WO2022062968A1 - 一种自训练方法、系统、装置、电子设备及存储介质 - Google Patents

一种自训练方法、系统、装置、电子设备及存储介质 Download PDF

Info

Publication number
WO2022062968A1
WO2022062968A1 PCT/CN2021/118262 CN2021118262W WO2022062968A1 WO 2022062968 A1 WO2022062968 A1 WO 2022062968A1 CN 2021118262 W CN2021118262 W CN 2021118262W WO 2022062968 A1 WO2022062968 A1 WO 2022062968A1
Authority
WO
WIPO (PCT)
Prior art keywords
object detection
data
detection model
model
accuracy
Prior art date
Application number
PCT/CN2021/118262
Other languages
English (en)
French (fr)
Inventor
董茂飞
Original Assignee
杭州海康威视数字技术股份有限公司
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 杭州海康威视数字技术股份有限公司 filed Critical 杭州海康威视数字技术股份有限公司
Priority to EP21871339.4A priority Critical patent/EP4220479A4/en
Publication of WO2022062968A1 publication Critical patent/WO2022062968A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/774—Generating sets of 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
    • G06F18/21—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
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/40—Information retrieval; Database structures therefor; File system structures therefor of multimedia data, e.g. slideshows comprising image and additional audio data
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/768—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using context analysis, e.g. recognition aided by known co-occurring patterns
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/776—Validation; Performance evaluation

Definitions

  • the present application relates to the technical field of artificial intelligence, and in particular, to a self-training method, system, device, electronic device and storage medium.
  • object detection models can be used to detect objects in data such as images and audios.
  • the input data of object detection models are mainly concentrated in the data of objects containing certain properties.
  • the attributes of the object include trucks, trucks, commercial vehicles, cars, etc.
  • the sample data collected can be images of trucks, trucks, commercial vehicles, and cars. If the vehicle detection model obtained by training is applied in an urban traffic road scene, the input data of the above vehicle detection model are mainly images of commercial vehicles and cars.
  • the sample data used in training the above object detection model is a variety of data including objects with different attributes
  • the data of objects including specific attributes is relatively less targeted.
  • the input data of the object detection model is generally data of objects containing specific attributes. Therefore, when the above object detection model is used for object detection, the accuracy of the detection results obtained is low.
  • the purpose of the embodiments of the present application is to provide a self-training method, system, device, electronic device and storage medium, so as to improve the accuracy of object detection by an object detection model.
  • the specific technical solutions are as follows:
  • an embodiment of the present application provides a self-training method, the method comprising:
  • data including objects with target attributes is selected from each multimedia data as first sample data, wherein the target attributes are: the first object detection When the model detects the object in the multimedia data, the accuracy of the detection result satisfies the attributes of the object in the multimedia data with preset self-training conditions;
  • the first object detection model is trained by using the first sample data to obtain a second object detection model.
  • the method further includes:
  • test data includes existing test data and/or new test data
  • the existing test data is: Used data for testing the accuracy of object detection performed by the first object detection model
  • new test data is: unused test data
  • test result it is determined whether to update the first object detection model in the data acquisition device to the second object detection model.
  • the new test data includes at least one of the following data:
  • the data obtained by marking the candidate data by the user, wherein the candidate data is: data extracted from the multimedia data collected by the data collection device.
  • the accuracy of the object detection performed by the second object detection model is tested based on the test data, and a test result is obtained, including:
  • the method further includes:
  • the format of the second object detection model after quantization is converted into a format supported by the data acquisition device, and the The first object detection model in the data acquisition device is updated to a format-converted second object detection model.
  • the method further includes:
  • the training of the first object detection model by using the first sample data to obtain a second object detection model includes:
  • the first object detection model is trained to obtain a second object detection model, wherein the second sample data is: multimedia data input by the user through the client .
  • the target attribute is obtained in the following manner:
  • candidate attributes are: attribute information of objects whose confidence is lower than a preset confidence threshold in the result of object detection performed on each multimedia data by the first object detection model;
  • a candidate attribute whose occurrence frequency reaches a preset frequency condition is determined as a target attribute.
  • an embodiment of the present application provides a self-training system, the system includes a self-training server and a data acquisition device, wherein:
  • the data acquisition device is used to collect multimedia data, detect attribute information of objects in each multimedia data based on the first object detection model, and send the collected multimedia data and the attribute information of objects in each multimedia data to the self-training server , wherein the first object detection model is: the model currently used by the data collection device;
  • the self-training server is used to receive the multimedia data sent by the data acquisition device and the attribute information of the objects in each multimedia data, and based on the obtained attribute information of the objects in each
  • the data of the object of the target attribute is used as the first sample data
  • the first object detection model is trained by using the first sample data to obtain the second object detection model, wherein the target attribute is: the When the first object detection model detects the object in the multimedia data, the accuracy of the detection result satisfies the attributes of the object in the multimedia data under the preset self-training condition.
  • an embodiment of the present application provides a self-training device, the device comprising:
  • a data acquisition module configured to acquire multimedia data collected by a data collection device, and obtain attribute information obtained by detecting objects in each multimedia data by a first object detection model, wherein the first object detection model is: the data collection the model currently used by the device;
  • the data selection module is used to select, based on the obtained attribute information of the object in each multimedia data, the data containing the object with the target attribute from each multimedia data, as the first sample data, wherein the target attribute is:
  • the first object detection model detects an object in the multimedia data, the accuracy of the detection result satisfies the attributes of the object in the multimedia data with preset self-training conditions;
  • a model training module configured to use the first sample data to train the first object detection model to obtain a second object detection model.
  • the device further includes:
  • An accuracy test module configured to test the accuracy of the object detection performed by the second object detection model based on the test data, and obtain a test result, wherein the test data includes existing test data and/or new test data, the The existing test data is: used data for testing the accuracy of object detection performed by the first object detection model, and the new test data is: unused test data;
  • a result determination module configured to determine whether to update the first object detection model in the data acquisition device to the second object detection model according to the test result.
  • the new test data includes at least one of the following data:
  • the accuracy test module is specifically used for:
  • the device further includes a first model update module, which is specifically used for:
  • the format of the second object detection model after quantization is converted into a format supported by the data acquisition device, and the The first object detection model in the data acquisition device is updated to a format-converted second object detection model.
  • the device further includes a second model updating module, which is specifically used for:
  • model training module is specifically used for:
  • the first object detection model is trained to obtain a second object detection model, wherein the second sample data is: multimedia data input by the user through the client .
  • the apparatus further includes a target attribute determination module, which is specifically configured to obtain the target attribute in the following manner:
  • candidate attributes are: attribute information of objects whose confidence is lower than a preset confidence threshold in the result of object detection performed on each multimedia data by the first object detection model;
  • a candidate attribute whose occurrence frequency reaches a preset frequency condition is determined as a target attribute.
  • an embodiment of the present application provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus;
  • the processor is configured to implement any one of the method steps of the first aspect when executing the program stored in the memory.
  • an embodiment of the present application provides a computer-readable storage medium, where a computer program is stored in the computer-readable storage medium, and when the computer program is executed by a processor, any one of the methods described in the first aspect is implemented step.
  • the embodiments of the present application also provide a computer program product including instructions, which, when run on a computer, enables the computer to execute any one of the self-training methods described above.
  • multimedia data collected by the data acquisition device is first obtained, and attribute information obtained by the first object detection model for detecting objects in each multimedia data is obtained, wherein the first The object detection model is: the model currently used by the data acquisition device, based on the obtained attribute information of the object in each multimedia data, selects the data containing the object with the target attribute from each multimedia data, as the first sample data, wherein , the target attribute is: when the first object detection model detects the object in the multimedia data, the accuracy of the detection result satisfies the preset self-training condition The attribute of the object in the multimedia data, the first sample data is used to detect the first The object detection model is trained to obtain a second object detection model.
  • the first object detection model can be trained according to the multimedia data collected by the data collection device. Since the object in the first sample data has a target attribute, and the target attribute can be understood as the attribute of the object in the multimedia data whose detection result is less accurate when the first object detection model detects the object in the multimedia data, That is to say, the accuracy of the first object detection model for object detection on multimedia data with target attributes is relatively low. In this way, using the first sample data to train the first object detection model can improve the ability of the first object detection model to detect objects with target attributes. Attributes of multimedia data for object detection accuracy. It can be seen that, by applying the solutions provided by the embodiments of the present application, the accuracy of object detection by the object detection model can be improved.
  • FIG. 1 is a schematic flowchart of a self-training method provided by an embodiment of the present application.
  • FIG. 2 is a schematic flowchart of another self-training method provided by an embodiment of the present application.
  • FIG. 3 is a schematic flowchart of a model update judgment method provided by an embodiment of the present application.
  • FIG. 4 is a schematic flowchart of a model testing method provided by an embodiment of the present application.
  • FIG. 5 is a schematic diagram of a self-training startup process provided by an embodiment of the present application.
  • FIG. 6 is a schematic diagram of a process of querying a self-training state provided by an embodiment of the present application
  • FIG. 7 is a schematic structural diagram of a first self-training system provided by an embodiment of the present application.
  • FIG. 8 is a schematic diagram of a first self-training process provided by an embodiment of the present application.
  • FIG. 9 is a schematic structural diagram of a second self-training system provided by an embodiment of the present application.
  • FIG. 10 is a schematic structural diagram between a client and a data collection device according to an embodiment of the present application.
  • FIG. 11 is a schematic diagram of a second self-training process provided by an embodiment of the present application.
  • FIG. 12 is a schematic structural diagram of a third self-training system provided by an embodiment of the present application.
  • FIG. 13 is a schematic structural diagram of a self-training device provided by an embodiment of the application.
  • FIG. 14 is a schematic structural diagram of an electronic device according to an embodiment of the present application.
  • the embodiments of the present application provide a self-training method, system, device, electronic device, and storage medium, which will be described in detail below.
  • FIG. 1 is a schematic flowchart of a self-training method provided by an embodiment of the present application.
  • the method can be applied to electronic equipment, the electronic equipment can be a data acquisition equipment with object detection function, an object detection model can be deployed in the data acquisition equipment, and the deployed object detection model can be directly used to analyze the collected multimedia data. object to be detected.
  • the above-mentioned data collection device may refer to a device that collects multimedia data.
  • the above-mentioned data collection device may be a device used to collect images, such as IPC (IP Camera, network camera), etc., or a mobile phone with an image collection function, Tablet computer, etc.; the above data acquisition device can also be a device used to collect audio, such as a pickup, or a mobile phone, tablet computer, etc. with an audio acquisition function.
  • the above-mentioned data acquisition device can also be a storage device that indirectly obtains multimedia data, such as NVR (Network Video Recorder, network hard disk video recorder), DVR (Digital Video Recorder, digital video recorder), etc.
  • the multimedia data is sent to the storage device, and the storage device obtains the multimedia data.
  • the above electronic device may also be a server or the like.
  • the above self-training method includes the following steps 101 to 103 .
  • Step 101 Obtain multimedia data collected by a data collection device, and obtain attribute information obtained by detecting objects in each multimedia data by a first object detection model.
  • the first object detection model is: the model currently used by the data acquisition device.
  • the above-mentioned multimedia data may be images, videos, audios, and the like.
  • the first object detection model may be a model for detecting objects in the video or image.
  • the first object detection model may be a vehicle A detection model, a vehicle classification model, etc.; in the case that the above-mentioned object is a human face, the first object detection model may be a face recognition model; in the case that the multimedia data is audio, the first object detection model may be used for audio.
  • a model for object detection for example, in the case that the object is a speech segment, the first object detection model may be a speech recognition model or the like.
  • the above-mentioned objects are objects to be detected by the first object detection model, and the attribute information is information describing the attributes of the objects in the multimedia data.
  • the attribute information of the face may include: the position of the face in the image, whether the face wears accessories, such as masks, Hats, glasses, etc.
  • the first object detection model is a vehicle detection model
  • the above-mentioned object is a vehicle
  • the attribute information of the vehicle may include: the position of the vehicle in the image, the color of the vehicle, and the size of the vehicle.
  • the data collection device can collect multimedia data, and the electronic device can obtain the multimedia data, and then use the first object detection model to detect objects in the multimedia data.
  • process information and detection result information during the detection can be obtained, and the attributes used to describe the objects in the multimedia data can be selected from the above process information and detection result information. information as attribute information.
  • all multimedia data collected by the data collection device can be obtained, and multimedia data can also be extracted from the data collected by the data collection device.
  • the multimedia data can be extracted according to a preset time interval, or the multimedia data can be randomly extracted. data, etc., which are not limited in the embodiments of the present application.
  • Step 102 based on the obtained attribute information of the object in each multimedia data, data including the object with the target attribute is selected from each multimedia data as the first sample data.
  • the target attribute is: when the first object detection model detects the object in the multimedia data, the accuracy of the detection result satisfies the attribute of the object in the multimedia data of the preset self-training condition.
  • the target attribute can be understood as: the attribute of the object in the multimedia data with low accuracy detected by the first object detection model, that is, when the first object detection model performs object detection on the multimedia data containing the object with the target attribute, The accuracy of the obtained detection results is low. For example, assuming that the vehicle detection model has low object detection accuracy for images with the vehicle located in the upper right corner of the image, the vehicle location in the upper right corner of the image can be used as the target attribute.
  • the above-mentioned target attribute may be an attribute of an object preset by a user.
  • the above-mentioned self-training conditions may be: the accuracy of the detection result is lower than the preset accuracy threshold; it may also be: the frequency of occurrence of the attributes of the object in the multimedia data whose accuracy of the detection result is lower than the accuracy threshold reaches the predetermined threshold.
  • the set frequency threshold it can also be: the frequency of occurrence of the attribute of the object in the multimedia data whose accuracy of the detection result is lower than the accuracy threshold, accounts for the ratio of the frequency of occurrence of the attribute of the object in all multimedia data, and reaches the predetermined frequency. ratio thresholds, etc.
  • the accuracy of the above detection result may be the confidence level of the detection result output by the first object detection model, or may be the accuracy of the detection result calculated by using a preset evaluation algorithm, or the user may score the detection result, according to the Scoring gives the accuracy of the detection results.
  • the first object detection model detects objects in the multimedia data in advance
  • the attributes of the objects in the multimedia data whose accuracy of the detection result satisfies the above self-training conditions can be counted, and the target attributes can be determined according to the statistical results.
  • the data including the object with the target attribute can be selected from the obtained multimedia data as first sample data.
  • the image can be used as the first sample data.
  • the attribute information may further include an object score
  • the above-mentioned object score is used to represent: the probability that the content included in the multimedia data is the object.
  • its object score represents the probability that the content contained in the image is a face.
  • Step 103 using the first sample data to train the first object detection model to obtain a second object detection model.
  • the first sample data can be used to perform unsupervised training on the first object detection model to obtain the second object detection model.
  • the first object detection model can be trained by using the first sample data and the second sample data to obtain the second object detection model.
  • the second sample data is: multimedia data input by the user through the client.
  • the second sample data may be data obtained from a multimedia database, or may be data collected by other data collection devices, or the like.
  • the second sample data may be data including objects with target attributes, or may be data including objects with other attributes.
  • the above-mentioned client can be a mobile phone, an electronic computer, a tablet computer, etc., and the user can realize data interaction and instruction interaction with the electronic device through the client.
  • the multimedia data input by the user through the client may be data carrying label information, or may be data not carrying label information.
  • the first sample data may be used to perform unsupervised training on the first object detection model
  • the second sample data may be used to perform supervised training on the first object detection model
  • a second object detection model is obtained; in the case where the second sample data is data that does not carry label information, the first object detection model can be unsupervised by using the first sample data and the second sample data to obtain the second object detection model.
  • the second sample data is the data input by the user through the client
  • the user when the user inputs the second sample data through the client, the user can select more targeted multimedia data for training the first object detection model, thereby further improving the training The accuracy of the post-second object detection model.
  • the detection model is trained.
  • the selection can also be made according to the attribute information of the object in the first sample data. For example, when the attribute information of the object in the first sample data indicates that the first sample data is a vehicle image , the vehicle detection model can be selected as the first object detection model, and when the attribute information of the object of the first sample data indicates that the first sample data is a face image, the face recognition model can be selected as the first object detection model .
  • the multimedia data collected by the data acquisition device is first obtained, and the attribute information obtained by the first object detection model for detecting the objects in each multimedia data is obtained, wherein the first object
  • the detection model is: the model currently used by the data collection device, based on the obtained attribute information of the object in each multimedia data, selects the data containing the object with the target attribute from each multimedia data, as the first sample data, wherein,
  • the target attribute is: when the first object detection model detects the object in the multimedia data, the accuracy of the detection result meets the attribute of the object in the multimedia data of the preset self-training condition, and the first sample data is used to detect the first object.
  • the detection model is trained to obtain a second object detection model.
  • the first object detection model can be trained according to the multimedia data collected by the data collection device. Since the object in the first sample data has a target attribute, and the target attribute can be understood as the attribute of the object in the multimedia data whose detection result is less accurate when the first object detection model detects the object in the multimedia data, That is to say, the accuracy of object detection by the first object detection model for multimedia data with target attributes is low. In this way, using the first sample data to train the first object detection model can improve the accuracy of the first object detection model for objects with target attributes. Attributes of multimedia data for object detection accuracy. It can be seen that, by applying the solutions provided by the above embodiments, the accuracy of object detection by the object detection model can be improved.
  • the first sample data is selected from the multimedia data collected by the image acquisition device, and the selected first sample data is used to train the first object detection model, so that no Collecting sample data from each media data platform can reduce the difficulty of obtaining sample data.
  • the selected first sample data is data containing objects with target attributes, and the selection of the first sample data is not subject to the subjective influence of the staff, thereby improving the accuracy of the selected first sample data, thereby improving training The accuracy of the resulting object detection model.
  • the model training is performed based on the selected first sample data, there is no need to label the first sample data, which reduces the workload of labeling the sample data and improves the efficiency of model training.
  • the occurrence frequency of the candidate attribute can be obtained, and the candidate attribute whose occurrence frequency reaches a preset frequency condition is determined as the target attribute.
  • the candidate attributes are: attribute information of objects whose confidence is lower than a preset confidence threshold in the result of object detection performed by the first object detection model on each multimedia data.
  • the above candidate attributes can be understood as: in each multimedia data, the confidence of the result obtained by using the first object detection model to detect the object is lower than the preset reliability threshold of the attribute information of the object in the data.
  • the frequency condition may be that the occurrence frequency reaches a preset number of times threshold, for example, the number of times threshold may be 100 times, 500 times, 1000 times, and so on.
  • the frequency condition may also be that the ratio of the occurrence frequency of each candidate attribute to the total occurrence frequency of all candidate attributes reaches a preset ratio threshold.
  • the candidate attributes include black vehicles, gray vehicles, and green vehicles
  • the ratio threshold is 60%, where the occurrence frequency of black vehicles is 30 times, the occurrence frequency of gray vehicles is 80 times, and the occurrence frequency of green vehicles is 80 times. It can be seen that the appearance frequency of gray vehicles accounts for 61% of the total number of appearance frequencies, which reaches the above ratio threshold. Therefore, it can be considered that the appearance frequency of gray vehicles reaches the preset frequency condition, and gray vehicles can be used as the target attribute.
  • a detection result and a confidence level of the detection result can be obtained.
  • the attributes of the objects in the multimedia data whose confidence level of the detection result is lower than the confidence threshold value can be recorded as candidate attributes.
  • the candidate attributes can be used as the target. Attributes.
  • step 104 after the second object detection model is obtained by training in the above step 103, the following step 104 may be further included.
  • Step 104 based on the test data, test the accuracy of object detection by the second object detection model, obtain a test result, and determine whether to update the first object detection model in the data acquisition device to the second object detection model according to the test result.
  • the test result represents the accuracy of the object detection performed by the second object detection model.
  • the test data includes existing test data and/or new test data, and the existing test data is: used data for testing the accuracy of object detection performed by the first object detection model.
  • the above-mentioned existing test data may be pre-stored multimedia data.
  • the above existing test data may carry label information, which facilitates testing the second object detection model by using the existing test data.
  • the above-mentioned existing test data may also not carry label information, which is not limited in this embodiment of the present application.
  • the new test data is: unused test data, that is, data for which the accuracy of the first object detection model has not been tested before.
  • the new test data may be unused data in pre-obtained data, or may be obtained after obtaining a first test that indicates that the accuracy of object detection performed by the second object detection model is higher than the accuracy of object detection performed by the first object detection model. After the results, the data is obtained.
  • the above-mentioned new test data may carry label information, or may not carry label information, etc., which is not limited in this embodiment of the present application.
  • the accuracy of object detection by the second object detection model can be tested, and according to the detected accuracy, it is determined whether to update the first object detection model in the data acquisition device to the second object detection model.
  • the accuracy of the object detection performed by the second object detection model can be detected based on the test data.
  • the model is updated to the second object detection model; in the case of low accuracy, the first object detection model in the data acquisition device may not be updated, or the second object detection model may continue to be trained until the second object detection
  • the accuracy of object detection performed by the model is higher than the accuracy of object detection performed by the first object detection model.
  • the test data in order to obtain the test result, may be input into the first object detection model to perform object detection, and the first output result may be obtained, and the test data may be input into the second object detection model for object detection, and the second object detection model may be obtained.
  • Output the result calculate the first difference value of the first output result relative to the annotation information, and calculate the second difference value of the second output result relative to the annotation information, and determine the second object detection model based on the first difference value and the second difference value. Whether the accuracy of object detection is higher than the accuracy of object detection performed by the first object detection model is obtained, and a judgment result is obtained as a test result.
  • the labeling information is: information that labels the test data in advance.
  • the above-mentioned first difference value and second difference value can be obtained by calculating Euclidean distance, cosine similarity, Manhattan distance, and the like.
  • the cosine similarity of the first output result relative to the annotation information may be calculated as the first difference value.
  • the first difference value is greater than the second difference value, it means that the first output result obtained by the first object detection model detecting the objects in the test data is quite different from the labeling information, and the second object detection model is not suitable for the test.
  • the difference between the second output result obtained by detecting the object in the data and the labeling information is small, which further indicates that the accuracy of detecting the object by the first object detection model is relatively low, and the accuracy of detecting the object by the second object detection model is relatively high. Therefore, it can be judged that the accuracy of object detection performed by the second object detection model is higher than the accuracy of object detection performed by the first object detection model;
  • the first difference value is smaller than the second difference value
  • the first output result obtained by the first object detection model detecting the object in the test data has a small difference with the labeling information
  • the second object detection model detects the object in the test data.
  • the second output result obtained by the detection is quite different from the annotation information, which further indicates that the accuracy of the first object detection model to detect the object is high, and the accuracy of the second object detection model to detect the object is low, so it can be judged
  • the accuracy of object detection performed by the first object detection model is higher than the accuracy of object detection performed by the second object detection model;
  • the above judgment result is used as the test result, and then it can be determined whether to update the first object detection model in the data acquisition device to the second object detection model according to the test result.
  • only the test data may be input into the second object detection model, a third output result may be obtained, and a third difference value of the third output result relative to the annotation information may be calculated, and if the third difference value is less than a preset value
  • the difference degree threshold it is considered that the accuracy of the object detection performed by the second object detection model is relatively high, and in the case that the third difference value is not less than the preset difference degree threshold value, it is considered that the second object detection model performs object detection. less accurate.
  • the above-mentioned difference threshold can be manually set based on experience, or can be obtained by calculating the average value of the difference between the test results obtained by the first object detection model detecting the objects in the test data and the marked information, and the above-mentioned average value can be an arithmetic average. value, weighted average, etc.
  • the above-mentioned new test data may be data with label information input by a user through a client, specifically data obtained from a multimedia database, or data collected by other data collection devices.
  • the user can mark the multimedia data, and input the marked multimedia data into the electronic device through the client to obtain new test data.
  • the above-mentioned new test data may also be data obtained by marking candidate data by a user, wherein the candidate data is: data extracted from multimedia data collected by a data collection device.
  • the electronic device obtains the multimedia data collected by the data collection device in the above step 101, it can extract part of the data from the multimedia data and send it to the client.
  • the user marks the above data through the client, and the client then marks the marked data. Return to the electronics for new test data.
  • candidate data from multimedia data it can be randomly selected or extracted according to a preset data collection interval.
  • the multimedia data is an image
  • an image can be extracted every 60 frames as candidate data.
  • the extracted data may be data used for model training or data other than model training data.
  • part of the data may also be extracted from the first sample data, the extracted data may be marked by the user, and the marked data may be used as the above-mentioned new test data. It can be understood that, based on the obtained attribute information of the object in each multimedia data, the data including the object with the target attribute is selected from the respective multimedia data. Divide the selected data into two parts, one part is used as sample data for training the first object detection model; the other part of the data is used as test data after labeling, used to test the second object detection model obtained after training .
  • the test data when the test data does not carry label information, the test data can be input into the second object detection model to obtain the confidence level of the detection result output by the model, and the above confidence level can be used as the second object detection model The accuracy of object detection is performed to obtain the test result.
  • the second object detection model may also be scored according to the accuracy of the second object detection model obtained by the test, and the scoring result may be displayed to the user through the client. Based on the scoring result, the user can send an instruction through the client to indicate whether to update the first object detection model in the data collection device to the second object detection model.
  • the second object detection model when it is determined to update the first object detection model in the data acquisition device to the second object detection model, the second object detection model is quantified according to a preset quantization method, and the preset quantization method is used to quantify the second object detection model.
  • the verification data is to verify the consistency of the second object detection model after quantization relative to the second object detection model before quantization. In the case where the verification result indicates that the second object detection model after quantization is consistent with the second object detection model before quantization,
  • the format of the quantized second object detection model is converted into a format supported by the data acquisition device, and the first object detection model in the data acquisition device is updated to the format converted second object detection model.
  • the above-mentioned quantization method may be weight quantization, activation value quantization, or the like.
  • the quantization of the second object detection model can make the second object detection model more streamlined, so that memory can be saved when the second object detection model is run, and the running efficiency of the model can be improved.
  • the consistency of the second object detection model after quantization relative to the second object detection model before quantization can be checked by using the test data.
  • the verification data may be input into the second object detection model after quantization to obtain the first verification result
  • the verification data may be input into the second object detection model before quantization to obtain the second verification result
  • the first verification result may be calculated.
  • the difference value between the verification result and the second verification result, and the consistency of the second object detection model after quantization relative to the second object detection model before quantization is determined according to the above difference value.
  • the detection result obtained by the second object detection model after the quantization detects the objects in the multimedia data
  • the second object detection model before the quantization detects the objects in the multimedia data.
  • the obtained detection results have a small degree of difference, so it can be considered that the second object detection model after quantization is consistent with the second object detection model before quantization.
  • the format conversion of the second object detection model after quantization can be performed.
  • the formats of models supported by different data collection devices may be different, and the above formats include the bit width, arrangement, and carry system of the data in the model.
  • the processors deployed by different data acquisition equipment may be produced by different manufacturers, such as NVIDIA, Intel, AMD (Advanced Micro Devices, Advanced Micro Devices), etc., and the models supported by processors produced by different manufacturers are in different formats.
  • the format of the model supported by the data acquisition device can be determined first, and then the quantized second object detection model can be format-converted according to the determined format, so that the format-converted second object detection model can be run on the data acquisition device. , it is convenient to update the first object detection model in the data acquisition device to the format-converted second object detection model.
  • the first object detection model in the data acquisition device of the same type may be updated to the second object Detection model.
  • the similar data acquisition device is: a data acquisition device using the first object detection model.
  • the first object detection model is a face recognition model
  • the similar data collection device is a data collection device using the face recognition model.
  • the second object detection model can be explained.
  • the accuracy of object detection is higher than the accuracy of the first object detection model, that is, the accuracy of the second object detection model is higher than that of the first object detection model used in each similar data acquisition device, so each similar data can be collected.
  • the first object detection model in the device is updated to the second object detection model.
  • the following steps 1041 to 1043 may be included.
  • Step 1041 based on the existing test data, test the accuracy of object detection performed by the second object detection model to obtain a first test result.
  • the existing test data carries the annotation information.
  • the existing test data may be input into the first object detection model for object detection to obtain the first output result
  • the existing test data may be input into the second object detection model for object detection to obtain the second output
  • the first difference value of the first output result relative to the annotation information is calculated
  • the second difference value of the second output result relative to the annotation information is calculated, and based on the first difference value and the second difference value, it is determined that the second object detection model performs an object Whether the detection accuracy is higher than the accuracy of the object detection performed by the first object detection model is obtained, and the judgment result is obtained as the first test result.
  • the first difference value is greater than the second difference value, it means that the first output result obtained by the first object detection model detecting the objects in the existing test data is quite different from the annotation information, and the second object detection model has a great effect on the test data.
  • the difference between the second output result obtained by detecting the middle object and the labeling information is small, which further indicates that the accuracy of detecting the object by the first object detection model is relatively low, and the accuracy of detecting the object by the second object detection model is relatively high;
  • the first difference value is smaller than the second difference value, it means that the difference between the first output result obtained by the first object detection model detecting the objects in the existing test data and the label information is small, and the second object detection model has little difference between the test data and the test data.
  • the second output result obtained by detecting the middle object is quite different from the labeling information, which further indicates that the accuracy of the first object detection model for detecting the object is high, and the accuracy of the second object detection model for detecting the object is low.
  • only the existing test data may be input into the second object detection model to obtain a third output result, and a third difference value of the third output result relative to the annotation information may be calculated.
  • the set difference degree threshold it is considered that the accuracy of the object detection performed by the second object detection model is high, and when the third difference value is not less than the preset difference degree threshold, it is considered that the second object detection model performs object detection. The detection accuracy is low.
  • the above-mentioned difference threshold can be manually set based on experience, or can be obtained by calculating the average value of the difference between the test results obtained by the first object detection model detecting the objects in the existing test data and the labeling information, and the above-mentioned average value can be Arithmetic mean, weighted mean, etc.
  • Step 1042 in the case that the first test result indicates that the accuracy of the object detection performed by the second object detection model is higher than the accuracy of the object detection performed by the first object detection model, re-test the second object detection model based on the new test data. The accuracy of the object detection, the second test result is obtained.
  • the new test data can be used to perform the first test again.
  • the two-object detection model is tested to obtain a second test result.
  • the testing method is similar to the above-mentioned step 1041, which is not repeated here.
  • Step 1043 determine whether to update the first object detection model in the data acquisition device to the second object detection model.
  • the second test result indicates that the accuracy of the second object detection model for object detection is higher than the accuracy of the first object detection model for object detection
  • the second object detection model is updated, otherwise, the first object detection model in the data acquisition device is not updated.
  • the second test result can also be displayed to the user, and the user determines whether to update the first object detection model in the data acquisition device to the second object detection model, and in the case of receiving the update instruction sent by the user through the client, The first object detection model in the data acquisition device is updated to the second object detection model.
  • the accuracy of object detection by the second object detection model is firstly tested based on the existing test data, and when the first test result indicates that the accuracy of the second object detection model is high, the second object detection model is further tested based on the new test data. The accuracy of the model is tested, and then whether to update the first object detection model in the data acquisition device to the second object detection model is determined according to the second test result, which can make the test result more accurate.
  • the new test data sent by the user through the client is not obtained, it can be directly determined according to the first test result whether to update the first object detection model in the data collection device to the second object Detection model.
  • FIG. 4 is a schematic flowchart of a model testing method provided by an embodiment of the present application. The method includes the following steps 401 to 406 .
  • Step 401 using the existing test data to test the accuracy of the object detection performed by the second object detection model to obtain a first test result
  • Step 402 based on the first test result, determine whether the accuracy of the object detection performed by the second object detection model is higher than the accuracy of the object detection performed by the first object detection model, if no, go to step 403, if so, go to step 404 ;
  • Step 403 continue to train the second object detection model
  • Step 404 judge whether to obtain new test data, if yes, go to step 405, if no, go to step 406;
  • Step 405 using the new test data to test the accuracy of the object detection performed by the second object detection model to obtain a second test result
  • Step 406 displaying the test result through the client.
  • FIG. 5 is a schematic diagram of a self-training startup process provided by an embodiment of the present application.
  • the client can display a GUI (Graphical User Interface, Graphical User Interface) to the user, and the user determines to enable the self-training by operating the GUI interface.
  • GUI Graphical User Interface
  • the client obtains the instruction to enable self-training through the GUI, and then sends the instruction to enable self-training to the electronic device.
  • the electronic device may contain an application program that executes self-training. After receiving the instruction to enable self-training, it first checks whether the instruction is correct. If it is incorrect, an error code message indicating an error occurred in the instruction is generated and returned as the result information to The client, if correct, initializes the self-training function and determines whether the initialization is successful.
  • the client sends an instruction to enable self-training to the electronic device, it detects whether it receives the result information returned by the electronic device. If it is no, it waits to receive the above-mentioned result information. If it is, it determines whether the self-training is successfully enabled according to the result information. . If it is no, the user can be prompted that the self-training fails, and the user can be prompted according to the error code information in the returned result. Specifically, the user can be prompted in the form of a pop-up window. If yes, the user can be prompted to open the self-training successfully. .
  • the user can independently choose whether to enable and execute self-training through the client terminal, thereby improving the degree of freedom of enabling self-training and bringing a better experience to the user.
  • the electronic device may also enable self-training when it detects that the first object detection model used has a low accuracy in detecting objects in multimedia data containing objects with a certain attribute , and use the above attributes as target attributes, thereby improving the accuracy of the first object detection model for detecting objects in multimedia data including objects with any attribute, and improving the automation degree of self-training.
  • FIG. 6 is a schematic diagram of a process of querying a self-training state according to an embodiment of the present application.
  • the user can query the self-training status by operating the GUI interface provided by the client.
  • the client can send the instruction for querying the self-training status to the electronic device.
  • the electronic device receives the above instruction and queries the self-training status. state, the queried self-training state is used as the result information and returned to the client.
  • the client detects whether the result information is received. If it is no, it continues to wait for the result information to be received. If it is, it judges whether the self-training state has changed according to the result information. If no, the user is prompted that the self-training state has not changed. If yes, the self-training status is updated according to the result information, and it is judged whether the self-training is completed. If yes, the user is prompted that the self-training is over. Returns the query self-training status.
  • the self-training method provided by the embodiments of the present application can be applied to a self-training system.
  • the self-training method can be applied to the self-training system.
  • the above-mentioned data acquisition device may be an IPC, NVR, DVR with an analysis function, or a server with an analysis function.
  • the self-training method can also be applied to the data acquisition device, which will be described in detail below.
  • FIG. 7 is a schematic structural diagram of a first self-training system provided by an embodiment of the present application.
  • the self-training system may include a client, a self-training server and multiple data acquisition devices, wherein the client and the self-training server can be connected through a network, and the self-training server and each data acquisition device are connected can also be connected via the network.
  • FIG. 8 is a schematic diagram of a first self-training process provided by an embodiment of the present application. As shown in Figure 8:
  • the data acquisition device can obtain the multimedia data, and use the object detection model to perform object detection on the multimedia data, and obtain the attribute information of the objects in the multimedia data as the detection result;
  • the self-training server can obtain multimedia data and attribute information of objects in the multimedia data from the data acquisition device, and select data including objects with target attributes from the multimedia data as the first sample data for model training.
  • the self-training server can also select the corresponding first object detection model from the stored multiple object detection models according to the attribute information of the multimedia data;
  • the self-training server when the self-training server obtains multimedia data and attribute information, it can obtain multimedia data and attribute information from multiple similar data acquisition devices;
  • the client can also obtain the second sample data, mark the second sample data, and send the marked second sample data to the self-training server;
  • the self-training server can manage the first sample data and the second sample data, and use the first sample data and the second sample data to perform self-training on the selected first object detection model to obtain the second object detection model . Then use the test data to evaluate the accuracy of the second object detection model;
  • the self-training server can select unlabeled data from the managed sample data and send it to the client.
  • the user can mark the above data through the client to obtain new test data, and send the new test data to the self-training server.
  • the training server can use the new test data to evaluate the accuracy of the second object detection model;
  • the self-training server can continue to obtain multimedia data and attribute information from the data acquisition device to continue training the model; after evaluating the accuracy of the second object detection model
  • the model can be transformed, including the quantification and format transformation of the second object detection model, and the consistency of the second object detection model after the transformation with the second object detection model before the transformation is checked. If the verification results are inconsistent, you can continue to obtain multimedia data and attribute information from the data acquisition device to continue training the model.
  • the detection model is updated to the second object detection model.
  • the self-training server may also store the second object detection model obtained by training, which facilitates subsequent self-training of the second object detection model.
  • FIG. 9 is a schematic structural diagram of a second self-training system provided by an embodiment of the present application.
  • the self-training system may include a client and multiple data collection devices, wherein the client and multiple data collection devices may be connected through a network.
  • FIG. 10 is a schematic structural diagram between a client and a data collection device according to an embodiment of the present application.
  • the client is connected to the data acquisition device, and the data acquisition device includes a main control unit and a plurality of computing units.
  • the main control unit can be used to read data, call computing units, etc., and each computing unit can Used for model running, model training, etc.
  • the data acquisition device may further include a storage unit for storing the object detection model deployed by the data acquisition device, the collected multimedia data, the trained object detection model, and the like. In this way, the control unit can read the sample data and the object detection model from the storage unit, use the read sample data to train the read object detection model, and store the trained object detection model in the storage unit.
  • the client can control the main control unit in the data acquisition device to call the specified computing unit for model training.
  • FIG. 11 is a schematic diagram of a second self-training process provided by this embodiment of the application, as shown in FIG. 11:
  • the data acquisition device can obtain the multimedia data, and use the object detection model to perform object detection on the multimedia data, and obtain attribute information of the objects in the multimedia data as the detection result.
  • the multimedia data and the attribute information of the objects in the multimedia data are obtained from the detection result, and then the data including the object with the target attribute is selected from the multimedia data as the first sample data for model training. It is also possible to select the corresponding first object detection model from the stored multiple object detection models according to the attribute information of the multimedia data;
  • the multimedia data and attribute information when obtaining multimedia data and attribute information, can be obtained from a plurality of similar data collection devices;
  • the client can also obtain the second sample data, mark the second sample data, and send the marked second sample data to the data collection device;
  • the data acquisition device can manage the first sample data and the second sample data, and use the first sample data and the second sample data to perform self-training on the selected first object detection model to obtain a second object detection model . Then use the test data to evaluate the accuracy of the second object detection model;
  • the data acquisition device can select unmarked data from the managed sample data and send it to the client.
  • the user can mark the above data through the client to obtain new test data, and send the new test data to the data acquisition device.
  • the acquisition device can use the new test data to evaluate the accuracy of the second object detection model;
  • the data acquisition device can continue to obtain multimedia data and attribute information to continue training the model; when the accuracy of evaluating the second object detection model meets the update conditions
  • the model can be transformed, including the quantification and format transformation of the second object detection model, and the consistency of the second object detection model after the transformation relative to the second object detection model before the transformation.
  • FIG. 12 is a schematic structural diagram of a third self-training system provided by an embodiment of the present application.
  • the system includes a self-training server 1201 and a data acquisition device 1202, wherein:
  • the data collection device 1202 is used to collect multimedia data, detect attribute information of objects in each multimedia data based on the first object detection model, and send the collected multimedia data and the information of the objects in each multimedia data to the self-training server 1201. attribute information, wherein the first object detection model is: the model currently used by the data collection device 1202;
  • the self-training server 1201 is used to receive the multimedia data sent by the data acquisition device 1202 and the attribute information of the objects in each multimedia data, and based on the obtained attribute information of the objects in each multimedia data, select from each multimedia data.
  • the data containing the object with the target attribute is used as the first sample data, and the first object detection model is trained by using the first sample data to obtain the second object detection model, wherein the target attribute is:
  • the first object detection model detects an object in the multimedia data, the accuracy of the detection result satisfies the attributes of the object in the multimedia data with preset self-training conditions.
  • the self-training server 1201 is further used for:
  • test data includes existing test data and/or new test data
  • the existing test data is: Used data for testing the accuracy of object detection performed by the first object detection model
  • new test data is: unused test data
  • test result it is determined whether to update the first object detection model in the data acquisition device 1202 to the second object detection model.
  • the new test data includes at least one of the following data:
  • the self-training server 1201 is specifically used for:
  • the self-training server 1201 is further used for:
  • the format of the second object detection model after quantization is converted into a format supported by the data acquisition device 1202, and the The first object detection model in the data acquisition device 1202 is updated to the format-converted second object detection model.
  • the self-training server 1201 is further used for:
  • the second object detection model wherein the data collection device 1202 of the same type is: a data collection device 1202 using the first object detection model.
  • the self-training server 1201 is specifically used for:
  • the first object detection model is trained to obtain a second object detection model, wherein the second sample data is: multimedia data input by the user through the client .
  • the self-training server 1201 is configured to obtain the target attribute in the following manner:
  • candidate attributes are: attribute information of objects whose confidence is lower than a preset confidence threshold in the result of object detection performed on each multimedia data by the first object detection model;
  • a candidate attribute whose occurrence frequency reaches a preset frequency condition is determined as a target attribute.
  • FIG. 13 is a schematic structural diagram of a self-training device provided by an embodiment of the present application, and the device includes:
  • the data obtaining module 1301 is used to obtain multimedia data collected by a data collection device, and obtain attribute information obtained by detecting objects in each multimedia data by a first object detection model, wherein the first object detection model is: the data The model currently used by the acquisition device;
  • the data selection module 1302 is used to select, based on the obtained attribute information of the object in each multimedia data, the data containing the object with the target attribute from each multimedia data, as the first sample data, wherein the target attribute is :
  • the first object detection model detects an object in the multimedia data, the accuracy of the detection result satisfies the attributes of the object in the multimedia data with preset self-training conditions;
  • the model training module 1303 is configured to use the first sample data to train the first object detection model to obtain a second object detection model.
  • the device further includes:
  • An accuracy test module configured to test the accuracy of the object detection performed by the second object detection model based on the test data, and obtain a test result, wherein the test data includes existing test data and/or new test data, the The existing test data is: used data for testing the accuracy of object detection performed by the first object detection model, and the new test data is: unused test data;
  • a result determination module configured to determine whether to update the first object detection model in the data acquisition device to the second object detection model according to the test result.
  • the new test data includes at least one of the following data:
  • the accuracy test module is specifically used for:
  • the device further includes a first model update module, which is specifically used for:
  • the format of the second object detection model after quantization is converted into a format supported by the data acquisition device, and the The first object detection model in the data acquisition device is updated to a format-converted second object detection model.
  • the device further includes a second model updating module, which is specifically used for:
  • model training module 1303 is specifically used for:
  • the first object detection model is trained to obtain a second object detection model, wherein the second sample data is: multimedia data input by the user through the client .
  • the apparatus further includes a target attribute determination module, which is specifically configured to obtain the target attribute in the following manner:
  • candidate attributes are: attribute information of objects whose confidence is lower than a preset confidence threshold in the result of object detection performed on each multimedia data by the first object detection model;
  • a candidate attribute whose occurrence frequency reaches a preset frequency condition is determined as a target attribute.
  • the multimedia data collected by the data acquisition device is first obtained, and the attribute information obtained by the first object detection model for detecting the objects in each multimedia data is obtained, wherein the first object
  • the detection model is: the model currently used by the data collection device, based on the obtained attribute information of the object in each multimedia data, selects the data containing the object with the target attribute from each multimedia data, as the first sample data, wherein,
  • the target attribute is: when the first object detection model detects the object in the multimedia data, the accuracy of the detection result meets the attribute of the object in the multimedia data of the preset self-training condition, and the first sample data is used to detect the first object.
  • the detection model is trained to obtain a second object detection model.
  • the first object detection model can be trained according to the multimedia data collected by the data collection device. Since the object in the first sample data has a target attribute, and the target attribute can be understood as the attribute of the object in the multimedia data whose detection result is less accurate when the first object detection model detects the object in the multimedia data, That is to say, the accuracy of object detection by the first object detection model for multimedia data with target attributes is low. In this way, using the first sample data to train the first object detection model can improve the accuracy of the first object detection model for objects with target attributes. Attributes of multimedia data for object detection accuracy. It can be seen that, by applying the solutions provided by the above embodiments, the accuracy of object detection by the object detection model can be improved.
  • An embodiment of the present application also provides an electronic device, as shown in FIG. 14 , including a processor 1401 , a communication interface 1402 , a memory 1403 and a communication bus 1404 , wherein the processor 1401 , the communication interface 1402 , and the memory 1403 pass through the communication bus 1404 complete communication with each other,
  • the memory 1403 is used to store computer programs
  • the processor 1401 is configured to implement the steps of the above self-training method when executing the program stored in the memory 1403 .
  • the communication bus mentioned in the above electronic device may be a peripheral component interconnect standard (Peripheral Component Interconnect, PCI) bus or an Extended Industry Standard Architecture (Extended Industry Standard Architecture, EISA) bus or the like.
  • PCI peripheral component interconnect standard
  • EISA Extended Industry Standard Architecture
  • the communication bus can be divided into an address bus, a data bus, a control bus, and the like. For ease of presentation, only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus.
  • the communication interface is used for communication between the above electronic device and other devices.
  • the memory may include random access memory (Random Access Memory, RAM), and may also include non-volatile memory (Non-Volatile Memory, NVM), such as at least one disk storage.
  • RAM Random Access Memory
  • NVM non-Volatile Memory
  • the memory may also be at least one storage device located away from the aforementioned processor.
  • the above-mentioned processor can be a general-purpose processor, including a central processing unit (Central Processing Unit, CPU), a network processor (Network Processor, NP), etc.; it can also be a digital signal processor (Digital Signal Processing, DSP), dedicated integrated Circuit (Application Specific Integrated Circuit, ASIC), Field-Programmable Gate Array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.
  • CPU Central Processing Unit
  • NP Network Processor
  • DSP Digital Signal Processing
  • ASIC Application Specific Integrated Circuit
  • FPGA Field-Programmable Gate Array
  • FPGA Field-Programmable Gate Array
  • a computer-readable storage medium is also provided, where a computer program is stored in the computer-readable storage medium, and when the computer program is executed by a processor, any one of the above self-training methods is implemented A step of.
  • the above-mentioned embodiments it may be implemented in whole or in part by software, hardware, firmware or any combination thereof.
  • software it can be implemented in whole or in part in the form of a computer program product.
  • the computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present application are generated.
  • the computer may be a general purpose computer, special purpose computer, computer network, or other programmable device.
  • the computer instructions may be stored in or transmitted from one computer readable storage medium to another computer readable storage medium, for example, the computer instructions may be downloaded from a website site, computer, server or data center Transmission to another website site, computer, server, or data center is by wire (eg, coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (eg, infrared, wireless, microwave, etc.).
  • the computer-readable storage medium may be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. that includes an integration of one or more available media.
  • the usable media may be magnetic media (eg, floppy disks, hard disks, magnetic tapes), optical media (eg, DVD), or semiconductor media (eg, Solid State Disk (SSD)), and the like.
  • the multimedia data collected by the data acquisition device is first obtained, and the attribute information obtained by the first object detection model for detecting the objects in each multimedia data is obtained, wherein the first object
  • the detection model is: the model currently used by the data collection device, based on the obtained attribute information of the object in each multimedia data, selects the data containing the object with the target attribute from each multimedia data, as the first sample data, wherein,
  • the target attribute is: when the first object detection model detects the object in the multimedia data, the accuracy of the detection result meets the attribute of the object in the multimedia data of the preset self-training condition, and the first sample data is used to detect the first object.
  • the detection model is trained to obtain a second object detection model.
  • the first object detection model can be trained according to the multimedia data collected by the data collection device. Since the object in the first sample data has a target attribute, and the target attribute can be understood as the attribute of the object in the multimedia data whose detection result is less accurate when the first object detection model detects the object in the multimedia data, That is to say, the accuracy of the first object detection model for object detection on multimedia data with target attributes is relatively low. In this way, using the first sample data to train the first object detection model can improve the ability of the first object detection model to detect objects with target attributes. Attributes of multimedia data for object detection accuracy. It can be seen that, by applying the solutions provided by the above embodiments, the accuracy of object detection by the object detection model can be improved.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Computation (AREA)
  • Databases & Information Systems (AREA)
  • Multimedia (AREA)
  • Medical Informatics (AREA)
  • General Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Image Analysis (AREA)

Abstract

本申请实施例提供了一种自训练方法、系统、装置、电子设备及存储介质，涉及人工智能技术领域，包括：获得数据采集设备采集的多媒体数据，并获得第一对象检测模型对各个多媒体数据中对象进行检测得到的属性信息；基于所获得的每一多媒体数据中对象的属性信息，从各个多媒体数据中选择包含具有目标属性的对象的数据，作为第一样本数据，其中，目标属性为：第一对象检测模型对多媒体数据中对象进行检测时，检测结果的准确度满足预设的自训练条件的多媒体数据中对象所具有的属性；利用第一样本数据对第一对象检测模型进行训练，得到第二对象检测模型。应用本申请实施例提供的方案，可以提高对象检测模型进行对象检测的准确度。

Description

一种自训练方法、系统、装置、电子设备及存储介质
本申请要求于2020年09月24日提交中国专利局、申请号为202011018277.9发明名称为“一种自训练方法、系统、装置、电子设备及存储介质”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及人工智能技术领域，特别是涉及一种自训练方法、系统、装置、电子设备及存储介质。
背景技术
随着人工智能技术的快速发展，对象检测模型的应用越来越广泛，其中，上述对象检测模型可以用于对图像、音频等数据中的对象进行检测。
现有技术中，在获得对象检测模型时，通常需要搜集大量样本数据，利用所搜集的样本数据对初始模型进行训练，从而得到训练后的模型，作为对象检测模型。为得到能够兼顾各种应用场景的对象检测模型，在搜集样本数据时，一般需要从多个场景中分别搜集包含不同属性的对象的多种样本数据。然而在实际应用场景中，对象检测模型的输入数据主要集中在包含某些属性的对象的数据中。例如，在上述对象为车辆时，对象的属性包括货车、卡车、商务车、轿车等，在对车辆检测模型进行训练时，所搜集的样本数据可以是货车图像、卡车图像、商务车图像、轿车图像等，若训练得到的车辆检测模型应用在城市交通道路场景中，则上述车辆检测模型的输入数据主要为商务车图像、轿车图像。
由于训练上述对象检测模型时，所采用的样本数据为包含不同属性的对象的多种数据，这样对于包含特定属性的对象的数据而言，针对性相对较弱。而实际应用场景中对象检测模型的输入数据一般为包含特定属性的对象的数据，所以，应用上述对象检测模型进行对象检测时，得到的检测结果准确度较低。
发明内容
本申请实施例的目的在于提供一种自训练方法、系统、装置、电子设备及存储介质，以提高对象检测模型进行对象检测的准确度。具体技术方案如下：
第一方面，本申请实施例提供了一种自训练方法，所述方法包括：
获得数据采集设备采集的多媒体数据，并获得第一对象检测模型对各个多媒体数据中对象进行检测得到的属性信息，其中，所述第一对象检测模型为：所述数据采集设备当前使用的模型；
基于所获得的每一多媒体数据中对象的属性信息，从各个多媒体数据中选择包含具有目标属性的对象的数据，作为第一样本数据，其中，所述目标属性为：所述第一对象检测模型对多媒体数据中对象进行检测时，检测结果的准确度满足预设的自训练条件的多媒体数据中对象所具有的属性；
利用所述第一样本数据对所述第一对象检测模型进行训练，得到第二对象检测模型。
本申请的一个实施例中，所述方法还包括：
基于测试数据，测试所述第二对象检测模型进行对象检测的准确度，得到测试结果，其中，所述测试数据包括已有测试数据和/或新测试数据，所述已有测试数据为：已使用过的、用于测试所述第一对象检测模型进行对象检测的准确度的数据，所述新测试数据为：未使用过的测试数据；
根据所述测试结果，确定是否将所述数据采集设备中的所述第一对象检测模型更新为所述第二对象检测模型。
本申请的一个实施例中，所述新测试数据包括以下数据中的至少一种：
用户通过客户端输入的带有标注信息的数据；
由用户对候选数据进行标注得到的数据，其中，所述候选数据为：从所述数据采集设备采集的 多媒体数据中抽取的数据。
本申请的一个实施例中，所述基于测试数据，测试所述第二对象检测模型进行对象检测的准确度，得到测试结果，包括：
将测试数据输入所述第一对象检测模型进行对象检测，得到第一输出结果，并将所述测试数据输入所述第二对象检测模型进行对象检测，得到第二输出结果；
计算所述第一输出结果相对标注信息的第一差异值，并计算所述第二输出结果相对所述标注信息的第二差异值，其中，所述标注信息为：预先对所述测试数据进行标注的信息；
基于所述第一差异值与第二差异值，判断所述第二对象检测模型进行对象检测的准确度是否高于所述第一对象检测模型进行对象检测的准确度，得到判断结果，作为测试结果。
本申请的一个实施例中，所述方法还包括：
在确定将所述数据采集设备中的所述第一对象检测模型更新为所述第二对象检测模型的情况下，按照预设的量化方式对所述第二对象检测模型进行量化；
利用预设的校验数据，校验量化后第二对象检测模型相对量化前第二对象检测模型的一致性；
在检验结果表征量化后第二对象检测模型相对量化前第二对象检测模型一致的情况下，将量化后的第二对象检测模型的格式转化为所述数据采集设备支持的格式，并将所述数据采集设备中的所述第一对象检测模型更新为格式转化后的第二对象检测模型。
本申请的一个实施例中，所述方法还包括：
在确定将所述数据采集设备中的所述第一对象检测模型更新为所述第二对象检测模型的情况下，将同类数据采集设备中的所述第一对象检测模型更新为所述第二对象检测模型，其中，所述同类数据采集设备为：使用所述第一对象检测模型的数据采集设备。
本申请的一个实施例中，所述利用所述第一样本数据对所述第一对象检测模型进行训练，得到第二对象检测模型，包括：
利用所述第一样本数据和第二样本数据，对所述第一对象检测模型进行训练，得到第二对象检测模型，其中，所述第二样本数据为：用户通过客户端输入的多媒体数据。
本申请的一个实施例中，通过以下方式获得所述目标属性：
获得候选属性的出现频次，其中，所述候选属性为：所述第一对象检测模型对各多媒体数据进行对象检测的结果中置信度低于预设置信度阈值的对象的属性信息；
确定出现频次达到预设的频次条件的候选属性，作为目标属性。
第二方面，本申请实施例提供了一种自训练系统，所述系统包括自训练服务器和数据采集设备，其中：
所述数据采集设备，用于采集多媒体数据，基于第一对象检测模型检测各个多媒体数据中对象的属性信息，并向所述自训练服务器发送所采集的多媒体数据及各个多媒体数据中对象的属性信息，其中，所述第一对象检测模型为：所述数据采集设备当前使用的模型；
所述自训练服务器，用于接收所述数据采集设备发送的多媒体数据及各个多媒体数据中对象的属性信息，基于所获得的每一多媒体数据中对象的属性信息，从各个多媒体数据中选择包含具有目标属性的对象的数据，作为第一样本数据，利用所述第一样本数据对所述第一对象检测模型进行训练，得到第二对象检测模型，其中，所述目标属性为：所述第一对象检测模型对多媒体数据中对象进行检测时，检测结果的准确度满足预设的自训练条件的多媒体数据中对象所具有的属性。
第三方面，本申请实施例提供了一种自训练装置，所述装置包括：
数据获得模块，用于获得数据采集设备采集的多媒体数据，并获得第一对象检测模型对各个多媒体数据中对象进行检测得到的属性信息，其中，所述第一对象检测模型为：所述数据采集设备当前使用的模型；
数据选择模块，用于基于所获得的每一多媒体数据中对象的属性信息，从各个多媒体数据中选择包含具有目标属性的对象的数据，作为第一样本数据，其中，所述目标属性为：所述第一对象检测模型对多媒体数据中对象进行检测时，检测结果的准确度满足预设的自训练条件的多媒体数据中对象所具有的属性；
模型训练模块，用于利用所述第一样本数据对所述第一对象检测模型进行训练，得到第二对象检测模型。
本申请的一个实施例中，所述装置还包括：
准确度测试模块，用于基于测试数据，测试所述第二对象检测模型进行对象检测的准确度，得到测试结果，其中，所述测试数据包括已有测试数据和/或新测试数据，所述已有测试数据为：已使用过的、用于测试所述第一对象检测模型进行对象检测的准确度的数据，所述新测试数据为：未使用过的测试数据；
结果确定模块，用于根据所述测试结果，确定是否将所述数据采集设备中的所述第一对象检测模型更新为所述第二对象检测模型。
本申请的一个实施例中，所述新测试数据包括以下数据中的至少一种：
用户通过客户端输入的带有标注信息的数据；
由用户对候选数据进行标注得到的数据，其中，所述候选数据为：从所述数据采集设备采集的多媒体数据中抽取的数据。
本申请的一个实施例中，所述准确度测试模块，具体用于：
将测试数据输入所述第一对象检测模型进行对象检测，得到第一输出结果，并将所述测试数据输入所述第二对象检测模型进行对象检测，得到第二输出结果；
计算所述第一输出结果相对标注信息的第一差异值，并计算所述第二输出结果相对所述标注信息的第二差异值，其中，所述标注信息为：预先对所述测试数据进行标注的信息；
基于所述第一差异值与第二差异值，判断所述第二对象检测模型进行对象检测的准确度是否高于所述第一对象检测模型进行对象检测的准确度，得到判断结果，作为测试结果。
本申请的一个实施例中，所述装置还包括第一模型更新模块，具体用于：
在确定将所述数据采集设备中的所述第一对象检测模型更新为所述第二对象检测模型的情况下，按照预设的量化方式对所述第二对象检测模型进行量化；
利用预设的校验数据，校验量化后第二对象检测模型相对量化前第二对象检测模型的一致性；
在检验结果表征量化后第二对象检测模型相对量化前第二对象检测模型一致的情况下，将量化后的第二对象检测模型的格式转化为所述数据采集设备支持的格式，并将所述数据采集设备中的所述第一对象检测模型更新为格式转化后的第二对象检测模型。
本申请的一个实施例中，所述装置还包括第二模型更新模块，具体用于：
在确定将所述数据采集设备中的所述第一对象检测模型更新为所述第二对象检测模型的情况下，将同类数据采集设备中的所述第一对象检测模型更新为所述第二对象检测模型，其中，所述同类数据采集设备为：使用所述第一对象检测模型的数据采集设备。
本申请的一个实施例中，所述模型训练模块，具体用于：
利用所述第一样本数据和第二样本数据，对所述第一对象检测模型进行训练，得到第二对象检测模型，其中，所述第二样本数据为：用户通过客户端输入的多媒体数据。
本申请的一个实施例中，所述装置还包括目标属性确定模块，具体用于通过以下方式获得所述目标属性：
获得候选属性的出现频次，其中，所述候选属性为：所述第一对象检测模型对各多媒体数据进行对象检测的结果中置信度低于预设置信度阈值的对象的属性信息；
确定出现频次达到预设的频次条件的候选属性，作为目标属性。
第四方面，本申请实施例提供了一种电子设备，包括处理器、通信接口、存储器和通信总线，其中，处理器，通信接口，存储器通过通信总线完成相互间的通信；
存储器，用于存放计算机程序；
处理器，用于执行存储器上所存放的程序时，实现第一方面任一所述的方法步骤。
第五方面，本申请实施例提供了一种计算机可读存储介质，所述计算机可读存储介质内存储有计算机程序，所述计算机程序被处理器执行时实现第一方面任一所述的方法步骤。
本申请实施例还提供了一种包含指令的计算机程序产品，当其在计算机上运行时，使得计算机执行上述任一所述的自训练方法。
本申请实施例有益效果：
应用本申请实施例提供的自训练方案训练对象检测模型时，首先获得数据采集设备采集的多媒体数据，并获得第一对象检测模型对各个多媒体数据中对象进行检测得到的属性信息，其中，第一对象检测模型为：数据采集设备当前使用的模型，基于所获得的每一多媒体数据中对象的属性信息，从各个多媒体数据中选择包含具有目标属性的对象的数据，作为第一样本数据，其中，目标属性为：第一对象检测模型对多媒体数据中对象进行检测时，检测结果的准确度满足预设的自训练条件的多媒体数据中对象所具有的属性，利用第一样本数据对第一对象检测模型进行训练，得到第二对象检测模型。这样可以在第一对象检测模型应用的过程中，根据数据采集设备采集的多媒体数据，对第一对象检测模型进行训练。由于第一样本数据内的对象具有目标属性，而目标属性可以理解为第一对象检测模型对多媒体数据中对象进行检测时，检测结果的准确度较低的多媒体数据中对象所具有的属性，也就说明第一对象检测模型对具有目标属性的多媒体数据进行对象检测的准确度较低，这样利用第一样本数据对第一对象检测模型进行训练，可以提高第一对象检测模型对具有目标属性的多媒体数据进行对象检测的准确度。由此可见，应用本申请实施例提供的方案，可以提高对象检测模型进行对象检测的准确度。
附图说明
为了更清楚地说明本申请实施例或现有技术中的技术方案，下面将对实施例或现有技术描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本申请的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的实施例。
图1为本申请实施例提供的一种自训练方法的流程示意图；
图2为本申请实施例提供的另一种自训练方法的流程示意图；
图3为本申请实施例提供的一种模型更新判断方法的流程示意图；
图4为本申请实施例提供的一种模型测试方法的流程示意图；
图5为本申请实施例提供的一种自训练开启过程的示意图；
图6为本申请实施例提供的一种查询自训练状态过程的示意图；
图7为本申请实施例提供的第一种自训练系统的结构示意图；
图8为本申请实施例提供的第一种自训练过程的示意图；
图9为本申请实施例提供的第二种自训练系统的结构示意图；
图10为本申请实施例提供的客户端与数据采集设备之间的结构示意图；
图11为本申请实施例提供的第二种自训练过程的示意图；
图12为本申请实施例提供的第三种自训练系统的结构示意图；
图13为本申请实施例提供的一种自训练装置的结构示意图；
图14为本申请实施例提供的一种电子设备的结构示意图。
具体实施方式
下面将结合本申请实施例中的附图，对本申请实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例仅仅是本申请一部分实施例，而不是全部的实施例。基于本申请中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都属于本申请保护的范围。
为提高对象检测模型进行对象检测的准确度，本申请实施例提供了一种自训练方法、系统、装置、电子设备及存储介质，下面分别进行详细说明。
参见图1，图1为本申请实施例提供的一种自训练方法的流程示意图。该方法可以应用于电子设备，上述电子设备可以是具有对象检测功能的数据采集设备，上述数据采集设备中可以部署有对象检测模型，可以直接利用所部署的对象检测模型对所采集的多媒体数据中的对象进行检测。上述数据采集设备可以是指采集多媒体数据的设备，具体的，上述数据采集设备可以是用于采集图像的设备，如IPC(IP Camera，网络摄像机)等，或者是带有图像采集功能的手机、平板电脑等；上述数据采集设备也可以是用来采集音频的设备，如拾音器，也可以是带有音频采集功能的手机、平板电脑等。上述数据采集设备还可以是间接获得多媒体数据的存储设备，如NVR(Network Video Recorder，网络硬盘录像机)、DVR(Digital Video Recorder，数字视频录像机)等，具体的，摄像机、拾音器等设备可以将采集的多媒体数据发送至存储设备，存储设备从而获得多媒体数据。上述电子设备还可以是服务器等。
上述自训练方法包括如下步骤101至步骤103。
步骤101，获得数据采集设备采集的多媒体数据，并获得第一对象检测模型对各个多媒体数据中对象进行检测得到的属性信息。
其中，第一对象检测模型为：数据采集设备当前使用的模型。
上述多媒体数据可以是图像、视频、音频等。在多媒体数据为视频或图像的情况下，第一对象检测模型可以是用于对视频或图像中对象进行检测的模型，例如，在上述对象为车辆的情况下，第一对象检测模型可以是车辆检测模型、车辆分类模型等；在上述对象为人脸的情况下，第一对象检测模型可以是人脸识别模型；在多媒体数据为音频的情况下，第一对象检测模型可以是用于对音频中对象进行检测的模型，例如，在对象为语音片段的情况下，第一对象检测模型可以是语音识别模型等。
上述对象为第一对象检测模型所要检测的对象，属性信息为描述多媒体数据中对象所具有的属性的信息。例如，在第一检测模型为人脸识别模型的情况下，上述对象即为图像中的人脸，人脸的属性信息可以包括：人脸在图像中的位置、人脸是否佩戴饰物，如口罩、帽子、眼镜等；在第一对象检测模型为车辆检测模型的情况下，上述对象即为车辆，车辆的属性信息可以包括：车辆在图像中的位置、车辆的颜色、车辆的大小等。
具体的，数据采集设备可以对多媒体数据进行采集，电子设备可以获得上述多媒体数据，然后利用第一对象检测模型对多媒体数据中对象进行检测。在利用第一对象检测模型对多媒体数据中对象进行检测时，可以得到检测时的过程信息以及检测结果信息，从上述过程信息和检测结果信息中可以选择用于描述多媒体数据中对象所具有的属性的信息，作为属性信息。
本申请的一个实施例中，可以获得数据采集设备采集的全部多媒体数据，也可以从数据采集设备采集的数据中抽取多媒体数据，具体可以按照预设的时间间隔抽取多媒体数据，也可以随机抽取多媒体数据等，本申请实施例并不对此进行限定。
步骤102，基于所获得的每一多媒体数据中对象的属性信息，从各个多媒体数据中选择包含具有目标属性的对象的数据，作为第一样本数据。
其中，目标属性为：第一对象检测模型对多媒体数据中对象进行检测时，检测结果的准确度满 足预设的自训练条件的多媒体数据中对象所具有的属性。目标属性可以理解为：第一对象检测模型检测的准确度较低的多媒体数据中对象所具有的属性，也就是，第一对象检测模型对包含具有目标属性的对象的多媒体数据进行对象检测时，所得的检测结果的准确度较低。例如，假设车辆检测模型对车辆位于图像右上角的图像进行对象检测的准确度较低，则可以将车辆位于图像右上角作为目标属性。上述目标属性可以是用户预先设定的对象的属性。
上述自训练条件可以是：检测结果的准确度低于预设的准确度阈值；也可以是：检测结果的准确度低于准确度阈值的多媒体数据中对象所具有的属性的出现频次，达到预设的频次阈值；还可以是：检测结果的准确度低于准确度阈值的多媒体数据中对象所具有的属性的出现频次，占所有多媒体数据中对象所具有的属性的出现频次的比率，达到预设的比率阈值等。
上述检测结果的准确度可以是第一对象检测模型输出的检测结果的置信度，也可以是利用预设的评估算法计算得到的检测结果的准确度，还可以由用户对检测结果进行评分，根据评分得到检测结果的准确度。
具体的，可以预先在第一对象检测模型对多媒体数据中对象进行检测时，统计检测结果的准确度满足上述自训练条件的多媒体数据中对象所具有的属性，根据统计结果确定目标属性。在选择第一样本数据时，由于在步骤101中获得了多媒体数据中所包含的对象所具有的属性信息，因此可以从所获得的多媒体数据中，选择包含具有目标属性的对象的数据，作为第一样本数据。
例如，假设所确定的目标属性为白色的车辆，若所获得的图像中车辆具有的属性信息包括颜色为白色，则可以将该图像作为第一样本数据。
本申请的一个实施例中，属性信息中还可以包括对象评分，上述对象评分用于表征：多媒体数据中所包含的内容为该对象的概率。例如，针对人脸图像，其对象评分表征：该图像中所包含的内容为人脸的概率。这样在选择第一样本数据时，可以选择包含具有目标属性的对象、且对象评分达到预设的评分阈值的多媒体数据，作为第一样本数据。
步骤103，利用第一样本数据对第一对象检测模型进行训练，得到第二对象检测模型。
本申请的一个实施例中，由于所选择的第一样本数据中未携带标注信息，因此可以利用第一样本数据对第一对象检测模型进行无监督训练，得到第二对象检测模型。
本申请的一个实施例中，可以利用第一样本数据和第二样本数据，对第一对象检测模型进行训练，得到第二对象检测模型。
其中，第二样本数据为：用户通过客户端输入的多媒体数据。第二样本数据可以是从多媒体数据库中获取的数据，也可以是其他数据采集设备采集的数据等。第二样本数据可以是包含具有目标属性的对象的数据，也可以是包含具有其他属性的对象的数据。
上述客户端可以是手机、电子计算机、平板电脑等，用户通过客户端可以实现与电子设备的数据交互和指令交互。
具体的，用户通过客户端输入的多媒体数据可以是携带标注信息的数据，也可以是未携带标注信息的数据。在第二样本数据是携带标注信息的数据的情况下，可以利用第一样本数据对第一对象检测模型进行无监督训练，并利用第二样本数据对第一对象检测模型进行有监督训练，最终得到第二对象检测模型；在第二样本数据是未携带标注信息的数据的情况下，可以利用第一样本数据与第二样本数据对第一对象检测模型进行无监督训练，得到第二对象检测模型。
由于第二样本数据为用户通过客户端输入的数据，用户在通过客户端输入第二样本数据时，可以选择更有针对性的多媒体数据用于对第一对象检测模型进行训练，从而进一步提高训练后第二对象检测模型的准确度。
本申请的一个实施例中，在电子设备中存储有多个对象检测模型时，需要从所存储的对象检测模型中选择第一对象检测模型，利用第一样本数据对所选择的第一对象检测模型进行训练。在选择 第一对象检测模型时，也可以根据第一样本数据中的对象的属性信息进行选择，例如，在第一样本数据的对象的属性信息表示该第一样本数据为车辆图像时，则可以选择车辆检测模型作为第一对象检测模型，在第一样本数据的对象的属性信息表示该第一样本数据为人脸图像时，则可以选择人脸识别模型作为第一对象检测模型。
应用上述实施例提供的自训练方案训练对象检测模型时，首先获得数据采集设备采集的多媒体数据，并获得第一对象检测模型对各个多媒体数据中对象进行检测得到的属性信息，其中，第一对象检测模型为：数据采集设备当前使用的模型，基于所获得的每一多媒体数据中对象的属性信息，从各个多媒体数据中选择包含具有目标属性的对象的数据，作为第一样本数据，其中，目标属性为：第一对象检测模型对多媒体数据中对象进行检测时，检测结果的准确度满足预设的自训练条件的多媒体数据中对象所具有的属性，利用第一样本数据对第一对象检测模型进行训练，得到第二对象检测模型。这样可以在第一对象检测模型应用的过程中，根据数据采集设备采集的多媒体数据，对第一对象检测模型进行训练。由于第一样本数据内的对象具有目标属性，而目标属性可以理解为第一对象检测模型对多媒体数据中对象进行检测时，检测结果的准确度较低的多媒体数据中对象所具有的属性，也就说明第一对象检测模型对具有目标属性的多媒体数据进行对象检测的准确度较低，这样利用第一样本数据对第一对象检测模型进行训练，可以提高第一对象检测模型对具有目标属性的多媒体数据进行对象检测的准确度。由此可见，应用上述实施例提供的方案，可以提高对象检测模型进行对象检测的准确度。
除此之外，上述实施例提供的方案中，从图像采集设备采集的多媒体数据中选择第一样本数据，利用所选择的第一样本数据对第一对象检测模型进行训练，这样无需从各个对媒体数据平台搜集样本数据，可以降低获得样本数据的困难度。所选择的第一样本数据为包含具有目标属性的对象的数据，在选择第一样本数据时不受工作人员主观影响，从而提高所选择的第一样本数据的准确度，进而提高训练得到的对象检测模型的准确度。并且在基于所筛选的第一样本数据进行模型训练时，无需对第一样本数据进行标注，减少对样本数据进行标注的工作量，提高模型训练的效率。
本申请的一个实施例中，在获得目标属性时，可以获得候选属性的出现频次，确定出现频次达到预设的频次条件的候选属性，作为目标属性。
其中，候选属性为：第一对象检测模型对各多媒体数据进行对象检测的结果中置信度低于预设置信度阈值的对象的属性信息。上述候选属性可以理解为：各多媒体数据中，利用第一对象检测模型进行对象检测所得结果的置信度低于预设置信度阈值的数据中对象的属性信息。
频次条件可以是出现频次达到预设的次数阈值，例如，次数阈值可以是100次、500次、1000次等。频次条件也可以是各个候选属性出现频次占所有候选属性出现频次总数的比率达到预设的比率阈值。例如，假设对象为车辆，候选属性包括黑色车辆、灰色车辆、绿色车辆，比率阈值为60％，其中，黑色车辆的出现频次为30次，灰色车辆的出现频次为80次，绿色车辆的出现频次为20次，可见，灰色车辆的出现频次占出现频次总数的比率为61％，达到了上述比率阈值，因此可以认为灰色车辆的出现频次达到预设的频次条件，可以将灰色车辆作为目标属性。
具体的，第一对象检测模型在使用过程中，对输入的多媒体数据中对象进行检测时，可以得到检测结果，以及检测结果的置信度。可以记录检测结果的置信度低于置信度阈值的多媒体数据中对象所具有的属性，作为候选属性，当所记录的候选属性的出现频次达到预设的频次条件时，则可以将该候选属性作为目标属性。
参见图2，本申请的一个实施例中，在上述步骤103训练得到第二对象检测模型后，还可以包括如下步骤104。
步骤104，基于测试数据，测试第二对象检测模型进行对象检测的准确度，得到测试结果，根据测试结果，确定是否将数据采集设备中的第一对象检测模型更新为第二对象检测模型。
其中，测试结果表征第二对象检测模型进行对象检测的准确度。测试数据包括已有测试数据和/或新测试数据，已有测试数据为：已使用过的、用于测试第一对象检测模型进行对象检测的准确度的数据。上述已有测试数据可以是预先存储的多媒体数据。
本申请的一个实施例中，上述已有测试数据中可以携带标注信息，这样便于利用已有测试数据对第二对象检测模型进行测试。上述已有测试数据中也可以不携带标注信息，本申请实施例并不对此进行限定。
新测试数据为：未使用过的测试数据，也就是之前未对第一对象检测模型的准确度进行测试过的数据。新测试数据可以是预先获得的数据中未使用过的数据，也可以是在得到表征第二对象检测模型进行对象检测的准确度高于第一对象检测模型进行对象检测的准确度的第一测试结果后，获得的数据。
上述新测试数据中可以携带标注信息，也可以不携带标注信息等，本申请实施例并不对此进行限定。
上述方案中，可以测试第二对象检测模型进行对象检测的准确度，并根据检测出的准确度，确定是否将数据采集设备中的第一对象检测模型更新为第二对象检测模型。
具体的，对于训练得到的第二对象检测模型，可以基于测试数据检测第二对象检测模型进行对象检测的准确度，在准确度较高的情况下，可以将数据采集设备中的第一对象检测模型更新为第二对象检测模型；在准确度较低的情况下，可以不对数据采集设备中的第一对象检测模型进行更新，也可以继续对第二对象检测模型进行训练，直至第二对象检测模型进行对象检测的准确度高于第一对象检测模型进行对象检测的准确度。
本申请的一个实施例中，为得到测试结果，可以将测试数据输入第一对象检测模型进行对象检测，得到第一输出结果，并将测试数据输入第二对象检测模型进行对象检测，得到第二输出结果，计算第一输出结果相对标注信息的第一差异值，并计算第二输出结果相对标注信息的第二差异值，基于第一差异值与第二差异值，判断第二对象检测模型进行对象检测的准确度是否高于第一对象检测模型进行对象检测的准确度，得到判断结果，作为测试结果。
其中，标注信息为：预先对测试数据进行标注的信息。
上述第一差异值和第二差异值可以通过计算欧氏距离、余弦相似度、曼哈顿距离等得到。例如，可以计算第一输出结果相对标注信息的余弦相似度，作为第一差异值。
具体的，在第一差异值大于第二差异值的情况下，说明第一对象检测模型对测试数据中对象进行检测得到的第一输出结果与标注信息相差较大，第二对象检测模型对测试数据中对象进行检测得到的第二输出结果与标注信息相差较小，进而说明第一对象检测模型对对象进行检测的准确度较低，第二对象检测模型对对象进行检测的准确度较高，因此可以判断第二对象检测模型进行对象检测的准确度高于第一对象检测模型进行对象检测的准确度；
在第一差异值小于第二差异值的情况下，说明第一对象检测模型对测试数据中对象进行检测得到的第一输出结果与标注信息相差较小，第二对象检测模型对测试数据中对象进行检测得到的第二输出结果与标注信息相差较大，进而说明第一对象检测模型对对象进行检测的准确度较高，第二对象检测模型对对象进行检测的准确度较低，因此可以判断第一对象检测模型进行对象检测的准确度高于第二对象检测模型进行对象检测的准确度；
将上述判断结果作为测试结果，进而可以根据测试结果确定是否将数据采集设备中的第一对象检测模型更新为第二对象检测模型。
本申请的一个实施例中，还可以仅将测试数据输入第二对象检测模型，得到第三输出结果，计算第三输出结果相对标注信息的第三差异值，在第三差异值小于预设的差异度阈值的情况下，认为第二对象检测模型进行对象检测的准确度较高，在第三差异值不小于于预设的差异度阈值的情况下， 认为第二对象检测模型进行对象检测的准确度较低。
其中，上述差异度阈值可以由人工根据经验设定，也可以通过计算第一对象检测模型对测试数据中对象进行检测得到的测试结果相对标注信息的差异值的均值得到，上述均值可以是算数平均值、加权平均值等。
本申请的一个实施例中，上述新测试数据可以是用户通过客户端输入的带有标注信息的数据，具体可以是从多媒体数据库中获取的数据，也可以是其他数据采集设备采集的数据等。在获得上述多媒体数据后，可以由用户对多媒体数据进行标注，将标注后的多媒体数据通过客户端输入电子设备，从而得到新测试数据。
上述新测试数据还可以是由用户对候选数据进行标注得到的数据，其中，候选数据为：从数据采集设备采集的多媒体数据中抽取的数据。具体的，电子设备在上述步骤101获得数据采集设备采集的多媒体数据后，可以从多媒体数据中抽取部分数据发送至客户端，用户通过客户端对上述数据进行标注，客户端再将标注后的数据返回至电子设备，从而得到新测试数据。
其中，在从多媒体数据中抽取候选数据时，可以随机抽取，也可以按照预设的数据采集间隔抽取，例如，在多媒体数据为图像时，可以每间隔60帧抽取一图像，作为候选数据。也可以根据多媒体数据所包含对象的属性信息，抽取满足预设属性的数据，作为候选数据。所抽取的数据可以是用于进行模型训练的数据，也可以是进行模型训练的数据以外的数据。
本申请的一个实施例中，还可以从第一样本数据中抽取部分数据，由用户对所抽取的数据进行标注，将标注后的数据作为上述新测试数据。可以理解为，基于所获得的每一多媒体数据中对象的属性信息，从各个多媒体数据中选择包含具有目标属性的对象的数据。将所选择的数据分为两部分，一部分作为样本数据，用于对第一对象检测模型进行训练；另一部分数据在标注后作为测试数据，用于对训练后得到的第二对象检测模型进行测试。
本申请的一个实施例中，在测试数据不携带标注信息的情况下，可以将测试数据输入第二对象检测模型，得到模型输出的检测结果的置信度，将上述置信度作为第二对象检测模型进行对象检测的准确度，从而得到测试结果。
本申请的一个实施例中，还可以根据测试得到的第二对象检测模型的准确度，对第二对象检测模型进行评分，并将评分结果通过客户端向用户进行展示。用户基于评分结果，可以通过客户端发送指令，指示是否将数据采集设备中的第一对象检测模型更新为第二对象检测模型。
本申请的一个实施例中，在确定将数据采集设备中的第一对象检测模型更新为第二对象检测模型的情况下，按照预设的量化方式对第二对象检测模型进行量化，利用预设的校验数据，校验量化后第二对象检测模型相对量化前第二对象检测模型的一致性，在检验结果表征量化后第二对象检测模型相对量化前第二对象检测模型一致的情况下，将量化后的第二对象检测模型的格式转化为数据采集设备支持的格式，并将数据采集设备中的第一对象检测模型更新为格式转化后的第二对象检测模型。
其中，上述量化方式可以是权重量化、激活值量化等。对第二对象检测模型进行量化可以使得第二对象检测模型更加精简，这样在运行第二对象检测模型时可以节省内存，提高模型运行效率。
具体的，针对量化后的第二对象检测模型，可以利用检验数据对量化后第二对象检测模型相对量化前第二对象检测模型的一致性进行检验。具体的，可以将校验数据输入量化后的第二对象检测模型，得到第一校验结果，并将校验数据输入量化前的第二对象检测模型，得到第二校验结果，计算第一校验结果相对第二校验结果的差异值，根据上述差异值判断量化后第二对象检测模型相对量化前第二对象检测模型的一致性。在上述差异值小于预设的差异值阈值的情况下，说明量化后第二对象检测模型对多媒体数据中对象进行检测得到的检测结果，相对量化前第二对象检测模型对多媒体数据中对象进行检测得到的检测结果差异度较小，因此可以认为量化后第二对象检测模型相对量 化前第二对象检测模型一致。
本申请的一个实施例中，在检验结果表征量化后第二对象检测模型相对量化前第二对象检测模型一致的情况下，可以对量化后第二对象检测模型进行格式转化。
具体的，不同的数据采集设备支持的模型的格式可能不同，上述格式包括模型中数据的位宽、排列方式、进位制等。例如，不同的数据采集设备部署的处理器可能由不同的厂商生产得到，如英伟达、英特尔、AMD(Advanced Micro Devices，超微半导体)等，不同厂商生产的处理器所支持的模型的格式不同。可以首先确定数据采集设备所支持的模型的格式，然后按照所确定的格式对量化后的第二对象检测模型进行格式转化，这样可以使得格式转化后的第二对象检测模型可以运行在数据采集设备中，便于将数据采集设备中的第一对象检测模型更新为格式转化后的第二对象检测模型。
本申请的一个实施例中，在确定将数据采集设备中的第一对象检测模型更新为第二对象检测模型的情况下，可以将同类数据采集设备中的第一对象检测模型更新为第二对象检测模型。
其中，同类数据采集设备为：使用第一对象检测模型的数据采集设备。例如，假设第一对象检测模型为人脸识别模型，则同类数据采集设备为使用人脸识别模型的数据采集设备。
具体的，由于各个同类数据采集设备中使用的第一对象检测模型相同，在确定将数据采集设备中的第一对象检测模型更新为第二对象检测模型的情况下，可以说明第二对象检测模型进行对象检测的准确度高于第一对象检测模型的准确度，也就是第二对象检测模型的准确度高于各个同类数据采集设备中使用的第一对象检测模型，因此可以将各个同类数据采集设备中的第一对象检测模型更新为第二对象检测模型。
参见图3，本申请的一个实施例中，对于上述步骤104在确定是否对数据采集设备中的第一对象检测模型进行更新时，可以包括如下步骤1041至步骤1043。
步骤1041，基于已有测试数据，测试第二对象检测模型进行对象检测的准确度，得到第一测试结果。
其中，已有测试数据中携带有标注信息。
本申请的一个实施例中，可以将已有测试数据输入第一对象检测模型进行对象检测，得到第一输出结果，并将已有测试数据输入第二对象检测模型进行对象检测，得到第二输出结果，计算第一输出结果相对标注信息的第一差异值，并计算第二输出结果相对标注信息的第二差异值，基于第一差异值与第二差异值，判断第二对象检测模型进行对象检测的准确度是否高于第一对象检测模型进行对象检测的准确度，得到判断结果，作为第一测试结果。
在第一差异值大于第二差异值的情况下，说明第一对象检测模型对已有测试数据中对象进行检测得到的第一输出结果与标注信息相差较大，第二对象检测模型对测试数据中对象进行检测得到的第二输出结果与标注信息相差较小，进而说明第一对象检测模型对对象进行检测的准确度较低，第二对象检测模型对对象进行检测的准确度较高；
在第一差异值小于第二差异值的情况下，说明第一对象检测模型对已有测试数据中对象进行检测得到的第一输出结果与标注信息相差较小，第二对象检测模型对测试数据中对象进行检测得到的第二输出结果与标注信息相差较大，进而说明第一对象检测模型对对象进行检测的准确度较高，第二对象检测模型对对象进行检测的准确度较低。
本申请的一个实施例中，还可以仅将已有测试数据输入第二对象检测模型，得到第三输出结果，计算第三输出结果相对标注信息的第三差异值，在第三差异值小于预设的差异度阈值的情况下，认为第二对象检测模型进行对象检测的准确度较高，在第三差异值不小于于预设的差异度阈值的情况下，认为第二对象检测模型进行对象检测的准确度较低。其中，上述差异度阈值可以由人工根据经验设定，也可以通过计算第一对象检测模型对已有测试数据中对象进行检测得到的测试结果相对标 注信息的差异值的均值得到，上述均值可以是算数平均值、加权平均值等。
步骤1042，在第一测试结果表征第二对象检测模型进行对象检测的准确度高于第一对象检测模型进行对象检测的准确度的情况下，基于新测试数据，重新测试第二对象检测模型进行对象检测的准确度，得到第二测试结果。
具体的，在基于已有测试数据得到表征第二对象检测模型进行对象检测的准确度高于第一对象检测模型进行对象检测的准确度的第一测试结果后，可以利用新测试数据再次对第二对象检测模型进行测试，得到第二测试结果。其中测试方式同上述步骤1041类似，在此不做赘述。
步骤1043，根据第二测试结果，确定是否将数据采集设备中的第一对象检测模型更新为第二对象检测模型。
具体的，可以在第二测试结果表征第二对象检测模型进行对象检测的准确度高于第一对象检测模型进行对象检测的准确度的情况下，确定将数据采集设备中的第一对象检测模型更新为第二对象检测模型，否则，不对数据采集设备中的第一对象检测模型进行更新。
还可以将第二测试结果通过展示给用户，由用户确定是否将数据采集设备中的第一对象检测模型更新为第二对象检测模型，在接收到用户通过客户端发送的更新指令的情况下，将数据采集设备中的第一对象检测模型更新为第二对象检测模型。
这样首先基于已有测试数据对第二对象检测模型进行对象检测的准确度进行测试，在第一测试结果表征第二对象检测模型的准确度较高时，进一步基于新测试数据对第二对象检测模型的准确度进行测试，再根据第二测试结果确定是否将数据采集设备中的第一对象检测模型更新为第二对象检测模型，可以使得测试结果的准确度更高。
本申请的一个实施例中，在未获得用户通过客户端发送的新测试数据的情况下，则可以直接根据第一测试结果确定是否将数据采集设备中的第一对象检测模型更新为第二对象检测模型。
参见图4，图4为本申请实施例提供的一种模型测试方法的流程示意图，该方法包括如下步骤401至步骤406。
步骤401，利用已有测试数据对第二对象检测模型进行对象检测的准确度进行测试，得到第一测试结果；
步骤402，基于第一测试结果判断第二对象检测模型进行对象检测的准确度是否高于第一对象检测模型进行对象检测的准确度，若为否，执行步骤403，若为是，执行步骤404；
步骤403，继续对第二对象检测模型进行训练；
步骤404，判断是否获得新测试数据，若为是，执行步骤405，若为否，执行步骤406；
步骤405，利用新测试数据对第二对象检测模型进行对象检测的准确度进行测试，得到第二测试结果；
步骤406，通过客户端对测试结果进行展示。
参见图5，图5为本申请实施例提供的一种自训练开启过程的示意图，客户端可以将GUI(Graphical User Interface，图形用户接口)展示给用户，用户通过操作GUI界面中确定开启自训练，客户端通过GUI获得开启自训练的指令后，再向电子设备发送开启自训练的指令。电子设备中可以包含执行自训练的应用程序，在接收到开启自训练的指令后，首先校验该指令是否正确，若不正确，则生成表征指令发生错误的错误码信息，作为结果信息返回至客户端，若正确，则对自训练功能进行初始化，并判断是否初始化成功。若为否，则生成表征初始化失败的错误码信息，作为结果信息返回至客户端，若为是，则将初始化成功的信息作为结果信息返回至客户端。客户端在向电子设备发送开启自训练的指令后，检测是否收到电子设备返回的结果信息，若为否，则等待接收上述结果信息，若为是，则根据结果信息判断是否成功开启自训练。若为否，则可以提示用户开启自训练失败，并根据返回结果中的错误码信息提示用户，具体可以以弹窗的形式对用户做出提示， 若为是，则可以提示用户开启自训练成功。
应用上述实施例提供的方案，用户可以通过客户端自主选择是否开启执行自训练，提高开启自训练的自由度，给用户带来更好的体验。
本申请的一个实施例中，也可以是电子设备在检测到使用的第一对象检测模型对包含具有某一属性的对象的多媒体数据中对象进行检测的准确度较低的情况下，开启自训练，并将上述属性作为目标属性，从而提高第一对象检测模型对包含具有任一属性的对象的多媒体数据中对象进行检测的准确度，并且提高自训练的自动化程度。
参见图6，图6为本申请实施例提供的一种查询自训练状态过程的示意图。用户可以通过操作客户端提供的GUI界面查询自训练状态，客户端在接收到查询自训练状态的指令后，可以向电子设备发送查询自训练状态的指令，电子设备接收上述指令，并查询自训练状态，将查询到的自训练状态作为结果信息，并返回至客户端。客户端检测是否接收到结果信息，若为否，则继续等待接收结果信息，若为是，则根据结果信息判断自训练状态是否发生改变。若为否，则提示用户自训练状态未发生改变，若为是，则根据结果信息更新自训练状态，并判断自训练是否完成，若为是，则提示用户自训练结束，若为否，则返回查询自训练状态。
本申请实施例提供的自训练方法可以应用在自训练系统中，一种情况下，在自训练系统包括客户端、自训练服务器、数据采集设备的情况下，自训练方法可以应用在自训练系统的自训练服务器中，上述数据采集设备可以是具有分析功能的IPC、NVR、DVR，也可以是具有分析功能的服务器等。另一种情况下，在自训练系统包括客户端和数据采集设备的情况下，自训练方法也可以应用在数据采集设备中，下面分别进行详细说明。
参见图7，图7为本申请实施例提供的第一种自训练系统的结构示意图。如图7所示，自训练系统中可以包括客户端、自训练服务器和多个数据采集设备，其中，客户端与自训练服务器之间可以通过网络进行连接，自训练服务器与各个数据采集设备之间也可以通过网络进行连接。
参见图8，图8为本申请实施例提供的第一种自训练过程的示意图。如图8所示：
数据采集设备可以获得多媒体数据，并利用对象检测模型对多媒体数据进行对象检测，得到多媒体数据中对象的属性信息，作为检测结果；
自训练服务器可以从数据采集设备中获得多媒体数据，以及多媒体数据中对象的属性信息，并从多媒体数据中选择包含具有目标属性的对象的数据，作为用于进行模型训练的第一样本数据。自训练服务器还可以根据多媒体数据的属性信息，从存储的多个对象检测模型中选择对应的第一对象检测模型；
其中，自训练服务器在获得多媒体数据和属性信息时，可以从多个同类数据采集设备中获得多媒体数据和属性信息；
客户端也可以获得第二样本数据，并对第二样本数据进行标注，将标注后的第二样本数据发送至自训练服务器；
自训练服务器可以对第一样本数据和第二样本数据进行管理，并利用第一样本数据和第二样本数据，对所选择的第一对象检测模型进行自训练，得到第二对象检测模型。然后利用测试数据对第二对象检测模型的准确度进行评估；
自训练服务器可以从所管理的样本数据中选择未标注的数据发送至客户端，由用户通过客户端对上述数据进行标注，得到新测试数据，并将新测试数据发送至自训练服务器，这样自训练服务器可以利用新测试数据对第二对象检测模型的准确度进行评估；
自训练服务器在评估第二对象检测模型的准确度未满足更新条件的情况下，可以继续从数据采集设备中获得多媒体数据及属性信息来继续对模型进行训练；在评估第二对象检测模型的准确度满足更新条件的情况下，可以对模型进行转化，包括对第二对象检测模型进行量化和格式转化，并校 验转化后第二对象检测模型相对转化前第二对象检测模型的一致性，在校验结果不一致的情况下，可以继续从数据采集设备中获得多媒体数据及属性信息来继续对模型进行训练，在校验结果一致的情况下，可以进行模型更新，将数据采集设备中第一对象检测模型更新为第二对象检测模型。
除此之外，自训练服务器还可以对训练得到的第二对象检测模型进行存储，这样便于后续针对第二对象检测模型进行自训练。
参见图9，图9为本申请实施例提供的第二种自训练系统的结构示意图。如图9所示，自训练系统中可以包括客户端和多个数据采集设备，其中，客户端与多个数据采集设备之间可以通过网络进行连接。
参见图10，图10为本申请实施例提供的客户端与数据采集设备之间的结构示意图。如图10所示，客户端与数据采集设备相连接，数据采集设备中包含主控单元和多个计算单元，其中，主控单元可以用于读取数据、调用计算单元等，各个计算单元可以用于模型运行、模型训练等。数据采集设备中还可以包括存储单元，用于对数据采集设备所部署的对象检测模型、所采集得到的多媒体数据、训练后的对象检测模型等进行存储。这样控制单元可以从存储单元中读取样本数据及对象检测模型，利用所读取的样本数据对所读取的对象检测模型进行训练，并将训练后的对象检测模型存储至存储单元。通过客户端可以控制数据采集设备中的主控单元调用指定的计算单元来进行模型训练。
参见图11，图11为本申请实施例提供的第二种自训练过程的示意图，如图11所示：
数据采集设备可以获得多媒体数据，并利用对象检测模型对多媒体数据进行对象检测，得到多媒体数据中对象的属性信息，作为检测结果。从检测结果中获得多媒体数据，以及多媒体数据中对象的属性信息，然后从多媒体数据中选择包含具有目标属性的对象的数据，作为用于进行模型训练的第一样本数据。还可以根据多媒体数据的属性信息，从存储的多个对象检测模型中选择对应的第一对象检测模型；
其中，在获得多媒体数据和属性信息时，可以从多个同类数据采集设备中获得多媒体数据和属性信息；
客户端也可以获得第二样本数据，并对第二样本数据进行标注，将标注后的第二样本数据发送至数据采集设备；
数据采集设备可以对第一样本数据和第二样本数据进行管理，并利用第一样本数据和第二样本数据，对所选择的第一对象检测模型进行自训练，得到第二对象检测模型。然后利用测试数据对第二对象检测模型的准确度进行评估；
数据采集设备可以从所管理的样本数据中选择未标注的数据发送至客户端，由用户通过客户端对上述数据进行标注，得到新测试数据，并将新测试数据发送至数据采集设备，这样数据采集设备可以利用新测试数据对第二对象检测模型的准确度进行评估；
数据采集设备在评估第二对象检测模型的准确度未满足更新条件的情况下，可以继续获得多媒体数据及属性信息来继续对模型进行训练；在评估第二对象检测模型的准确度满足更新条件的情况下，可以对模型进行转化，包括对第二对象检测模型进行量化和格式转化，并校验转化后第二对象检测模型相对转化前第二对象检测模型的一致性，在校验结果不一致的情况下，可以继续获得多媒体数据及属性信息来继续对模型进行训练，在校验结果一致的情况下，可以进行模型更新，将自身的第一对象检测模型更新为第二对象检测模型。
参见图12，图12为本申请实施例提供的第三种自训练系统的结构示意图，所述系统包括自训练服务器1201和数据采集设备1202，其中：
所述数据采集设备1202，用于采集多媒体数据，基于第一对象检测模型检测各个多媒体数据中对象的属性信息，并向所述自训练服务器1201发送所采集的多媒体数据及各个多媒体数据中对 象的属性信息，其中，所述第一对象检测模型为：所述数据采集设备1202当前使用的模型；
所述自训练服务器1201，用于接收所述数据采集设备1202发送的多媒体数据及各个多媒体数据中对象的属性信息，基于所获得的每一多媒体数据中对象的属性信息，从各个多媒体数据中选择包含具有目标属性的对象的数据，作为第一样本数据，利用所述第一样本数据对所述第一对象检测模型进行训练，得到第二对象检测模型，其中，所述目标属性为：所述第一对象检测模型对多媒体数据中对象进行检测时，检测结果的准确度满足预设的自训练条件的多媒体数据中对象所具有的属性。
本申请的一个实施例中，所述自训练服务器1201还用于：
基于测试数据，测试所述第二对象检测模型进行对象检测的准确度，得到测试结果，其中，所述测试数据包括已有测试数据和/或新测试数据，所述已有测试数据为：已使用过的、用于测试所述第一对象检测模型进行对象检测的准确度的数据，所述新测试数据为：未使用过的测试数据；
根据所述测试结果，确定是否将所述数据采集设备1202中的所述第一对象检测模型更新为所述第二对象检测模型。
本申请的一个实施例中，所述新测试数据包括以下数据中的至少一种：
用户通过客户端输入的带有标注信息的数据；
由用户对候选数据进行标注得到的数据，其中，所述候选数据为：从所述数据采集设备1202采集的多媒体数据中抽取的数据。
本申请的一个实施例中，所述自训练服务器1201，具体用于：
将测试数据输入所述第一对象检测模型进行对象检测，得到第一输出结果，并将所述测试数据输入所述第二对象检测模型进行对象检测，得到第二输出结果；
计算所述第一输出结果相对标注信息的第一差异值，并计算所述第二输出结果相对所述标注信息的第二差异值，其中，所述标注信息为：预先对所述测试数据进行标注的信息；
基于所述第一差异值与第二差异值，判断所述第二对象检测模型进行对象检测的准确度是否高于所述第一对象检测模型进行对象检测的准确度，得到判断结果，作为测试结果。
本申请的一个实施例中，所述自训练服务器1201还用于：
在确定将所述数据采集设备1202中的所述第一对象检测模型更新为所述第二对象检测模型的情况下，按照预设的量化方式对所述第二对象检测模型进行量化；
利用预设的校验数据，校验量化后第二对象检测模型相对量化前第二对象检测模型的一致性；
在检验结果表征量化后第二对象检测模型相对量化前第二对象检测模型一致的情况下，将量化后的第二对象检测模型的格式转化为所述数据采集设备1202支持的格式，并将所述数据采集设备1202中的所述第一对象检测模型更新为格式转化后的第二对象检测模型。
本申请的一个实施例中，所述自训练服务器1201还用于：
在确定将所述数据采集设备1202中的所述第一对象检测模型更新为所述第二对象检测模型的情况下，将同类数据采集设备1202中的所述第一对象检测模型更新为所述第二对象检测模型，其中，所述同类数据采集设备1202为：使用所述第一对象检测模型的数据采集设备1202。
本申请的一个实施例中，所述自训练服务器1201，具体用于：
利用所述第一样本数据和第二样本数据，对所述第一对象检测模型进行训练，得到第二对象检测模型，其中，所述第二样本数据为：用户通过客户端输入的多媒体数据。
本申请的一个实施例中，所述自训练服务器1201用于通过以下方式获得所述目标属性：
获得候选属性的出现频次，其中，所述候选属性为：所述第一对象检测模型对各多媒体数据进行对象检测的结果中置信度低于预设置信度阈值的对象的属性信息；
确定出现频次达到预设的频次条件的候选属性，作为目标属性。
参见图13，图13为本申请实施例提供的一种自训练装置的结构示意图，所述装置包括：
数据获得模块1301，用于获得数据采集设备采集的多媒体数据，并获得第一对象检测模型对各个多媒体数据中对象进行检测得到的属性信息，其中，所述第一对象检测模型为：所述数据采集设备当前使用的模型；
数据选择模块1302，用于基于所获得的每一多媒体数据中对象的属性信息，从各个多媒体数据中选择包含具有目标属性的对象的数据，作为第一样本数据，其中，所述目标属性为：所述第一对象检测模型对多媒体数据中对象进行检测时，检测结果的准确度满足预设的自训练条件的多媒体数据中对象所具有的属性；
模型训练模块1303，用于利用所述第一样本数据对所述第一对象检测模型进行训练，得到第二对象检测模型。
本申请的一个实施例中，所述装置还包括：
准确度测试模块，用于基于测试数据，测试所述第二对象检测模型进行对象检测的准确度，得到测试结果，其中，所述测试数据包括已有测试数据和/或新测试数据，所述已有测试数据为：已使用过的、用于测试所述第一对象检测模型进行对象检测的准确度的数据，所述新测试数据为：未使用过的测试数据；
结果确定模块，用于根据所述测试结果，确定是否将所述数据采集设备中的所述第一对象检测模型更新为所述第二对象检测模型。
本申请的一个实施例中，所述新测试数据包括以下数据中的至少一种：
用户通过客户端输入的带有标注信息的数据；
由用户对候选数据进行标注得到的数据，其中，所述候选数据为：从所述数据采集设备采集的多媒体数据中抽取的数据。
本申请的一个实施例中，所述准确度测试模块，具体用于：
将测试数据输入所述第一对象检测模型进行对象检测，得到第一输出结果，并将所述测试数据输入所述第二对象检测模型进行对象检测，得到第二输出结果；
计算所述第一输出结果相对标注信息的第一差异值，并计算所述第二输出结果相对所述标注信息的第二差异值，其中，所述标注信息为：预先对所述测试数据进行标注的信息；
基于所述第一差异值与第二差异值，判断所述第二对象检测模型进行对象检测的准确度是否高于所述第一对象检测模型进行对象检测的准确度，得到判断结果，作为测试结果。
本申请的一个实施例中，所述装置还包括第一模型更新模块，具体用于：
在确定将所述数据采集设备中的所述第一对象检测模型更新为所述第二对象检测模型的情况下，按照预设的量化方式对所述第二对象检测模型进行量化；
利用预设的校验数据，校验量化后第二对象检测模型相对量化前第二对象检测模型的一致性；
在检验结果表征量化后第二对象检测模型相对量化前第二对象检测模型一致的情况下，将量化后的第二对象检测模型的格式转化为所述数据采集设备支持的格式，并将所述数据采集设备中的所述第一对象检测模型更新为格式转化后的第二对象检测模型。
本申请的一个实施例中，所述装置还包括第二模型更新模块，具体用于：
在确定将所述数据采集设备中的所述第一对象检测模型更新为所述第二对象检测模型的情况下，将同类数据采集设备中的所述第一对象检测模型更新为所述第二对象检测模型，其中，所述同类数据采集设备为：使用所述第一对象检测模型的数据采集设备。
本申请的一个实施例中，所述模型训练模块1303，具体用于：
利用所述第一样本数据和第二样本数据，对所述第一对象检测模型进行训练，得到第二对象检测模型，其中，所述第二样本数据为：用户通过客户端输入的多媒体数据。
本申请的一个实施例中，所述装置还包括目标属性确定模块，具体用于通过以下方式获得所述目标属性：
获得候选属性的出现频次，其中，所述候选属性为：所述第一对象检测模型对各多媒体数据进行对象检测的结果中置信度低于预设置信度阈值的对象的属性信息；
确定出现频次达到预设的频次条件的候选属性，作为目标属性。
应用上述实施例提供的自训练方案训练对象检测模型时，首先获得数据采集设备采集的多媒体数据，并获得第一对象检测模型对各个多媒体数据中对象进行检测得到的属性信息，其中，第一对象检测模型为：数据采集设备当前使用的模型，基于所获得的每一多媒体数据中对象的属性信息，从各个多媒体数据中选择包含具有目标属性的对象的数据，作为第一样本数据，其中，目标属性为：第一对象检测模型对多媒体数据中对象进行检测时，检测结果的准确度满足预设的自训练条件的多媒体数据中对象所具有的属性，利用第一样本数据对第一对象检测模型进行训练，得到第二对象检测模型。这样可以在第一对象检测模型应用的过程中，根据数据采集设备采集的多媒体数据，对第一对象检测模型进行训练。由于第一样本数据内的对象具有目标属性，而目标属性可以理解为第一对象检测模型对多媒体数据中对象进行检测时，检测结果的准确度较低的多媒体数据中对象所具有的属性，也就说明第一对象检测模型对具有目标属性的多媒体数据进行对象检测的准确度较低，这样利用第一样本数据对第一对象检测模型进行训练，可以提高第一对象检测模型对具有目标属性的多媒体数据进行对象检测的准确度。由此可见，应用上述实施例提供的方案，可以提高对象检测模型进行对象检测的准确度。
本申请实施例还提供了一种电子设备，如图14所示，包括处理器1401、通信接口1402、存储器1403和通信总线1404，其中，处理器1401，通信接口1402，存储器1403通过通信总线1404完成相互间的通信，
存储器1403，用于存放计算机程序；
处理器1401，用于执行存储器1403上所存放的程序时，实现上述自训练方法的步骤。
上述电子设备提到的通信总线可以是外设部件互连标准(Peripheral Component Interconnect，PCI)总线或扩展工业标准结构(Extended Industry Standard Architecture，EISA)总线等。该通信总线可以分为地址总线、数据总线、控制总线等。为便于表示，图中仅用一条粗线表示，但并不表示仅有一根总线或一种类型的总线。
通信接口用于上述电子设备与其他设备之间的通信。
存储器可以包括随机存取存储器(Random Access Memory，RAM)，也可以包括非易失性存储器(Non-Volatile Memory，NVM)，例如至少一个磁盘存储器。可选的，存储器还可以是至少一个位于远离前述处理器的存储装置。
上述的处理器可以是通用处理器，包括中央处理器(Central Processing Unit，CPU)、网络处理器(Network Processor，NP)等；还可以是数字信号处理器(Digital Signal Processing，DSP)、专用集成电路(Application Specific Integrated Circuit，ASIC)、现场可编程门阵列(Field-Programmable Gate Array，FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件。
在本申请提供的又一实施例中，还提供了一种计算机可读存储介质，该计算机可读存储介质内存储有计算机程序，所述计算机程序被处理器执行时实现上述任一自训练方法的步骤。
在本申请提供的又一实施例中，还提供了一种包含指令的计算机程序产品，当其在计算机上运行时，使得计算机执行上述实施例中任一自训练方法。
在上述实施例中，可以全部或部分地通过软件、硬件、固件或者其任意组合来实现。当使用软件实现时，可以全部或部分地以计算机程序产品的形式实现。所述计算机程序产品包括一个或多个计算机指令。在计算机上加载和执行所述计算机程序指令时，全部或部分地产生按照本申请实施例 所述的流程或功能。所述计算机可以是通用计算机、专用计算机、计算机网络、或者其他可编程装置。所述计算机指令可以存储在计算机可读存储介质中，或者从一个计算机可读存储介质向另一个计算机可读存储介质传输，例如，所述计算机指令可以从一个网站站点、计算机、服务器或数据中心通过有线(例如同轴电缆、光纤、数字用户线(DSL))或无线(例如红外、无线、微波等)方式向另一个网站站点、计算机、服务器或数据中心进行传输。所述计算机可读存储介质可以是计算机能够存取的任何可用介质或者是包含一个或多个可用介质集成的服务器、数据中心等数据存储设备。所述可用介质可以是磁性介质，(例如，软盘、硬盘、磁带)、光介质(例如，DVD)、或者半导体介质(例如固态硬盘Solid State Disk(SSD))等。
应用上述实施例提供的自训练方案训练对象检测模型时，首先获得数据采集设备采集的多媒体数据，并获得第一对象检测模型对各个多媒体数据中对象进行检测得到的属性信息，其中，第一对象检测模型为：数据采集设备当前使用的模型，基于所获得的每一多媒体数据中对象的属性信息，从各个多媒体数据中选择包含具有目标属性的对象的数据，作为第一样本数据，其中，目标属性为：第一对象检测模型对多媒体数据中对象进行检测时，检测结果的准确度满足预设的自训练条件的多媒体数据中对象所具有的属性，利用第一样本数据对第一对象检测模型进行训练，得到第二对象检测模型。这样可以在第一对象检测模型应用的过程中，根据数据采集设备采集的多媒体数据，对第一对象检测模型进行训练。由于第一样本数据内的对象具有目标属性，而目标属性可以理解为第一对象检测模型对多媒体数据中对象进行检测时，检测结果的准确度较低的多媒体数据中对象所具有的属性，也就说明第一对象检测模型对具有目标属性的多媒体数据进行对象检测的准确度较低，这样利用第一样本数据对第一对象检测模型进行训练，可以提高第一对象检测模型对具有目标属性的多媒体数据进行对象检测的准确度。由此可见，应用上述实施例提供的方案，可以提高对象检测模型进行对象检测的准确度。
需要说明的是，在本文中，诸如第一和第二等之类的关系术语仅仅用来将一个实体或者操作与另一个实体或操作区分开来，而不一定要求或者暗示这些实体或操作之间存在任何这种实际的关系或者顺序。而且，术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、物品或者设备不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、方法、物品或者设备所固有的要素。在没有更多限制的情况下，由语句“包括一个……”限定的要素，并不排除在包括所述要素的过程、方法、物品或者设备中还存在另外的相同要素。
本说明书中的各个实施例均采用相关的方式描述，各个实施例之间相同相似的部分互相参见即可，每个实施例重点说明的都是与其他实施例的不同之处。尤其，对于装置实施例、系统实施例、电子设备实施例、计算机可读存储介质实施例、计算机程序产品实施例而言，由于其基本相似于方法实施例，所以描述的比较简单，相关之处参见方法实施例的部分说明即可。
以上所述仅为本申请的较佳实施例，并非用于限定本申请的保护范围。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，均包含在本申请的保护范围内。

Claims (13)

  1. 一种自训练方法，其特征在于，所述方法包括：
    获得数据采集设备采集的多媒体数据，并获得第一对象检测模型对各个多媒体数据中对象进行检测得到的属性信息，其中，所述第一对象检测模型为：所述数据采集设备当前使用的模型；
    基于所获得的每一多媒体数据中对象的属性信息，从各个多媒体数据中选择包含具有目标属性的对象的数据，作为第一样本数据，其中，所述目标属性为：所述第一对象检测模型对多媒体数据中对象进行检测时，检测结果的准确度满足预设的自训练条件的多媒体数据中对象所具有的属性；
    利用所述第一样本数据对所述第一对象检测模型进行训练，得到第二对象检测模型。
  2. 根据权利要求1所述的方法，其特征在于，所述方法还包括：
    基于测试数据，测试所述第二对象检测模型进行对象检测的准确度，得到测试结果，其中，所述测试数据包括已有测试数据和/或新测试数据，所述已有测试数据为：已使用过的、用于测试所述第一对象检测模型进行对象检测的准确度的数据，所述新测试数据为：未使用过的测试数据；
    根据所述测试结果，确定是否将所述数据采集设备中的所述第一对象检测模型更新为所述第二对象检测模型。
  3. 根据权利要求2所述的方法，其特征在于，所述新测试数据包括以下数据中的至少一种：
    用户通过客户端输入的带有标注信息的数据；
    由用户对候选数据进行标注得到的数据，其中，所述候选数据为：从所述数据采集设备采集的多媒体数据中抽取的数据。
  4. 根据权利要求2所述的方法，其特征在于，所述基于测试数据，测试所述第二对象检测模型进行对象检测的准确度，得到测试结果，包括：
    将测试数据输入所述第一对象检测模型进行对象检测，得到第一输出结果，并将所述测试数据输入所述第二对象检测模型进行对象检测，得到第二输出结果；
    计算所述第一输出结果相对标注信息的第一差异值，并计算所述第二输出结果相对所述标注信息的第二差异值，其中，所述标注信息为：预先对所述测试数据进行标注的信息；
    基于所述第一差异值与第二差异值，判断所述第二对象检测模型进行对象检测的准确度是否高于所述第一对象检测模型进行对象检测的准确度，得到判断结果，作为测试结果。
  5. 根据权利要求2所述的方法，其特征在于，所述方法还包括：
    在确定将所述数据采集设备中的所述第一对象检测模型更新为所述第二对象检测模型的情况下，按照预设的量化方式对所述第二对象检测模型进行量化；
    利用预设的校验数据，校验量化后第二对象检测模型相对量化前第二对象检测模型的一致性；
    在检验结果表征量化后第二对象检测模型相对量化前第二对象检测模型一致的情况下，将量化后的第二对象检测模型的格式转化为所述数据采集设备支持的格式，并将所述数据采集设备中的所述第一对象检测模型更新为格式转化后的第二对象检测模型。
  6. 根据权利要求2所述的方法，其特征在于，所述方法还包括：
    在确定将所述数据采集设备中的所述第一对象检测模型更新为所述第二对象检测模型的情况下，将同类数据采集设备中的所述第一对象检测模型更新为所述第二对象检测模型，其中，所述同类数据采集设备为：使用所述第一对象检测模型的数据采集设备。
  7. 根据权利要求1-6中任一项所述的方法，其特征在于，所述利用所述第一样本数据对所述第一对象检测模型进行训练，得到第二对象检测模型，包括：
    利用所述第一样本数据和第二样本数据，对所述第一对象检测模型进行训练，得到第二对象检测模型，其中，所述第二样本数据为：用户通过客户端输入的多媒体数据。
  8. 根据权利要求1-6中任一项所述的方法，其特征在于，通过以下方式获得所述目标属性：
    获得候选属性的出现频次，其中，所述候选属性为：所述第一对象检测模型对各多媒体数据进行对象检测的结果中置信度低于预设置信度阈值的对象的属性信息；
    确定出现频次达到预设的频次条件的候选属性，作为目标属性。
  9. 一种自训练系统，其特征在于，所述系统包括自训练服务器和数据采集设备，其中：
    所述数据采集设备，用于采集多媒体数据，基于第一对象检测模型检测各个多媒体数据中对象的属性信息，并向所述自训练服务器发送所采集的多媒体数据及各个多媒体数据中对象的属性信息，其中，所述第一对象检测模型为：所述数据采集设备当前使用的模型；
    所述自训练服务器，用于接收所述数据采集设备发送的多媒体数据及各个多媒体数据中对象的属性信息，基于所获得的每一多媒体数据中对象的属性信息，从各个多媒体数据中选择包含具有目标属性的对象的数据，作为第一样本数据，利用所述第一样本数据对所述第一对象检测模型进行训练，得到第二对象检测模型，其中，所述目标属性为：所述第一对象检测模型对多媒体数据中对象进行检测时，检测结果的准确度满足预设的自训练条件的多媒体数据中对象所具有的属性。
  10. 一种自训练装置，其特征在于，所述装置包括：
    数据获得模块，用于获得数据采集设备采集的多媒体数据，并获得第一对象检测模型对各个多媒体数据中对象进行检测得到的属性信息，其中，所述第一对象检测模型为：所述数据采集设备当前使用的模型；
    数据选择模块，用于基于所获得的每一多媒体数据中对象的属性信息，从各个多媒体数据中选择包含具有目标属性的对象的数据，作为第一样本数据，其中，所述目标属性为：所述第一对象检测模型对多媒体数据中对象进行检测时，检测结果的准确度满足预设的自训练条件的多媒体数据中对象所具有的属性；
    模型训练模块，用于利用所述第一样本数据对所述第一对象检测模型进行训练，得到第二对象检测模型。
  11. 根据权利要求10所述的装置，其特征在于，所述装置还包括：
    准确度测试模块，用于基于测试数据，测试所述第二对象检测模型进行对象检测的准确度，得到测试结果，其中，所述测试数据包括已有测试数据和/或新测试数据，所述已有测试数据为：已使用过的、用于测试所述第一对象检测模型进行对象检测的准确度的数据，所述新测试数据为：未使用过的测试数据；
    结果确定模块，用于根据所述测试结果，确定是否将所述数据采集设备中的所述第一对象检测模型更新为所述第二对象检测模型；或
    所述新测试数据包括以下数据中的至少一种：户通过客户端输入的带有标注信息的数据；由用户对候选数据进行标注得到的数据，其中，所述候选数据为：从所述数据采集设备采集的多媒体数据中抽取的数据；或
    所述准确度测试模块，具体用于：将测试数据输入所述第一对象检测模型进行对象检测，得到第一输出结果，并将所述测试数据输入所述第二对象检测模型进行对象检测，得到第二输出结果；计算所述第一输出结果相对标注信息的第一差异值，并计算所述第二输出结果相对所述标注信息的第二差异值，其中，所述标注信息为：预先对所述测试数据进行标注的信息；基于所述第一差异值与第二差异值，判断所述第二对象检测模型进行对象检测的准确度是否高于所述第一对象检测模型进行对象检测的准确度，得到判断结果，作为测试结果；或
    所述装置还包括第一模型更新模块，具体用于：在确定将所述数据采集设备中的所述第一对象检测模型更新为所述第二对象检测模型的情况下，按照预设的量化方式对所述第二对象检测模型进行量化；利用预设的校验数据，校验量化后第二对象检测模型相对量化前第二对象检测模型的一致性；在检验结果表征量化后第二对象检测模型相对量化前第二对象检测模型一致的情况下，将量化 后的第二对象检测模型的格式转化为所述数据采集设备支持的格式，并将所述数据采集设备中的所述第一对象检测模型更新为格式转化后的第二对象检测模型；或
    所述装置还包括第二模型更新模块，具体用于：在确定将所述数据采集设备中的所述第一对象检测模型更新为所述第二对象检测模型的情况下，将同类数据采集设备中的所述第一对象检测模型更新为所述第二对象检测模型，其中，所述同类数据采集设备为：使用所述第一对象检测模型的数据采集设备；或
    所述模型训练模块，具体用于：利用所述第一样本数据和第二样本数据，对所述第一对象检测模型进行训练，得到第二对象检测模型，其中，所述第二样本数据为：用户通过客户端输入的多媒体数据；或
    所述装置还包括目标属性确定模块，具体用于通过以下方式获得所述目标属性：获得候选属性的出现频次，其中，所述候选属性为：所述第一对象检测模型对各多媒体数据进行对象检测的结果中置信度低于预设置信度阈值的对象的属性信息；确定出现频次达到预设的频次条件的候选属性，作为目标属性。
  12. 一种电子设备，其特征在于，包括处理器、通信接口、存储器和通信总线，其中，处理器，通信接口，存储器通过通信总线完成相互间的通信；
    存储器，用于存放计算机程序；
    处理器，用于执行存储器上所存放的程序时，实现权利要求1-8任一所述的方法步骤。
  13. 一种计算机可读存储介质，其特征在于，所述计算机可读存储介质内存储有计算机程序，所述计算机程序被处理器执行时实现权利要求1-8任一所述的方法步骤。
PCT/CN2021/118262 2020-09-24 2021-09-14 一种自训练方法、系统、装置、电子设备及存储介质 WO2022062968A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
EP21871339.4A EP4220479A4 (en) 2020-09-24 2021-09-14 METHOD, SYSTEM, APPARATUS, ELECTRONIC DEVICE AND RECORDING MEDIUM FOR SELF-TRAINING

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202011018277.9 2020-09-24
CN202011018277.9A CN112132220A (zh) 2020-09-24 2020-09-24 一种自训练方法、系统、装置、电子设备及存储介质

Publications (1)

Publication Number Publication Date
WO2022062968A1 true WO2022062968A1 (zh) 2022-03-31

Family

ID=73841111

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/118262 WO2022062968A1 (zh) 2020-09-24 2021-09-14 一种自训练方法、系统、装置、电子设备及存储介质

Country Status (3)

Country Link
EP (1) EP4220479A4 (zh)
CN (2) CN117786485A (zh)
WO (1) WO2022062968A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114880995A (zh) * 2022-06-30 2022-08-09 浙江大华技术股份有限公司 算法方案部署方法及相关装置、设备和存储介质

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117786485A (zh) * 2020-09-24 2024-03-29 杭州海康威视数字技术股份有限公司 一种自训练方法、系统、装置、电子设备及存储介质
CN118051828A (zh) * 2024-04-15 2024-05-17 腾讯科技（深圳）有限公司 模型训练方法、装置及存储介质

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190012575A1 (en) * 2017-07-04 2019-01-10 Beijing Baidu Netcom Science And Technology Co., Ltd. Method, apparatus and system for updating deep learning model
CN111210024A (zh) * 2020-01-14 2020-05-29 深圳供电局有限公司 模型训练方法、装置、计算机设备和存储介质
CN112132220A (zh) * 2020-09-24 2020-12-25 杭州海康威视数字技术股份有限公司 一种自训练方法、系统、装置、电子设备及存储介质

Family Cites Families (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7697026B2 (en) * 2004-03-16 2010-04-13 3Vr Security, Inc. Pipeline architecture for analyzing multiple video streams
CN108614858B (zh) * 2018-03-23 2019-07-05 北京达佳互联信息技术有限公司 图像分类模型优化方法、装置及终端
CN111310835B (zh) * 2018-05-24 2023-07-21 北京嘀嘀无限科技发展有限公司 目标对象的检测方法及装置
US10885370B2 (en) * 2018-12-16 2021-01-05 International Business Machines Corporation Selecting object detections or recognitions using correctness mappings
CN110046586A (zh) * 2019-04-19 2019-07-23 腾讯科技（深圳）有限公司 一种数据处理方法、设备及存储介质
CN110378420A (zh) * 2019-07-19 2019-10-25 Oppo广东移动通信有限公司 一种图像检测方法、装置及计算机可读存储介质
CN110598620B (zh) * 2019-09-06 2022-05-06 腾讯科技（深圳）有限公司 基于深度神经网络模型的推荐方法和装置
CN111538852B (zh) * 2020-04-23 2023-09-05 北京达佳互联信息技术有限公司 多媒体资源处理方法、装置、存储介质及设备

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190012575A1 (en) * 2017-07-04 2019-01-10 Beijing Baidu Netcom Science And Technology Co., Ltd. Method, apparatus and system for updating deep learning model
CN111210024A (zh) * 2020-01-14 2020-05-29 深圳供电局有限公司 模型训练方法、装置、计算机设备和存储介质
CN112132220A (zh) * 2020-09-24 2020-12-25 杭州海康威视数字技术股份有限公司 一种自训练方法、系统、装置、电子设备及存储介质

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of EP4220479A4 *

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114880995A (zh) * 2022-06-30 2022-08-09 浙江大华技术股份有限公司 算法方案部署方法及相关装置、设备和存储介质

Also Published As

Publication number Publication date
CN112132220A (zh) 2020-12-25
EP4220479A1 (en) 2023-08-02
EP4220479A4 (en) 2024-03-27
CN117786485A (zh) 2024-03-29

Similar Documents

Publication Publication Date Title
WO2022062968A1 (zh) 一种自训练方法、系统、装置、电子设备及存储介质
WO2020087974A1 (zh) 生成模型的方法和装置
EP3839942A1 (en) Quality inspection method, apparatus, device and computer storage medium for insurance recording
CN109447156B (zh) 用于生成模型的方法和装置
WO2017000109A1 (zh) 搜索方法、搜索装置、用户设备和计算机程序产品
CN110363220B (zh) 行为类别检测方法、装置、电子设备和计算机可读介质
CN111652087B (zh) 验车方法、装置、电子设备和存储介质
WO2015090215A1 (zh) 区分地域性口音的语音数据识别方法、装置和服务器
WO2021174812A1 (zh) 用于画像的数据的清洗方法、装置、介质及电子设备
WO2020029608A1 (zh) 用于检测电极片毛刺的方法和装置
US11501102B2 (en) Automated sound matching within an audio recording
CN111666898B (zh) 用于识别车辆所属类别的方法和装置
CN111885375A (zh) 双录视频的检验方法、装置、服务器及系统
CN111653274B (zh) 唤醒词识别的方法、装置及存储介质
CN113763348A (zh) 图像质量确定方法、装置、电子设备及存储介质
CN111984803B (zh) 多媒体资源处理方法、装置、计算机设备及存储介质
CN112766218A (zh) 基于非对称联合教学网络的跨域行人重识别方法和装置
WO2022062028A1 (zh) 酒标识别方法、酒品信息管理方法、装置、设备及存储介质
CN114881711A (zh) 基于请求行为进行异常分析的方法及电子设备
CN112989312B (zh) 验证码的识别方法、装置、电子设备和存储介质
CN111967450B (zh) 自动驾驶模型用样本获取方法、训练方法、装置及系统
CN112837640A (zh) 屏幕动态画面测试方法、系统、电子设备及存储介质
CN110826616A (zh) 信息处理方法及装置、电子设备、存储介质
WO2023045365A1 (zh) 视频质量评估方法、装置、电子设备及存储介质
CN110874554A (zh) 动作识别方法、终端设备、服务器、系统及存储介质

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: 21871339

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

ENP Entry into the national phase

Ref document number: 2021871339

Country of ref document: EP

Effective date: 20230424