步地，为了实现对执行图像信息识别的机器学习模型的训练优化，步骤203之前，还包括：基于对执行主任务训练处理的所述预设深度学习模型的模型精度、损失值进行递归计算，当所述模型精度、损失值不变时，确定执行主任务训练处理的所述预设深度学习模型处于局部最优、和/或鞍点状态，所述预设深度学习模型为神经网络模型。
本申请实施例中，为了确定预设深度学习模型是否处于局部最优、鞍点状态，通过递归计算执行主任务训练处理的预设深度学习模型的模型精度、损失值，当模型精度、损失值不变，则说明预设深度学习模型处于局部最优、鞍点状态。具体的，为了实现对机器学习模型的优化目的，并结合图像信息识别场景，本申请实施例中，预设深度学习模型为神经网络模型。针对神经网络模型而言，当损失值loss或模型精度acc不变时，说明神经网络模型的一阶导数都为零，不能再继续进行梯度下降，即神经网络模型不能继续优化，处于局部最优，存在鞍点。
需要说明的是，损失函数是用来评估模型的预测值与真实值之间的差异程度，计算得到损失值。另外，损失函数也是神经网络中优化的目标函数，神经网络训练或者优化的过程就是最小化损失函数的过程，损失函数越小，说明神经网络模型的预测值就越接近真实值，准确性也就越好。由于本申请实施例中仅仅是计算神经网络模型的损失值或模型精度，以便确定是否进入分支任务，因此，本申请实施例中，对于神经网络的层数不做具体限定，可以为一层，也可以为多层，例如，对于一层神经网络，若干个输入和一个输出的感知机模型。损失函数可以为平方损失函数，对数损失函数、交叉熵损失函数等不同形式的损失函数，以计算出神经网络的损失值，本申请实施例不做具体限定。
204、当执行分支任务训练处理的所述预设深度学习模型跳出所述局部最优以及所述鞍点状态，则根据所述主任务图像训练样本集对所述预设深度学习模型重新执行主任务训练处理。
对于本申请实施例，由于处于局部最优、鞍点状态的判断为基于损失值以及模型精度，对应的，配置跳出局部最优、鞍点状态的判断依据为符合预设损失阈值，或者到达预设训练时间、迭代次数，确定为跳出局部最优、鞍点状态。因此，停止分支任务训练，通过主任务图像训练样本集对已进行分支任务训练的神经网络模型重新执行主任务训练处理。
需要说明的是，在特征不够明显的样本分类任务中，神经网络模型收敛速度慢，需要大量神经元去拟合相似的特征，使得在初始阶段的分支任务训练过程中要迭代次数尽量少，进而，还可以设定迭代次数为停止条件，例如分支任务迭代10次后继续主任务的训练。在特征比较鲜明或者样本量足够大的训练任务中，网络收敛速度足够快，不需要特别考虑权重偏移的问题，这时可以利用预设训练时间、预设损失阈值进行判断。
205、当重新执行主任务训练处理的所述预设深度学习模型的模型精度匹配预设精度阈值时，确定完成所述预设深度学习模型的训练。
对于本申请实施例，为了实现对模型训练的优化，以满足训练过程的完成需求，当重新执行主任务训练处理的神经网络模型的模型精度匹配预设精度阈值时，则确定此时的神经网络模型为完成训练处理。
进一步地，为了实现对机器学习模型的优化为可循环、可迭代的过程，从而满足对图像信息进行识别时，利用已优化的预设深度学习模型进行识别的高精度需求，本申请实施例中还包括：当重新执行主任务训练处理后，所述所述预设深度学习模型处于局部最优、和/或鞍点状态，则再次基于所述分支任务图像训练样本集重新对所述预设深度学习模型执行分支任务训练处理的步骤；当执行分支任务训练处理的所述预设深度学习模型跳出所述局部最优以及所述鞍点状态后，再次执行根据所述主任务图像训练样本集对所述预设深度学习模型重新执行主任务训练处理的步骤，直至所述预设深度学习模型的模型精度匹配预设精度阈值确定完成所述预设深度学习模型的训练。
具体的，为了避免在进行分支任务训练后重新执行主任务训练的预设深度学习模型再次陷入局部最优、鞍点状态，则当重新执行主任务训练处理的预设深度学习模型处于局部最优、鞍点状态时，再次利用分支任务图像训练样本集重新执行分支任务训练处理，当跳出局部最优后，再一次执行主任务训练，以此循环，直至对预设深度学习模型执行的主任务训练时，模型精度匹配预设精度阈值，则确定完成预设深度学习模型的训练。另外，对于再次执行分支任务训练的分支任务图像样本集的选取，可以依照预设的顺序依次进行。例如，分支任务图像样本集包括a、b、c，对应的主任务图像训练样本集为S，当利用主任务图像训练样本集执行主任务训练处理后，预设深度学习模型陷入局部最优，则停止主任务训练，通过分支任务图像样本集a对已执行主任务训练的预设深度学习模型执行分支任务训练处理，当进行分支任务训练的预设深度学习模型跳出局部最优后，重新利用主任务图像训练样本集进行训练，当再次处于局部最优、或鞍点时，停止主任务训练的执行，并利用分支任务图像样本集b对重新执行主任务训练的预设深度学习模型再次执行分支任务训练，当利用预设损失阈值、或迭代次数、训练时间确定跳出局部最优、鞍点状态时，重新执行主任务训练，以此循环，直至执行主任务训练的预设深度学习模型完成训练。
206、基于完成训练的预设深度学习模型，对待识别的图像信息进行识别处理，得到所述图像信息的识别结果。
具体的，本申请实施例中，对于完成训练的神经网络模型，适用但不限定于人脸图像信息、卡证图像信息进行分类识别，例如可以在卡证鉴伪系统中判断18版香港身份证中全息图是否存在，还可以增加相应区域图像是否高亮、颜色鲜艳程度、底色是否符合正常背景识别。
进一步地，为了提高利用主任训练、分支任务训练进行切换训练实现图像信息识别的准确性，所述方法还包括：解析得到所述图像信息的识别结果，基于解析出的识别特征对所述主任务图像训练样本集、所述分支任务图像训练样本集进行更新。
具体的，对于图像信息的识别结果中包含有对图像信息进行识别特征的分类，因此，作为对图像训练集的补充及优化，基于解析识别结果中的识别特征对主任务图像训练样本集、分支任务图像训练样本集进行更新。更新包括对相同图像信息不同识别特征的删除、合并、替换等，根据图像训练集中图像信息的个数确定具体的更新方式，本申请实施例不做具体限定。
本申请实施例提供了另一种基于机器学习的图像信息识别的优化方法。与现有技术相比，本申请实施例通过 获取主任务图像训练样本集以及至少一个分支任务图像训练样本集，所述主任务图像训练样本集与所述分支任务图像训练样本集相匹配；基于所述主任务图像训练样本集、所述分支任务图像训练样本集对预设深度学习模型执行所述主任务训练及所述分支任务训练相切换的训练处理，以使根据模型精度、损失值确定所述预设深度学习模型处于非局部最优、和/或非鞍点状态下完成训练处理；基于完成训练的预设深度学习模型，对待识别的图像信息进行识别处理，得到所述图像信息的识别结果，避免因对深度学习模型进行模型训练陷入局部最优或鞍点使得模型精度较差，大大提高了对深度学习模型的优化效果，并有效的解决了深度学习模型跳出不接最优的问题，从而实现了对图像信息进行识别时，基于机器学习的识别高精度的需求，提高图像信息的识别效率。
进一步的，作为对上述图1所示方法的实现，本申请实施例提供了一种基于机器学习的图像信息识别的优化装置，如图3所示，该装置包括：获取模块31，用于获取主任务图像训练样本集以及至少一个分支任务图像训练样本集，所述主任务图像训练样本集与所述分支任务图像训练样本集相匹配；训练模块32，用于基于所述主任务图像训练样本集、所述分支任务图像训练样本集对预设深度学习模型执行所述主任务训练及所述分支任务训练相切换的训练处理，以使根据模型精度、损失值确定所述预设深度学习模型处于非局部最优、和/或非鞍点状态下完成训练处理；识别模块33，用于基于完成训练的预设深度学习模型，对待识别的图像信息进行识别处理，得到所述图像信息的识别结果。
本申请实施例提供了一种基于机器学习的图像信息识别的优化装置。与现有技术相比，本申请实施例通过 获取主任务图像训练样本集以及至少一个分支任务图像训练样本集，所述主任务图像训练样本集与所述分支任务图像训练样本集相匹配；基于所述主任务图像训练样本集、所述分支任务图像训练样本集对预设深度学习模型执行所述主任务训练及所述分支任务训练相切换的训练处理，以使根据模型精度、损失值确定所述预设深度学习模型处于非局部最优、和/或非鞍点状态下完成训练处理；基于完成训练的预设深度学习模型，对待识别的图像信息进行识别处理，得到所述图像信息的识别结果，避免因对深度学习模型进行模型训练陷入局部最优或鞍点使得模型精度较差，大大提高了对深度学习模型的优化效果，并有效的解决了深度学习模型跳出不接最优的问题，从而实现了对图像信息进行识别时，基于机器学习的识别高精度的需求，提高图像信息的识别效率。
进一步的，作为对上述图2所示方法的实现，本申请实施例提供了另一种基于机器学习的图像信息识别的优化装置，如图4所示，该装置包括：获取模块41，用于获取主任务图像训练样本集以及至少一个分支任务图像训练样本集，所述主任务图像训练样本集与所述分支任务图像训练样本集相匹配；训练模块42，用于基于所述主任务图像训练样本集、所述分支任务图像训练样本集对预设深度学习模型执行所述主任务训练及所述分支任务训练相切换的训练处理，以使根据模型精度、损失值确定所述预设深度学习模型处于非局部最优、和/或非鞍点状态下完成训练处理；识别模块43，用于基于完成训练的预设深度学习模型，对待识别的图像信息进行识别处理，得到所述图像信息的识别结果。
进一步地，所述训练模块42包括：第一训练单元4201，用于当执行主任务训练处理的所述预设深度学习模型处于局部最优、和/或鞍点状态，则根据与所述主任务图像训练样本集匹配的分支任务图像训练样本集对已执行主任务训练处理的所述预设深度学习模型执行分支任务训练处理；第二训练单元4202，用于当执行分支任务训练处理的所述预设深度学习模型跳出所述局部最优以及所述鞍点状态，则根据所述主任务图像训练样本集对所述预设深度学习模型重新执行主任务训练处理；第一确定单元4203，用于当重新执行主任务训练处理的所述预设深度学习模型的模型精度匹配预设精度阈值时，确定完成所述预设深度学习模型的训练。
进一步地，所述第一训练单元4201，还用于当重新执行主任务训练处理后，所述所述预设深度学习模型处于局部最优、和/或鞍点状态，则再次基于所述分支任务图像训练样本集重新对所述预设深度学习模型执行分支任务训练处理的步骤；所述第二训练单元4202，还用于当执行分支任务训练处理的所述预设深度学习模型跳出所述局部最优以及所述鞍点状态后，再次执行根据所述主任务图像训练样本集对所述预设深度学习模型重新执行主任务训练处理的步骤，直至所述预设深度学习模型的模型精度匹配预设精度阈值确定完成所述预设深度学习模型的训练。
进一步地，所述第一训练单元4201，具体用于若所述分支任务图像训练样本集为多个时，基于所述分支任务图像训练样本集匹配的识别特征，以及识别特征与图像信息映射对应关系确定执行分支任务训练处理的顺序，并按照所述顺序对已执行主任务训练处理的所述预设深度学习模型执行多个分支任务训练处理。
进一步地，所述训练模块还包括：第二确定单元4204，用于基于对执行主任务训练处理的所述预设深度学习模型的模型精度、损失值进行递归计算，当所述模型精度、损失值不变时，确定执行主任务训练处理的所述预设深度学习模型处于局部最优、和/或鞍点状态，所述预设深度学习模型为神经网络模型。
进一步地，所述装置还包括：构建模块44，用于基于待训练的图像训练集中各图像信息的识别特征，构建一个主任务图像训练样本集，以及至少一个分支任务图像训练样本集，所述识别特征至少包括图像信息中五官特征、性别特征、年龄特征、表情特征、文字特征、数字特征之一。
进一步地，所述装置还包括：更新模块45，用于解析得到所述图像信息的识别结果，基于解析出的识别特征对所述主任务图像训练样本集、所述分支任务图像训练样本集进行更新。
本申请实施例提供了另一种基于机器学习的图像信息识别的优化装置。与现有技术相比，本申请实施例通过 获取主任务图像训练样本集以及至少一个分支任务图像训练样本集，所述主任务图像训练样本集与所述分支任务图像训练样本集相匹配；基于所述主任务图像训练样本集、所述分支任务图像训练样本集对预设深度学习模型执行所述主任务训练及所述分支任务训练相切换的训练处理，以使根据模型精度、损失值确定所述预设深度学习模型处于非局部最优、和/或非鞍点状态下完成训练处理；基于完成训练的预设深度学习模型，对待识别的图像信息进行识别处理，得到所述图像信息的识别结果，避免因对深度学习模型进行模型训练陷入局部最优或鞍点使得模型精度较差，大大提高了对深度学习模型的优化效果，并有效的解决了深度学习模型跳出不接最优的问题，从而实现了对图像信息进行识别时，基于机器学习的识别高精度的需求，提高图像信息的识别效率。
根据本申请一个实施例提供了一种存储介质，所述存储介质存储有至少一可执行指令，该计算机可执行指令可执行上述任意方法实施例中的基于机器学习的图像信息识别的优化方法。
可选的，本申请涉及的存储介质可以是计算机可读存储介质，该存储介质如计算机可读存储介质可以是非易失性的，也可以是易失性的。
图5示出了根据本申请一个实施例提供的一种计算机设备的结构示意图，本申请具体实施例并不对计算机设备的具体实现做限定。
如图5所示，该计算机设备可以包括：处理器(processor)502、通信接口(Communications Interface)504、存储器(memory)506、以及通信总线508。
其中：处理器502、通信接口504、以及存储器506通过通信总线508完成相互间的通信。
通信接口504，用于与其它设备比如客户端或其它服务器等的网元通信。
处理器502，用于执行程序510，具体可以执行上述基于机器学习的图像信息识别的优化方法实施例中的相关步骤。
具体地，程序510可以包括程序代码，该程序代码包括计算机操作指令。
处理器502可能是中央处理器CPU，或者是特定集成电路ASIC（Application Specific Integrated Circuit），或者是被配置成实施本申请实施例的一个或多个集成电路。计算机设备包括的一个或多个处理器，可以是同一类型的处理器，如一个或多个CPU；也可以是不同类型的处理器，如一个或多个CPU以及一个或多个ASIC。
存储器506，用于存放程序510。存储器506可能包含高速RAM存储器，也可能还包括非易失性存储器（non-volatile memory），例如至少一个磁盘存储器。
程序510具体可以用于使得处理器502执行以下操作：获取主任务图像训练样本集以及至少一个分支任务图像训练样本集，所述主任务图像训练样本集与所述分支任务图像训练样本集相匹配；基于所述主任务图像训练样本集、所述分支任务图像训练样本集对预设深度学习模型执行主任务训练及分支任务训练相切换的训练处理；基于完成训练的预设深度学习模型，对待识别的图像信息进行识别处理，得到所述图像信息的识别结果。
显然，本领域的技术人员应该明白，上述的本申请的各模块或各步骤可以用通用的计算装置来实现，它们可以集中在单个的计算装置上，或者分布在多个计算装置所组成的网络上，可选地，它们可以用计算装置可执行的程序代码来实现，从而，可以将它们存储在存储装置中由计算装置来执行，并且在某些情况下，可以以不同于此处的顺序执行所示出或描述的步骤，或者将它们分别制作成各个集成电路模块，或者将它们中的多个模块或步骤制作成单个集成电路模块来实现。这样，本申请不限制于任何特定的硬件和软件结合。
以上所述仅为本申请的优选实施例而已，并不用于限制本申请，对于本领域的技术人员来说，本申请可以有各种更改和变化。凡在本申请的精神和原则之内，所作的任何修改、等同替换、改进等，均应包括在本申请的保护范围之内。

Claims (20)

  1. 一种基于机器学习的图像信息识别的优化方法，包括：
    获取主任务图像训练样本集以及至少一个分支任务图像训练样本集，所述主任务图像训练样本集与所述分支任务图像训练样本集相匹配；
    基于所述主任务图像训练样本集、所述分支任务图像训练样本集对预设深度学习模型执行主任务训练及分支任务训练相切换的训练处理，以使根据模型精度、损失值确定所述预设深度学习模型处于非局部最优、和/或非鞍点状态下完成训练处理；
    基于完成训练的预设深度学习模型，对待识别的图像信息进行识别处理，得到所述图像信息的识别结果。
  2. 根据权利要求1所述的方法，其中，所述基于所述主任务图像训练样本集、所述分支任务图像训练样本集对预设深度学习模型执行主任务训练及分支任务训练相切换的训练处理包括：
    当执行主任务训练处理的所述预设深度学习模型处于局部最优、和/或鞍点状态，则根据与所述主任务图像训练样本集匹配的分支任务图像训练样本集对已执行主任务训练处理的所述预设深度学习模型执行分支任务训练处理；
    当执行分支任务训练处理的所述预设深度学习模型跳出所述局部最优以及所述鞍点状态，则根据所述主任务图像训练样本集对所述预设深度学习模型重新执行主任务训练处理；
    当重新执行主任务训练处理的所述预设深度学习模型的模型精度匹配预设精度阈值时，确定完成所述预设深度学习模型的训练。
  3. 根据权利要求2所述的方法，其中，所述方法还包括：
    当重新执行主任务训练处理后，所述所述预设深度学习模型处于局部最优、和/或鞍点状态，则再次基于所述分支任务图像训练样本集重新对所述预设深度学习模型执行分支任务训练处理的步骤；
    当执行分支任务训练处理的所述预设深度学习模型跳出所述局部最优以及所述鞍点状态后，再次执行根据所述主任务图像训练样本集对所述预设深度学习模型重新执行主任务训练处理的步骤，直至所述预设深度学习模型的模型精度匹配预设精度阈值确定完成所述预设深度学习模型的训练。
  4. 根据权利要求2所述的方法，其中，所述根据与所述主任务图像训练样本集匹配的分支任务图像训练样本集对已执行主任务训练处理的所述预设深度学习模型执行分支任务训练处理包括：
    若所述分支任务图像训练样本集为多个时，基于所述分支任务图像训练样本集匹配的识别特征，以及识别特征与图像信息映射对应关系确定执行分支任务训练处理的顺序，并按照所述顺序对已执行主任务训练处理的所述预设深度学习模型执行多个分支任务训练处理。
  5. 根据权利要求2所述的方法，其中，所述方法还包括：
    基于对执行主任务训练处理的所述预设深度学习模型的模型精度、损失值进行递归计算，当所述模型精度、损失值不变时，确定执行主任务训练处理的所述预设深度学习模型处于局部最优、和/或鞍点状态，所述预设深度学习模型为神经网络模型。
  6. 根据权利要求1-5任一项所述的方法，其中，所述方法还包括：
    基于待训练的图像训练集中各图像信息的识别特征，构建一个主任务图像训练样本集，以及至少一个分支任务图像训练样本集，所述识别特征至少包括图像信息中五官特征、性别特征、年龄特征、表情特征、文字特征、数字特征之一。
  7. 根据权利要求6所述的方法，其中，所述方法还包括：
    解析得到所述图像信息的识别结果，基于解析出的识别特征对所述主任务图像训练样本集、所述分支任务图像训练样本集进行更新。
  8. 一种基于机器学习的图像信息识别的优化装置，包括：
    获取模块，用于获取主任务图像训练样本集以及至少一个分支任务图像训练样本集，所述主任务图像训练样本集与所述分支任务图像训练样本集相匹配；
    训练模块，用于基于所述主任务图像训练样本集、所述分支任务图像训练样本集对预设深度学习模型执行主任务训练及分支任务训练相切换的训练处理，以使根据模型精度、损失值确定所述预设深度学习模型处于非局部最优、和/或非鞍点状态下完成训练处理；
    识别模块，用于基于完成训练的预设深度学习模型，对待识别的图像信息进行识别处理，得到所述图像信息的识别结果。
  9. 一种存储介质，所述存储介质中存储有至少一可执行指令，其中，所述可执行指令使处理器执行以下方法：
    获取主任务图像训练样本集以及至少一个分支任务图像训练样本集，所述主任务图像训练样本集与所述分支任务图像训练样本集相匹配；
    基于所述主任务图像训练样本集、所述分支任务图像训练样本集对预设深度学习模型执行主任务训练及分支任务训练相切换的训练处理，以使根据模型精度、损失值确定所述预设深度学习模型处于非局部最优、和/或非鞍点状态下完成训练处理；
    基于完成训练的预设深度学习模型，对待识别的图像信息进行识别处理，得到所述图像信息的识别结果。
  10. 根据权利要求9所述的存储介质，其中，执行所述基于所述主任务图像训练样本集、所述分支任务图像训练样本集对预设深度学习模型执行主任务训练及分支任务训练相切换的训练处理包括：
    当执行主任务训练处理的所述预设深度学习模型处于局部最优、和/或鞍点状态，则根据与所述主任务图像训练样本集匹配的分支任务图像训练样本集对已执行主任务训练处理的所述预设深度学习模型执行分支任务训练处理；
    当执行分支任务训练处理的所述预设深度学习模型跳出所述局部最优以及所述鞍点状态，则根据所述主任务图像训练样本集对所述预设深度学习模型重新执行主任务训练处理；
    当重新执行主任务训练处理的所述预设深度学习模型的模型精度匹配预设精度阈值时，确定完成所述预设深度学习模型的训练。
  11. 根据权利要求10所述的存储介质，其中，所述可执行指令还使处理器执行：
    当重新执行主任务训练处理后，所述所述预设深度学习模型处于局部最优、和/或鞍点状态，则再次基于所述分支任务图像训练样本集重新对所述预设深度学习模型执行分支任务训练处理的步骤；
    当执行分支任务训练处理的所述预设深度学习模型跳出所述局部最优以及所述鞍点状态后，再次执行根据所述主任务图像训练样本集对所述预设深度学习模型重新执行主任务训练处理的步骤，直至所述预设深度学习模型的模型精度匹配预设精度阈值确定完成所述预设深度学习模型的训练。
  12. 根据权利要求10所述的存储介质，其中，所述可执行指令还使处理器执行：
    基于对执行主任务训练处理的所述预设深度学习模型的模型精度、损失值进行递归计算，当所述模型精度、损失值不变时，确定执行主任务训练处理的所述预设深度学习模型处于局部最优、和/或鞍点状态，所述预设深度学习模型为神经网络模型。
  13. 根据权利要求9-12任一项所述的存储介质，其中，所述可执行指令还使处理器执行：
    基于待训练的图像训练集中各图像信息的识别特征，构建一个主任务图像训练样本集，以及至少一个分支任务图像训练样本集，所述识别特征至少包括图像信息中五官特征、性别特征、年龄特征、表情特征、文字特征、数字特征之一。
  14. 根据权利要求13所述的存储介质，其中，所述可执行指令还使处理器执行：
    解析得到所述图像信息的识别结果，基于解析出的识别特征对所述主任务图像训练样本集、所述分支任务图像训练样本集进行更新。
  15. 一种计算机设备，包括：处理器、存储器、通信接口和通信总线，所述处理器、所述存储器和所述通信接口通过所述通信总线完成相互间的通信，
    所述存储器用于存放至少一可执行指令，其中，所述可执行指令使所述处理器执行以下方法：
    获取主任务图像训练样本集以及至少一个分支任务图像训练样本集，所述主任务图像训练样本集与所述分支任务图像训练样本集相匹配；
    基于所述主任务图像训练样本集、所述分支任务图像训练样本集对预设深度学习模型执行主任务训练及分支任务训练相切换的训练处理，以使根据模型精度、损失值确定所述预设深度学习模型处于非局部最优、和/或非鞍点状态下完成训练处理；
    基于完成训练的预设深度学习模型，对待识别的图像信息进行识别处理，得到所述图像信息的识别结果。
  16. 根据权利要求15所述的计算机设备，其中，执行所述基于所述主任务图像训练样本集、所述分支任务图像训练样本集对预设深度学习模型执行主任务训练及分支任务训练相切换的训练处理包括：
    当执行主任务训练处理的所述预设深度学习模型处于局部最优、和/或鞍点状态，则根据与所述主任务图像训练样本集匹配的分支任务图像训练样本集对已执行主任务训练处理的所述预设深度学习模型执行分支任务训练处理；
    当执行分支任务训练处理的所述预设深度学习模型跳出所述局部最优以及所述鞍点状态，则根据所述主任务图像训练样本集对所述预设深度学习模型重新执行主任务训练处理；
    当重新执行主任务训练处理的所述预设深度学习模型的模型精度匹配预设精度阈值时，确定完成所述预设深度学习模型的训练。
  17. 根据权利要求16所述的计算机设备，其中，所述可执行指令还使处理器执行：
    当重新执行主任务训练处理后，所述所述预设深度学习模型处于局部最优、和/或鞍点状态，则再次基于所述分支任务图像训练样本集重新对所述预设深度学习模型执行分支任务训练处理的步骤；
    当执行分支任务训练处理的所述预设深度学习模型跳出所述局部最优以及所述鞍点状态后，再次执行根据所述主任务图像训练样本集对所述预设深度学习模型重新执行主任务训练处理的步骤，直至所述预设深度学习模型的模型精度匹配预设精度阈值确定完成所述预设深度学习模型的训练。
  18. 根据权利要求16所述的计算机设备，其中，所述可执行指令还使处理器执行：
    基于对执行主任务训练处理的所述预设深度学习模型的模型精度、损失值进行递归计算，当所述模型精度、损失值不变时，确定执行主任务训练处理的所述预设深度学习模型处于局部最优、和/或鞍点状态，所述预设深度学习模型为神经网络模型。
  19. 根据权利要求15-18任一项所述的计算机设备，其中，所述可执行指令还使处理器执行：
    基于待训练的图像训练集中各图像信息的识别特征，构建一个主任务图像训练样本集，以及至少一个分支任务图像训练样本集，所述识别特征至少包括图像信息中五官特征、性别特征、年龄特征、表情特征、文字特征、数字特征之一。
  20. 根据权利要求19所述的计算机设备，其中，所述可执行指令还使处理器执行：
    解析得到所述图像信息的识别结果，基于解析出的识别特征对所述主任务图像训练样本集、所述分支任务图像训练样本集进行更新。
PCT/CN2021/083875 2020-11-02 2021-03-30 基于机器学习的图像信息识别的优化方法及装置 WO2021180243A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202011201765.3 2020-11-02
CN202011201765.3A CN112308149B (zh) 2020-11-02 2020-11-02 基于机器学习的图像信息识别的优化方法及装置

Publications (1)

Publication Number Publication Date
WO2021180243A1 true WO2021180243A1 (zh) 2021-09-16

Family

ID=74334256

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/083875 WO2021180243A1 (zh) 2020-11-02 2021-03-30 基于机器学习的图像信息识别的优化方法及装置

Country Status (2)

Country Link
CN (1) CN112308149B (zh)
WO (1) WO2021180243A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116453201A (zh) * 2023-06-19 2023-07-18 南昌大学 基于相邻边缘损失的人脸识别方法及系统

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112308149B (zh) * 2020-11-02 2023-10-24 平安科技（深圳）有限公司 基于机器学习的图像信息识别的优化方法及装置

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190205643A1 (en) * 2017-12-29 2019-07-04 RetailNext, Inc. Simultaneous Object Localization And Attribute Classification Using Multitask Deep Neural Networks
CN111091109A (zh) * 2019-12-24 2020-05-01 厦门瑞为信息技术有限公司 基于人脸图像进行年龄和性别预测的方法、系统和设备
CN111178432A (zh) * 2019-12-30 2020-05-19 武汉科技大学 多分支神经网络模型的弱监督细粒度图像分类方法
CN111506755A (zh) * 2020-04-22 2020-08-07 广东博智林机器人有限公司 图片集的分类方法和装置
CN112308149A (zh) * 2020-11-02 2021-02-02 平安科技（深圳）有限公司 基于机器学习的图像信息识别的优化方法及装置

Family Cites Families (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP3465537A2 (en) * 2016-05-25 2019-04-10 Metail Limited Method and system for predicting garment attributes using deep learning
CN106503669B (zh) * 2016-11-02 2019-12-10 重庆中科云丛科技有限公司 一种基于多任务深度学习网络的训练、识别方法及系统
CN109919209B (zh) * 2019-02-26 2020-06-19 中国人民解放军军事科学院国防科技创新研究院 一种领域自适应深度学习方法及可读存储介质
CN111128236B (zh) * 2019-12-17 2022-05-03 电子科技大学 一种基于辅助分类深度神经网络的主乐器识别方法
CN111476369A (zh) * 2020-05-11 2020-07-31 哈尔滨工程大学 一种神经网络模型的训练方法
CN111666873A (zh) * 2020-06-05 2020-09-15 汪金玲 一种基于多任务深度学习网络的训练方法、识别方法及系统
CN111862067B (zh) * 2020-07-28 2021-10-26 中山佳维电子有限公司 一种焊接缺陷检测方法、装置、电子设备以及存储介质

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190205643A1 (en) * 2017-12-29 2019-07-04 RetailNext, Inc. Simultaneous Object Localization And Attribute Classification Using Multitask Deep Neural Networks
CN111091109A (zh) * 2019-12-24 2020-05-01 厦门瑞为信息技术有限公司 基于人脸图像进行年龄和性别预测的方法、系统和设备
CN111178432A (zh) * 2019-12-30 2020-05-19 武汉科技大学 多分支神经网络模型的弱监督细粒度图像分类方法
CN111506755A (zh) * 2020-04-22 2020-08-07 广东博智林机器人有限公司 图片集的分类方法和装置
CN112308149A (zh) * 2020-11-02 2021-02-02 平安科技（深圳）有限公司 基于机器学习的图像信息识别的优化方法及装置

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116453201A (zh) * 2023-06-19 2023-07-18 南昌大学 基于相邻边缘损失的人脸识别方法及系统
CN116453201B (zh) * 2023-06-19 2023-09-01 南昌大学 基于相邻边缘损失的人脸识别方法及系统

Also Published As

Publication number Publication date
CN112308149B (zh) 2023-10-24
CN112308149A (zh) 2021-02-02

Similar Documents

Publication Publication Date Title
WO2020140386A1 (zh) 基于TextCNN知识抽取方法、装置、计算机设备及存储介质
CN112115267A (zh) 文本分类模型的训练方法、装置、设备及存储介质
WO2020232840A1 (zh) 基于神经网络结构搜索的车辆多属性识别方法、装置、介质
KR102167011B1 (ko) 샘플링 및 적응적으로 변경되는 임계치에 기초하여 뉴럴 네트워크를 학습하는데 이용되는 하드 네거티브 샘플을 추출하는 영상 학습 장치 및 상기 장치가 수행하는 방법
WO2021180243A1 (zh) 基于机器学习的图像信息识别的优化方法及装置
CN113128478B (zh) 模型训练方法、行人分析方法、装置、设备及存储介质
WO2021212681A1 (zh) 语义角色标注方法、装置、计算机设备及存储介质
CN113343677B (zh) 一种意图识别方法、装置、电子设备及存储介质
KR20200110064A (ko) 변환 모델을 이용한 인증 방법 및 장치
CN114896395A (zh) 语言模型微调方法、文本分类方法、装置及设备
CN111738017A (zh) 一种意图识别方法、装置、设备及存储介质
CN113705092B (zh) 基于机器学习的疾病预测方法及装置
He et al. On-device deep multi-task inference via multi-task zipping
CN117218667B (zh) 一种基于字根的中文文字识别方法及系统
CN114492601A (zh) 资源分类模型的训练方法、装置、电子设备及存储介质
CN113469338A (zh) 模型训练方法、模型训练装置、终端设备及存储介质
CN108573275B (zh) 一种在线分类微服务的构建方法
CN117648950A (zh) 神经网络模型的训练方法、装置、电子设备及存储介质
CN111738455B (zh) 一种基于集成域自适应的故障诊断方法及系统
CN116029261A (zh) 中文文本语法纠错方法及相关设备
KR20210044003A (ko) 단어 임베딩 방법 및 장치와 단어 검색 방법
CN111507195A (zh) 虹膜分割神经网络模型的训练方法、虹膜分割方法及装置
CN111695594B (zh) 图像类别识别方法、装置、计算机设备及介质
CN115292455B (zh) 图文匹配模型的训练方法及装置
CN114723050B (zh) 预训练模型提示向量的确定方法、装置及电子设备

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: 21768900

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: 21768900

Country of ref document: EP

Kind code of ref document: A1