CN110163265A - Data processing method, device and computer equipment - Google Patents
Data processing method, device and computer equipment Download PDFInfo
- Publication number
- CN110163265A CN110163265A CN201910362817.6A CN201910362817A CN110163265A CN 110163265 A CN110163265 A CN 110163265A CN 201910362817 A CN201910362817 A CN 201910362817A CN 110163265 A CN110163265 A CN 110163265A
- Authority
- CN
- China
- Prior art keywords
- feature
- sample
- candidate
- images
- data set
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; 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 OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
A kind of data processing method, device and computer equipment provided by the present application, difficult sample is excavated by the data set of multi-source, specifically from two class images in the first data set, screen target positive sample pair and the first negative sample pair, from in the first data set and the second data set in similar image, screen the second negative sample pair, the case where being intersected due to the image in the first data set and the second data set there is no identity, improve negative sample digging efficiency, be conducive to the promotion of human face recognition model performance, and then improve the recognition accuracy of human face recognition model.
Description
Technical Field
The present application relates to the field of face recognition technology, and in particular, to a data processing method, apparatus, and computer device.
Background
The face recognition is a biological recognition technology for performing identity recognition based on face feature information of a person, is widely applied to various video monitoring scenes at present, and in practical application, features of an image to be recognized are processed by using a pre-trained model to obtain identity information of the image to be recognized.
In the model training process, positive and negative samples are usually screened from a large number of images, the positive sample refers to two different images with the same identity, the negative sample refers to two images with different identities, then, the similarity of the positive sample is lower than a first threshold, the sample with the similarity of the negative sample higher than a second threshold is marked as a difficult sample, and then the difficult sample is input into a network model for training to obtain a model for identifying the identity of the image.
However, the inventor of the present application has noticed that the difficult samples required by the model training are often fixed and may possibly cause an over-fitting problem, which affects the accuracy of the output result of the model, i.e. reduces the accuracy of the face recognition result.
Disclosure of Invention
The embodiment of the application provides a data processing method, a data processing device and computer equipment, so that the excavation of a difficult sample of a multi-source data set is realized, the over-fitting problem is avoided, the excavation efficiency and the model performance of the difficult sample are improved, and the model identification accuracy is further improved.
In order to achieve the above purpose, the embodiments of the present application provide the following technical solutions:
the application provides a data processing method, which comprises the following steps:
acquiring a first data set and a second data set, wherein images corresponding to the same identity in the first data set at least comprise two types of images, and the identities of the images in the first data set are different from the identities of the images in the second data set;
screening a target positive sample pair and a first negative sample pair from the two types of images in the first data set;
screening a second negative sample pair from the same type of images in the first data set and the second data set, and combining the second negative sample pair and the first negative sample pair into a target negative sample pair;
and carrying out model training on the target positive sample pair and the target negative sample pair based on a neural network to obtain a face recognition model.
The present application also provides a data processing apparatus, the apparatus comprising:
the data acquisition module is used for acquiring a first data set and a second data set, wherein the images corresponding to the same identity in the first data set at least comprise two types of images, and the identities of the images in the first data set are different from the identities of the images in the second data set;
the first screening module is used for screening a target positive sample pair and a first negative sample pair from the two types of images in the first data set;
the second screening module is used for screening a second negative sample pair from the same type of images in the first data set and the second data set, and combining the second negative sample pair and the first negative sample pair into a target negative sample pair;
and the model training module is used for carrying out model training on the target positive sample pair and the target negative sample pair based on a neural network to obtain a face recognition model.
The present application further provides a computer device, comprising:
a communication interface;
a memory for storing a computer program implementing the data processing method as described above;
and the processor is used for recording and executing the computer program stored in the memory and realizing the steps of the data processing method.
The present application also provides a storage medium having stored thereon a computer program for execution by a processor to perform the steps of the data processing method as described above.
Based on the technical scheme, according to the data processing method, the data processing device and the computer equipment, the difficult samples are mined through the multi-source data set, the target positive sample pair and the first negative sample pair are specifically screened from two kinds of images in the first data set, the second negative sample pair is screened from the same kind of images in the first data set and the second data set, and due to the fact that the images in the first data set and the second data set do not have the condition of identity intersection, the negative sample mining efficiency is improved, the improvement of the performance of the face recognition model is facilitated, and the recognition accuracy of the face recognition model is further improved.
Drawings
In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced below, it is obvious that the drawings in the following description are only embodiments of the present application, and for those skilled in the art, other drawings can be obtained according to the provided drawings without creative efforts.
FIG. 1 is a block diagram of a data processing system according to an embodiment of the present application;
fig. 2 is a schematic flow chart illustrating a data processing method according to an embodiment of the present application;
FIG. 3 is a flow chart illustrating a difficult sample mining process in a data processing method according to an embodiment of the present application;
FIG. 4 is a flow chart illustrating another data processing method provided by an embodiment of the present application;
fig. 5 is a schematic structural diagram illustrating a data processing apparatus according to an embodiment of the present application;
FIG. 6 is a schematic structural diagram of another data processing apparatus provided in an embodiment of the present application;
FIG. 7 is a schematic structural diagram of another data processing apparatus provided in an embodiment of the present application;
fig. 8 shows a hardware structure diagram of a computer device according to an embodiment of the present application.
Detailed Description
In combination with the analysis of the above background art, in order to improve the accuracy of the model output result obtained by training, the inventor of the present application proposes a scheme for dynamically mining the difficult samples, that is, in the training process, the similarity between every two samples in the training samples used in the training of this time is calculated, and accordingly the difficult samples are determined, and the model training is completed. Compared with a fixed difficult sample, the dynamic difficult sample mining mode can adjust the distribution of the difficult sample to a certain extent, reduce the deviation in the training process, improve the reliability and accuracy of the model obtained by training and further improve the effect of the model in the actual scene.
However, the dynamic difficult sample mining method is limited to mining in a batch of training samples, the scale generated each time is small, and the samples are from a single data set, so that the mined difficult samples, especially the difficult negative samples, are often not real difficult samples, the effect on model training is small, and the improvement on network performance is limited, that is, the accuracy and reliability of the obtained model are not ideal when the model training is performed on the difficult samples obtained by the method.
In order to further improve the problems, more and more real difficult samples can be mined in one training, the inventor of the application proposes to expand the sample mining range, is not limited to one training batch any more, proposes to realize the mining of the difficult samples based on a plurality of data sets, and particularly mines the difficult negative samples so as to improve the mining efficiency of the difficult negative samples, improve the performance of the whole model, and further improve the accuracy and reliability of the output result of the model.
The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application, and it is obvious that the described embodiments are only a part of the embodiments of the present application, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present application.
Referring to fig. 1, in order to implement the system architecture of the data processing method provided in the present application, the system may include a server 11, a first data storage device 12, and a second data storage device 13, where:
the server 13 may be a service device providing a face recognition function, and in practical application, the server may receive an image to be recognized uploaded by a user side, and process the image to be recognized by using a pre-trained model to obtain identity information of the image to be recognized.
In the present application, the server 13 may be a single server, or may be a server set integrated by a plurality of servers, which is not limited in the present application.
The first data storage device 12 and the second data storage device 13 may be data storage devices for storing training data, and may specifically be a database server, and the like, and the specific product structure of the data storage devices is not limited in the present application.
In the present application, the first data storage device 12 may be an identity image for storing a large number of user identities, such as an image that can accurately indicate the user identity, for example, a certificate photo of a user, and the present application does not limit the manner of obtaining such an image, such as obtaining a face image on a user identity card from a public security system.
The second data storage device 13 may be configured to store an image obtained by terminal shooting, such as a self-shot by a camera, and the like.
Therefore, in the present application, the identity image in the first data storage device and the captured image in the second data storage device together form a training data set, and then a difficult sample required for model training is mined from the identity image in the first data storage device and the captured image in the second data storage device according to the data processing method described below, and the specific mining process may refer to the description of the corresponding part in the method embodiment described below, which is not described in detail in this embodiment.
It should be noted that, the components of the system architecture are not limited to the components given above, and may further include a client and the like as needed, and the details of the present application are not described herein.
Referring to fig. 2 in conjunction with the system architecture shown in fig. 1, there is a flowchart illustrating an embodiment of a data processing method provided in the present application, where the method may be applied to a service side, that is, the method may be implemented by being executed by a server, and as shown in fig. 2, the method may include, but is not limited to, the following steps:
step S101, screening a plurality of first identity images from a first data set to form a first candidate sample set, and screening a plurality of shot images to form a second candidate sample set;
step S102, screening a plurality of second identity images from a second data set to form a third candidate sample set;
in this embodiment, the first data set may be a face training data set, which may include identity images (i.e. ID photos, in particular, certificate images) of different identities, and a self-portrait image (i.e. a photographic image, such as Camera photos). The second data set may include identity images of a large number of identities, the scale of which may be in the millions, and the identity images in the second data set do not coincide with the identity images in the first data set, that is, the identities of the identity images in the two data sets do not intersect, and in order to distinguish the identity images in different data sets, the embodiment may mark the identity images in the first data set as the first identity images, and mark the identity images in the second data set as the second identity images.
The images in the first data set and the second data set can be obtained through business accumulation and data purchase, and the specific sources of the images are not limited in the application.
Based on this, in this embodiment, ID photos of partial identities may be randomly screened out from the first data set to form a first candidate sample set Sd, and meanwhile, photographed images of corresponding identities are screened out to form a second candidate sample set Sc; similarly, a plurality of identity images are randomly screened from the second data set to form a third candidate set Sn, and the screening process of the candidate samples of the three candidate sample sets is not described in detail in the present application.
Step S103, sequentially extracting the features of the respective candidate samples of the first candidate sample set, the second candidate sample set and the third candidate sample set to obtain a first feature set of the first candidate sample set, a second feature set of the second candidate sample set and a third feature set of the third candidate sample set;
in this embodiment, a deep convolutional neural network may be used to perform feature extraction on candidate samples (images), a specific implementation process is not described in detail in this embodiment, and an N-dimensional feature vector may be obtained for each candidate sample input to the convolutional neural network. It can be seen that each feature set obtained in this embodiment may be composed of a plurality of N-dimensional feature vectors, and the value of N may be determined based on the network structure of the convolutional neural network, and the specific value of the N is not limited in this application.
In the network training process, after a certain number of iterations, network parameters in the convolutional neural network can be updated, so that the convolutional neural network is optimized, and the accuracy of a feature extraction result is improved.
Through feature extraction of candidate samples included in each of the first candidate sample set, the second candidate sample set, and the third candidate sample set, the embodiment may record the corresponding obtained first feature set as Fd, second feature set as Fc, and third feature set as Fn.
Step S104, obtaining a target positive sample pair and a first negative sample pair according to the similarity between the feature vector in the first feature set and the feature vector in the second feature set;
in this embodiment, the positive and negative samples may be determined by calculating the similarity between feature vectors with the same identity in the first feature set and the second feature set, and the similarity between feature vectors without identities, for example, a candidate sample corresponding to a feature vector with a lower similarity in feature vectors with the same identity is determined as a positive sample P, and a candidate sample corresponding to a feature vector with a higher similarity in feature vectors with different identities is determined as a negative sample N1, which is not described in detail in this embodiment of the specific implementation process.
It should be understood that, in the present embodiment, when positive and negative samples are obtained, a candidate sample is respectively screened from two different candidate sample sets as a positive sample or a negative sample, so that each time a pair of candidate samples is actually screened, a pair of positive/negative samples is obtained.
Step S105, obtaining a second negative sample pair according to the similarity between the feature vector in the first feature set and the feature vector in the third feature set;
step S106, combining the first negative sample pair and the second negative sample pair into a target negative sample pair;
similar to the screening process of the first negative sample, the present embodiment may calculate the similarity between feature vectors included in the first feature set and the second feature set, and select a plurality of candidate samples corresponding to feature vectors with higher similarity as the second negative sample pair N2.
In combination with the above description of the process of acquiring the first candidate sample set, the second candidate sample set, and the third candidate sample set, the candidate samples in the first candidate sample set and the second candidate sample set are both from the face training data set, and specifically from different types of images in the face training data set, so that the candidate samples in the first candidate sample set are identity images of different users, and the candidate samples in the second candidate sample set are captured images of different users, and therefore, in the process of performing similarity on the feature vectors in the first feature set and the second feature set, the similarity between the feature vectors of different types of images of the same user and the similarity between the feature vectors of two types of images of different users are included.
Based on this, it should be understood that, for the same user, the higher the similarity between the feature vectors in the first feature set and the second feature set, the easier the identity of the candidate sample in the second candidate sample set is to be identified, and when a certain threshold is reached, the candidate sample can be considered as a different candidate sample of the same user; conversely, the lower the similarity, the more difficult the identity of the candidate sample in the second candidate sample set is to be identified, and the less likely the two candidate samples are to be the same identity. Therefore, in the process of mining the difficult samples, the present embodiment may use the candidate sample pair with the same identity and a lower similarity as the positive sample pair.
Similarly, for different users, the higher the similarity of the feature vectors from the two feature sets, the harder the identity of the candidate sample in the second candidate sample set is to be identified, but the easier the two candidate samples are to be identified as belonging to different identities, so that the candidate sample pair with the higher similarity in this case, i.e. different identities, can be regarded as a negative sample pair.
Similarly, because the third candidate sample set does not coincide with the identities of the candidate samples in the first candidate sample set and the first candidate sample set, and the feature vectors in the first feature set and the third feature set are actually the feature vectors corresponding to the identity images of different users, the candidate sample pair with higher similarity can be screened out as the negative sample pair through the similarity calculation of the feature vectors.
In summary, the target negative sample of the present embodiment is actually composed of two parts, and it can be known from the description of the source of each candidate sample set that the target negative sample actually comes from different data sets, that is, mining the difficult negative sample is realized through multiple source data sets, so that the screening range and scale of the difficult sample are expanded, the mining efficiency of the negative sample pair is improved, and the performance of the model is favorably improved.
And S107, respectively inputting the target positive sample pair and the target negative sample into a neural network for model training until the obtained positive sample similarity distribution and the obtained negative sample similarity distribution meet constraint conditions, and obtaining a face recognition model.
After the target positive/negative sample pairs are obtained according to the method, the target positive/negative sample pairs can be input to a neural network for model training to obtain the positive sample similarity distribution and the negative sample similarity distribution, wherein in the model training process, namely the continuous optimization process of the neural network model, the positive sample similarity distribution and the negative sample similarity distribution can be adjusted by using constraint conditions until the constraint conditions are met, and the content of the constraint conditions is not limited by the method.
The target positive/negative sample pairs are sequentially input into the neural network model for training, and the similarity of the obtained target positive sample pairs can be recorded as: spos＝(S1，S2，…，Sn) The similarity of the target negative sample pair can be recorded asAnd the positive sample similarity distribution is denoted as P (S)pos) Negative sample phaseSimilarity distribution is denoted as P (S)neg)。
Alternatively, for the optimization of the neural network model, it may be implemented by using a loss function, which may be defined as L ═ f (S)pos，Sneg) The form can be varied, such as defined directly asWherein,for averaging, α is an adjustable parameter, which represents the distance interval between the similarity of the target positive sample pair and the similarity of the target negative sample pair.
It should be noted that, regarding the model optimization method, the method is not limited to the above-mentioned loss function, and the gradient descent algorithm or various gradient-based variant optimization methods, such as adam and adagrad, may be used to continuously optimize the neural network structure until the constraint condition is satisfied. The model optimization method is not limited, and is not limited to the optimization methods listed herein, and the recognition accuracy of the face recognition model obtained through training can be improved and the false alarm rate can be reduced by continuously optimizing the model.
As another optional embodiment of the present application, for the constraint condition of the model training, the overlapping degree between the similarity distribution of the positive sample and the similarity distribution of the negative sample may be further obtained, and the similarity difference between the target positive sample pair and the target negative sample pair is pulled by reducing the overlapping degree region, so as to improve the model identification capability, where the overlapping degree may be obtained by an integral calculation method, which is not limited in the present application.
In summary, in the embodiment, candidate sample sets for different data sources are screened out through a multi-source data set, the candidate sample sets are not limited to data in a training batch, and the whole data set is concerned, meanwhile, in order to improve the scale of difficult negative sample pairs, the embodiment specifically screens the negative sample pairs for model training from the candidate sample sets with non-crossed identities, and compared with the prior art that the negative sample pairs are screened from the data set of a single source, the mining efficiency of the negative sample pairs is improved, the improvement of the performance of a face recognition model is facilitated, and the recognition accuracy of the face recognition model is further improved.
Based on the above description of the overall concept of the data processing method provided by the present application, the following mainly describes a difficult sample mining process, where the difficult samples may refer to a target positive sample pair and a target negative sample pair in the above embodiment, and refer to a flowchart shown in fig. 3. As shown in fig. 3, the excavation process of the difficult sample may include, but is not limited to, the following steps:
step S201, a first feature set of a first candidate sample set, a second feature set of a second candidate sample set and a third feature set of a third candidate sample set are obtained;
following the description of the above embodiment on the process of acquiring each candidate sample set, the candidate samples of the first candidate sample set are from ID photos of different identities in the face training data set, the candidate samples of the second candidate sample set are from self-photographs of different identities in the face training data set, the candidate samples of the third candidate sample set are from ID photos of different identities in the face registration base, and the ID photos in the face registration base and the ID photos in the face training data set do not intersect.
Step S202, calculating a first similarity between each feature vector in the first feature set and each feature vector in the second feature set;
for the similarity between the feature vectors, similarity algorithms such as cosine similarity calculation method and cosine distance can be used for realizing the similarity, and the specific similarity calculation method is not limited in the application.
Step S203, sorting according to the calculated first similarity, and determining a feature vector pair corresponding to the same identity in the first feature set and the second feature set and a feature vector pair corresponding to different identities according to a sorting result;
in this embodiment, if the similarity between two feature vectors (i.e., a feature vector pair) from different feature sets is greater than a certain threshold, the two feature vectors may be considered to represent the features of the same user image, that is, the candidate samples corresponding to the two feature vectors are images of the same user, so that it may be determined that the two feature vectors correspond to the same identity (i.e., the same user); conversely, the similarity of two feature vectors is smaller than the threshold, and the two feature vectors can be regarded as image feature vectors of different users, and the two feature vectors correspond to different identities. The present application does not limit the value of the threshold.
It should be noted that the method for confirming the identity corresponding to the feature vectors in the first feature set and the second feature set is not limited to the similarity meter algorithm method described above.
Step S204, selecting candidate sample pairs corresponding to a first number of feature vector pairs with smaller first similarity from the feature vector pairs corresponding to the same identity as target positive sample pairs;
step S205, selecting candidate sample pairs corresponding to a second number of feature vector pairs with larger first similarity from the feature vector pairs corresponding to different identities as first negative sample pairs;
in combination with the above explanations of positive and negative samples, the target positive sample pair is different candidate samples of the same identity, the target negative sample pair is candidate samples of different identities, and the first candidate sample set and the second candidate sample set are from the same face training data set, and both of them have candidate samples of the same identity and also have candidate samples of different identities.
Step S206, calculating a second similarity between each feature vector in the first feature set and each feature vector in the third feature set;
step S206 is similar to the similarity calculation method in step S202, and detailed description is omitted here.
Step S207, sorting according to the calculated second similarity, and selecting candidate sample pairs corresponding to a third number of eigenvector pairs with larger second similarity as second negative sample pairs;
in this embodiment, a specific implementation process of screening the target positive sample pair, the first negative sample pair, and the second negative sample pair based on the similarity ranking result is not described in detail, that is, specific values of the first number, the second number, and the third number are not limited, and may be preset values, at this time, the candidate sample pairs corresponding to the feature vector pairs of the corresponding number are directly screened according to the ranking result, or a similarity threshold value is preset, and the candidate sample pairs corresponding to the feature vector pairs whose similarity reaches the similarity threshold value are obtained by screening, at this time, the first number, the second number, and the third number may be determined based on the preset similarity threshold value.
It should be noted that, because the identities of the candidate samples corresponding to the feature vectors in the first feature set and the third feature set do not intersect, after the second similarity is obtained by calculation, the second negative sample pair can be screened directly according to the results sorted by the size of the similarity, and identity confirmation is not required.
And step S208, merging the first negative sample pair and the second negative sample pair to obtain a target negative sample pair.
Therefore, the target negative sample pair mined from the candidate samples comprises two parts, wherein one part of the target negative sample pair is derived from different identity sample images in the face training data set (namely the first data set), the other part of the target negative sample pair is derived from the face training data set and a sample image pair in the face registration base (namely the second data set), and the constructed sample pairs are negative sample pairs because the face training data set and the sample images in the face registration base do not have identity intersection.
Taking the first data set as the face training data set S1 and the second data set as the face registration base S2 as an example, the mining of the difficult samples and the training process of the face recognition model will be described below, and referring to the flow chart shown in fig. 4, the whole process may be divided into four stages, namely, a data reading stage, a feature extraction stage, a difficult sample mining stage, and a network training stage, where:
in the data reading stage, the face training data set S1 and the face registration database S2 can be accurately obtained, and a partial ID picture is randomly selected from S1 to form a first candidate sample set Sd, a corresponding self-photograph (i.e., Camera picture) to form a second candidate sample set Sc, and a partial ID picture is randomly selected from S2 to form a third candidate sample set Sn.
In the feature extraction stage, the feature extraction model may be used to extract respective image features of Sd, Sc, and Sn, where the feature extraction model may be obtained by using a face recognition model obtained through training, for example, by using a face recognition model obtained through optimization training, the feature extraction model is updated, the respective candidate samples of Sd, Sc, and Sn are respectively input into the updated feature extraction model, so as to obtain feature vectors corresponding to the respective candidate samples, that is, a first feature set Fd is formed by extracting image features (i.e., feature vectors) from the candidate samples in Sd, a second feature set Fc is formed by extracting image features (i.e., feature vectors) from the candidate samples in Sc, and a third feature set Fn is formed by extracting image features (i.e., feature vectors) from the candidate samples in Sn. The embodiment will not be described in detail about the specific implementation process of feature extraction, and the feature extraction model may be a convolutional neural network model.
In the difficult sample mining stage, as shown in fig. 4, the difficult sample mining may be performed in two parts, and the execution sequence of the two parts is not limited to the sequence described in the above embodiment, and may be adjusted as needed, or may be performed simultaneously. Specifically, the similarity of the feature vectors in Fd and Fc is calculated, from the feature vector pairs with the same identity, a candidate sample pair corresponding to a first number of feature vector pairs with lower similarity is screened as a target positive sample pair P, and from the feature vector pairs with different identities, a candidate sample pair corresponding to a second number of feature vector pairs with higher similarity is screened as a first negative sample pair N1; meanwhile, by calculating the similarity of the feature vectors in the Fd and the Fn, a third number of candidate sample pairs corresponding to feature vector pairs with higher similarity are screened as a second negative sample pair N2, and the sum of the first negative sample pair N1 and the second negative sample pair N2 is recorded as a target negative sample pair N.
In the network training stage, the difficult samples N and P obtained by mining are input into a neural network for processing, the similarity of P and N in the training is calculated, the similarity of a target negative sample pair is inhibited through the constraint of an objective function, the similarity of a target positive sample pair is improved, and the specific implementation process can refer to the description of the model training part of the embodiment.
In the continuous training of the model, whether the training result meets the constraint condition or not can be judged, if yes, the model obtained in the training is used as a face recognition module, and if not, the model obtained in the training can be used for replacing the feature extraction model for extracting the image features in the feature extraction stage. It should be understood that if the model obtained by the training is the same as the original feature extraction model, the original feature extraction model can be directly used to extract the image features without replacement.
Therefore, the method and the device have the advantages that the difficult positive and negative samples excavated through the multi-source data set are used for training the face recognition model, in the model training process, the model obtained by each training is used for realizing image feature extraction, the difficult sample excavation is further realized, the difficult sample excavation is not limited to a batch of training data, the excavation efficiency of the negative sample pair is improved by expanding the excavation range of the difficult sample, and the effects of improving the face recognition model performance and the recognition accuracy are achieved.
Referring to fig. 5, a schematic structural diagram of a data processing apparatus provided in an embodiment of the present application is a data processing apparatus, where the data processing apparatus may be applied to a server, and may specifically include, but is not limited to, the following virtual modules:
a data obtaining module 21, configured to obtain a first data set and a second data set, where images corresponding to a same identity in the first data set at least include two types of images, and an identity of an image in the first data set is different from an identity of an image in the second data set;
a first screening module 22, configured to screen a target positive sample pair and a first negative sample pair from the two types of images in the first data set;
a second screening module 23, configured to screen a second negative sample pair from the same type of image in the first data set and the second data set, and combine the second negative sample pair and the first negative sample pair into a target negative sample pair;
alternatively, as shown in fig. 6, the first screening module 22 may include:
a first filtering unit 221, configured to filter a plurality of first identity images from the first data set to form a first candidate sample set, and corresponding captured images to form a second candidate sample set;
a first feature extraction unit 222, configured to perform feature extraction on the first candidate sample set and the second candidate sample set respectively to obtain a first feature set of the first candidate sample set and a second feature set of the second candidate sample set;
a first similarity calculation unit 223, configured to obtain a target positive sample pair and a first negative sample pair according to a similarity between the feature vector in the first feature set and the feature vector in the second feature set;
accordingly, as shown in fig. 6, the second screening module 23 may include:
a second filtering unit 231, configured to filter a plurality of second identity images from the second data set to form a third candidate sample set;
the second feature extraction unit 232 is configured to perform feature extraction on the third candidate sample set to obtain a third feature set;
the second similarity calculation unit 233 is configured to obtain a second negative sample pair according to the similarity between the feature vector in the first feature set and the feature vector in the third feature set.
And the model training module 24 is configured to perform model training on the target positive sample pair and the target negative sample pair based on a neural network to obtain a face recognition model.
Optionally, the present embodiment may further include a model optimization module, configured to obtain the similarity distribution of the positive sample and the similarity distribution of the negative sample, and converge the similarity distribution of the positive sample and the similarity distribution of the negative sample, so as to adjust a model parameter of the face recognition model.
For the model optimization process, reference may be made to the description of the corresponding part of the above method embodiment, which is not described in detail in this embodiment.
Therefore, in the face recognition model training process, the difficult samples in the multi-source data set are dynamically constructed, the similarity distance between the difficult positive sample pair and the negative sample pair is adjusted, the similarity of the negative sample pair is restrained, the similarity of the positive sample pair is improved, and further the model recognition capability is improved. When the difficult negative sample pair is obtained, the data sets of different sources without cross identities are directly constructed, the cardinality of the negative sample pair is greatly improved, screening of the difficult negative sample pair is facilitated, the mining efficiency of the negative sample pair is improved, and then the model training efficiency is improved.
Optionally, as shown in fig. 7, the apparatus may further include:
the model updating module 25 is configured to update the feature extraction model by using the model obtained through the training when the model obtained through the training does not meet the constraint condition;
correspondingly, the first screening module 22 or the second screening module 23 is specifically configured to input the candidate sample into the updated feature extraction model to obtain a feature vector, and form a corresponding feature set from the feature vectors corresponding to the candidate samples belonging to the same candidate sample set, where the candidate sample is a candidate sample in the first candidate sample set, the second candidate sample set, or the third candidate sample set.
For the extraction process of the image features of the candidate sample, reference may be made to the description of the above method embodiment, and this embodiment is not repeated.
The embodiment of the present application further provides a storage medium, where a computer program is stored, where the computer program is executed by a processor to implement the steps of the data processing method, and as for a specific flow of the data processing method, reference may be made to the description of the corresponding part in the embodiment of the method, which is not described in detail in this embodiment.
An embodiment of the present application further provides a computer device, where a hardware structure of the computer device may be as shown in fig. 8, and in practical application, the computer device may be a server, and the hardware structure of the computer device may include: a communication interface 31, a memory 32, and a processor 33;
in the embodiment of the present application, the communication interface 31, the memory 32, and the processor 33 may implement communication with each other through a communication bus, and the number of the communication interface 31, the memory 32, the processor 33, and the communication bus may be at least one.
Optionally, the communication interface 31 may be an interface of a communication module, such as an interface of a GSM module, an interface of a WIFI module, and/or a serial port, and may be used to obtain data of other devices, such as a face image, and may also be used to implement data interaction between internal components of a computer device, and the type of the interface included in the communication interface 31 is not limited in this application.
The processor 33 may be a central processing unit CPU, or an application specific Integrated circuit asic, or one or more Integrated circuits configured to implement embodiments of the present application.
The memory 32 may comprise high-speed RAM memory and may also include non-volatile memory (non-volatile memory), such as at least one disk memory.
Wherein, the memory 32 stores a computer program, and the processor 33 calls the computer program stored in the memory 2 to realize the steps of the data processing method applied to the computer device;
optionally, the computer program is primarily operable to:
acquiring a first data set and a second data set, wherein images corresponding to the same identity in the first data set at least comprise two types of images, and the identities of the images in the first data set are different from the identities of the images in the second data set;
screening a target positive sample pair and a first negative sample pair from the two types of images in the first data set;
screening a second negative sample pair from the same type of images in the first data set and the second data set, and combining the second negative sample pair and the first negative sample pair into a target negative sample pair;
and carrying out model training on the target positive sample pair and the target negative sample pair based on a neural network to obtain a face recognition model.
Optionally, the processor executing the computer program may be specifically configured to:
screening a plurality of first identity images from the first data set to form a first candidate sample set, and enabling corresponding shot images to form a second candidate sample set;
respectively performing feature extraction on the first candidate sample set and the second candidate sample set to obtain a first feature set of the first candidate sample set and a second feature set of the second candidate sample set;
obtaining a target positive sample pair and a first negative sample pair according to the similarity between the feature vector in the first feature set and the feature vector in the second feature set;
screening a plurality of second identity images from the second data set to form a third candidate sample set;
performing feature extraction on the third candidate sample set to obtain a third feature set;
and obtaining a second negative sample pair according to the similarity between the feature vector in the first feature set and the feature vector in the third feature set.
Optionally, the processor executing the computer program may be further configured to:
in the model training process, if the model obtained by the training does not meet the constraint condition, updating the feature extraction model by using the model obtained by the training;
inputting a candidate sample into the updated feature extraction model to obtain a feature vector, wherein the candidate sample is a candidate sample in the first candidate sample set, the second candidate sample set or the third candidate sample set;
and forming a corresponding feature set by the feature vectors corresponding to the candidate samples belonging to the same candidate sample set.
Optionally, the processor executing the computer program may be specifically configured to:
calculating a first similarity between the feature vectors in the first feature set and the feature vectors in the second feature set;
selecting candidate sample pairs corresponding to a first number of feature vector pairs with smaller first similarity from feature vector pairs with the same identity as target positive sample pairs, and selecting candidate sample pairs corresponding to a second number of feature vector pairs with larger first similarity from feature vector pairs with different identities as first negative sample pairs;
calculating a second similarity between the feature vectors in the first feature set and the feature vectors in the third feature set;
and selecting candidate sample pairs corresponding to a third number of feature vector pairs with larger second similarity as second negative sample pairs.
It should be noted that, regarding the specific processes of the data processing method implemented by the computer program executed by the processor, reference may be made to the descriptions of the corresponding parts of the above method embodiments, and the steps described in the embodiment are not limited.
The embodiments in the present description are described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same and similar parts among the embodiments are referred to each other. The device and the computer equipment disclosed by the embodiment correspond to the method disclosed by the embodiment, so that the description is relatively simple, and the relevant points can be referred to the method part for description.
Those of skill would further appreciate that the various illustrative elements and algorithm steps described in connection with the embodiments disclosed herein may be implemented as electronic hardware, computer software, or combinations of both, and that the various illustrative components and steps have been described above generally in terms of their functionality in order to clearly illustrate this interchangeability of hardware and software. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the implementation. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present application.
The steps of a method or algorithm described in connection with the embodiments disclosed herein may be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. A software module may reside in Random Access Memory (RAM), memory, Read Only Memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.
The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the disclosure. Thus, the present application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims (10)
1. A method of data processing, the method comprising:
acquiring a first data set and a second data set, wherein images corresponding to the same identity in the first data set at least comprise two types of images, and the identities of the images in the first data set are different from the identities of the images in the second data set;
screening a target positive sample pair and a first negative sample pair from the two types of images in the first data set;
screening a second negative sample pair from the same type of images in the first data set and the second data set, and combining the second negative sample pair and the first negative sample pair into a target negative sample pair;
and carrying out model training on the target positive sample pair and the target negative sample pair based on a neural network to obtain a face recognition model.
2. The method of claim 1, wherein the screening a target positive sample pair and a first negative sample pair from two types of images in the first data set comprises:
screening a plurality of first identity images from the first data set to form a first candidate sample set, and enabling corresponding shot images to form a second candidate sample set;
respectively performing feature extraction on the first candidate sample set and the second candidate sample set to obtain a first feature set of the first candidate sample set and a second feature set of the second candidate sample set;
and obtaining a target positive sample pair and a first negative sample pair according to the similarity between the feature vector in the first feature set and the feature vector in the second feature set.
3. The method of claim 2, wherein said screening a second negative sample pair from homogeneous images in said first data set and said second data set comprises:
screening a plurality of second identity images from the second data set to form a third candidate sample set;
performing feature extraction on the third candidate sample set to obtain a third feature set;
and obtaining a second negative sample pair according to the similarity between the feature vector in the first feature set and the feature vector in the third feature set.
4. The method of claim 2, further comprising:
in the model training process, if the model obtained by the training does not meet the constraint condition, updating the feature extraction model by using the model obtained by the training;
inputting a candidate sample into the updated feature extraction model to obtain a feature vector, wherein the candidate sample is a candidate sample in the first candidate sample set, the second candidate sample set or the third candidate sample set;
and forming a corresponding feature set by the feature vectors corresponding to the candidate samples belonging to the same candidate sample set.
5. The method of claim 3, wherein obtaining the target positive sample pair and the first negative sample pair according to the similarity between the feature vector in the first feature set and the feature vector in the second feature set comprises:
calculating a first similarity between the feature vectors in the first feature set and the feature vectors in the second feature set;
selecting candidate sample pairs corresponding to a first number of feature vector pairs with smaller first similarity from feature vector pairs with the same identity as target positive sample pairs, and selecting candidate sample pairs corresponding to a second number of feature vector pairs with larger first similarity from feature vector pairs with different identities as first negative sample pairs;
obtaining a second negative sample pair according to the similarity between the feature vector in the first feature set and the feature vector in the third feature set, including:
calculating a second similarity between the feature vectors in the first feature set and the feature vectors in the third feature set;
and selecting candidate sample pairs corresponding to a third number of feature vector pairs with larger second similarity as second negative sample pairs.
6. The method of claim 3, wherein the identity image comprises a certificate photo of the corresponding user, and wherein the captured image comprises a self-photograph of the corresponding user.
7. A data processing apparatus, characterized in that the apparatus comprises:
the data acquisition module is used for acquiring a first data set and a second data set, wherein the images corresponding to the same identity in the first data set at least comprise two types of images, and the identities of the images in the first data set are different from the identities of the images in the second data set;
the first screening module is used for screening a target positive sample pair and a first negative sample pair from the two types of images in the first data set;
the second screening module is used for screening a second negative sample pair from the same type of images in the first data set and the second data set, and combining the second negative sample pair and the first negative sample pair into a target negative sample pair;
and the model training module is used for carrying out model training on the target positive sample pair and the target negative sample pair based on a neural network to obtain a face recognition model.
8. The apparatus of claim 7, wherein the first screening module comprises:
a first screening unit for screening a plurality of first identity images from the first data set to form a first candidate sample set, and corresponding captured images to form a second candidate sample set;
the first feature extraction unit is used for respectively performing feature extraction on the first candidate sample set and the second candidate sample set to obtain a first feature set of the first candidate sample set and a second feature set of the second candidate sample set;
the first similarity calculation unit is used for obtaining a target positive sample pair and a first negative sample pair according to the similarity between the feature vector in the first feature set and the feature vector in the second feature set;
the second screening module includes:
a second screening unit, configured to screen a plurality of second identity images from the second data set to form a third candidate sample set;
the second feature extraction unit is used for performing feature extraction on the third candidate sample set to obtain a third feature set;
and the second similarity calculation unit is used for obtaining a second negative sample pair according to the similarity between the feature vector in the first feature set and the feature vector in the third feature set.
9. The apparatus of claim 8, wherein during model training, the apparatus further comprises:
the model updating module is used for updating the feature extraction model by using the model obtained by the training under the condition that the model obtained by the training does not meet the constraint condition;
the first screening module or the second screening module is specifically configured to input a candidate sample into the updated feature extraction model to obtain a feature vector, and form a corresponding feature set from the feature vectors corresponding to candidate samples belonging to the same candidate sample set, where the candidate sample is a candidate sample in the first candidate sample set, the second candidate sample set, or the third candidate sample set.
10. A computer device, characterized in that the computer device comprises:
a communication interface;
a memory for storing a computer program for implementing the data processing method of any one of claims 1 to 6;
a processor for recording and executing the computer program stored in the memory, implementing the steps of the data processing method according to any one of claims 1 to 6.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910362817.6A CN110163265B (en) | 2019-04-30 | 2019-04-30 | Data processing method and device and computer equipment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910362817.6A CN110163265B (en) | 2019-04-30 | 2019-04-30 | Data processing method and device and computer equipment |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110163265A true CN110163265A (en) | 2019-08-23 |
| CN110163265B CN110163265B (en) | 2024-08-09 |
Family
ID=67633335
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910362817.6A Active CN110163265B (en) | 2019-04-30 | 2019-04-30 | Data processing method and device and computer equipment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110163265B (en) |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110705614A (en) * | 2019-09-20 | 2020-01-17 | 北京三快在线科技有限公司 | Model training method and device, electronic equipment and storage medium |
| CN113469204A (en) * | 2020-03-31 | 2021-10-01 | 日本电气株式会社 | Data processing method, device, equipment and computer storage medium |
| CN113706502A (en) * | 2021-08-26 | 2021-11-26 | 重庆紫光华山智安科技有限公司 | Method and device for evaluating quality of face image |
| CN113761105A (en) * | 2021-05-24 | 2021-12-07 | 腾讯科技（深圳）有限公司 | Text data processing method, apparatus, device and medium |
| CN113947701A (en) * | 2021-10-18 | 2022-01-18 | 北京百度网讯科技有限公司 | Training method, object recognition method, device, electronic device and storage medium |
| CN114020974A (en) * | 2022-01-04 | 2022-02-08 | 阿里巴巴达摩院(杭州)科技有限公司 | Sample data determination and dialogue intention identification method, storage medium, and program product |
| CN114764942A (en) * | 2022-05-20 | 2022-07-19 | 清华大学深圳国际研究生院 | Difficult positive and negative sample online mining method and face recognition method |
| CN115761434A (en) * | 2022-10-24 | 2023-03-07 | 北京百度网讯科技有限公司 | Training method, device, electronic equipment and storage medium of target detection model |
| CN117095469A (en) * | 2023-08-07 | 2023-11-21 | 中国工商银行股份有限公司 | Identity recognition method, model training method, device, equipment and storage medium |
| CN118351399A (en) * | 2024-06-13 | 2024-07-16 | 比亚迪股份有限公司 | Sample generation method, image recognition model training method and corresponding device |
| CN118447268A (en) * | 2023-09-22 | 2024-08-06 | 荣耀终端有限公司 | Identity recognition method, electronic device and storage medium |
| CN120071053A (en) * | 2025-02-08 | 2025-05-30 | 苏州天准科技股份有限公司 | Sample processing method and processing system |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105138523A (en) * | 2014-05-30 | 2015-12-09 | 富士通株式会社 | Method and device for determining semantic keywords in text |
| CN106951484A (en) * | 2017-03-10 | 2017-07-14 | 百度在线网络技术（北京）有限公司 | Picture retrieval method and device, computer equipment and computer-readable medium |
| WO2018000731A1 (en) * | 2016-06-28 | 2018-01-04 | 华南理工大学 | Method for automatically detecting curved surface defect and device thereof |
| CN107944020A (en) * | 2017-12-11 | 2018-04-20 | 深圳云天励飞技术有限公司 | Facial image lookup method and device, computer installation and storage medium |
| WO2018137357A1 (en) * | 2017-01-24 | 2018-08-02 | 北京大学 | Target detection performance optimization method |
| CN108596277A (en) * | 2018-05-10 | 2018-09-28 | 腾讯科技（深圳）有限公司 | A kind of testing vehicle register identification method, apparatus and storage medium |
| CN108805048A (en) * | 2018-05-25 | 2018-11-13 | 腾讯科技（深圳）有限公司 | A kind of method of adjustment of human face recognition model, device and storage medium |
| WO2019041519A1 (en) * | 2017-08-29 | 2019-03-07 | 平安科技（深圳）有限公司 | Target tracking device and method, and computer-readable storage medium |
| WO2019056470A1 (en) * | 2017-09-19 | 2019-03-28 | 平安科技（深圳）有限公司 | Driving model training method, driver recognition method and apparatus, device, and medium |
-
2019
- 2019-04-30 CN CN201910362817.6A patent/CN110163265B/en active Active
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105138523A (en) * | 2014-05-30 | 2015-12-09 | 富士通株式会社 | Method and device for determining semantic keywords in text |
| WO2018000731A1 (en) * | 2016-06-28 | 2018-01-04 | 华南理工大学 | Method for automatically detecting curved surface defect and device thereof |
| WO2018137357A1 (en) * | 2017-01-24 | 2018-08-02 | 北京大学 | Target detection performance optimization method |
| CN106951484A (en) * | 2017-03-10 | 2017-07-14 | 百度在线网络技术（北京）有限公司 | Picture retrieval method and device, computer equipment and computer-readable medium |
| WO2019041519A1 (en) * | 2017-08-29 | 2019-03-07 | 平安科技（深圳）有限公司 | Target tracking device and method, and computer-readable storage medium |
| WO2019056470A1 (en) * | 2017-09-19 | 2019-03-28 | 平安科技（深圳）有限公司 | Driving model training method, driver recognition method and apparatus, device, and medium |
| CN107944020A (en) * | 2017-12-11 | 2018-04-20 | 深圳云天励飞技术有限公司 | Facial image lookup method and device, computer installation and storage medium |
| CN108596277A (en) * | 2018-05-10 | 2018-09-28 | 腾讯科技（深圳）有限公司 | A kind of testing vehicle register identification method, apparatus and storage medium |
| CN108805048A (en) * | 2018-05-25 | 2018-11-13 | 腾讯科技（深圳）有限公司 | A kind of method of adjustment of human face recognition model, device and storage medium |
Non-Patent Citations (3)
| Title |
|---|
| 孙玉;刘贵全;汪中;: "基于不平衡分类的人脸检测系统", 计算机应用与软件, no. 12, 15 December 2012 (2012-12-15) * |
| 张星明;李和恒;: "基于独立负样本集和SVM的人脸确认算法", 计算机研究与发展, no. 12, 28 December 2006 (2006-12-28) * |
| 王大伟;陈章玲;: "基于LBP与卷积神经网络的人脸识别", 天津理工大学学报, no. 06, 15 December 2017 (2017-12-15) * |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110705614A (en) * | 2019-09-20 | 2020-01-17 | 北京三快在线科技有限公司 | Model training method and device, electronic equipment and storage medium |
| CN113469204A (en) * | 2020-03-31 | 2021-10-01 | 日本电气株式会社 | Data processing method, device, equipment and computer storage medium |
| CN113761105A (en) * | 2021-05-24 | 2021-12-07 | 腾讯科技（深圳）有限公司 | Text data processing method, apparatus, device and medium |
| CN113706502A (en) * | 2021-08-26 | 2021-11-26 | 重庆紫光华山智安科技有限公司 | Method and device for evaluating quality of face image |
| CN113706502B (en) * | 2021-08-26 | 2023-09-05 | 重庆紫光华山智安科技有限公司 | Face image quality assessment method and device |
| CN113947701B (en) * | 2021-10-18 | 2024-02-23 | 北京百度网讯科技有限公司 | Training method, object recognition method, device, electronic equipment and storage medium |
| CN113947701A (en) * | 2021-10-18 | 2022-01-18 | 北京百度网讯科技有限公司 | Training method, object recognition method, device, electronic device and storage medium |
| CN114020974A (en) * | 2022-01-04 | 2022-02-08 | 阿里巴巴达摩院(杭州)科技有限公司 | Sample data determination and dialogue intention identification method, storage medium, and program product |
| CN114020974B (en) * | 2022-01-04 | 2022-06-21 | 阿里巴巴达摩院(杭州)科技有限公司 | Sample data determination and dialogue intent recognition method, storage medium and program product |
| CN114764942A (en) * | 2022-05-20 | 2022-07-19 | 清华大学深圳国际研究生院 | Difficult positive and negative sample online mining method and face recognition method |
| CN115761434A (en) * | 2022-10-24 | 2023-03-07 | 北京百度网讯科技有限公司 | Training method, device, electronic equipment and storage medium of target detection model |
| CN117095469A (en) * | 2023-08-07 | 2023-11-21 | 中国工商银行股份有限公司 | Identity recognition method, model training method, device, equipment and storage medium |
| CN118447268A (en) * | 2023-09-22 | 2024-08-06 | 荣耀终端有限公司 | Identity recognition method, electronic device and storage medium |
| CN118447268B (en) * | 2023-09-22 | 2025-04-15 | 荣耀终端股份有限公司 | Identity recognition method, electronic device and storage medium |
| CN118351399A (en) * | 2024-06-13 | 2024-07-16 | 比亚迪股份有限公司 | Sample generation method, image recognition model training method and corresponding device |
| CN118351399B (en) * | 2024-06-13 | 2025-05-13 | 比亚迪股份有限公司 | Sample generation method, training method of image recognition model and corresponding devices |
| CN120071053A (en) * | 2025-02-08 | 2025-05-30 | 苏州天准科技股份有限公司 | Sample processing method and processing system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110163265B (en) | 2024-08-09 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110163265B (en) | Data processing method and device and computer equipment | |
| CN108961303B (en) | Image processing method and device, electronic equipment and computer readable medium | |
| CN108269254B (en) | Image quality evaluation method and device | |
| US9036905B2 (en) | Training classifiers for deblurring images | |
| EP3989158A1 (en) | Method, apparatus and device for video similarity detection | |
| CN111667001B (en) | Target re-identification method, device, computer equipment and storage medium | |
| Villalba et al. | Smartphone image clustering | |
| CN110941978B (en) | Face clustering method and device for unidentified personnel and storage medium | |
| CN111079816A (en) | Image auditing method and device and server | |
| CN112132279A (en) | Convolutional neural network model compression method, device, equipment and storage medium | |
| TW201447775A (en) | Method and system for recognizing information | |
| CN111882578B (en) | Foreground image acquisition method, foreground image acquisition device and electronic device | |
| CN112381104A (en) | Image identification method and device, computer equipment and storage medium | |
| CN111027450A (en) | Bank card information identification method and device, computer equipment and storage medium | |
| CN112116556A (en) | Passenger flow volume statistical method and device and computer equipment | |
| CN113408332A (en) | Video mirror splitting method, device, equipment and computer readable storage medium | |
| CN103353881A (en) | Method and device for searching application | |
| CN105516735A (en) | Representation frame acquisition method and representation frame acquisition apparatus | |
| CN109165572B (en) | Method and apparatus for generating information | |
| CN115393761A (en) | Video key frame extraction method, device, equipment and storage medium | |
| WO2022116668A1 (en) | Video filtering method and system based on identical content, and device | |
| CN117173232A (en) | Depth image acquisition method, device and equipment | |
| CN112749660A (en) | Method and equipment for generating video content description information | |
| CN110889438B (en) | Image processing method and device, electronic equipment and storage medium | |
| CN112883956A (en) | Text character recognition method, device and server |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |