WO2021018189A1 - 进行对象识别的方法和装置 - Google Patents

进行对象识别的方法和装置 Download PDF

Info

Publication number
WO2021018189A1
WO2021018189A1 PCT/CN2020/105500 CN2020105500W WO2021018189A1 WO 2021018189 A1 WO2021018189 A1 WO 2021018189A1 CN 2020105500 W CN2020105500 W CN 2020105500W WO 2021018189 A1 WO2021018189 A1 WO 2021018189A1
Authority
WO
WIPO (PCT)
Prior art keywords
feature
loss value
extraction model
model
extracted
Prior art date
Application number
PCT/CN2020/105500
Other languages
English (en)
French (fr)
Inventor
翟中华
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 EP20846719.1A priority Critical patent/EP4006775A4/en
Priority to US17/631,446 priority patent/US20220277588A1/en
Publication of WO2021018189A1 publication Critical patent/WO2021018189A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00Arrangements for image or video recognition or understanding
    • G06V10/70Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/74Image or video pattern matching; Proximity measures in feature spaces
    • G06V10/75Organisation of the matching processes, e.g. simultaneous or sequential comparisons of image or video features; Coarse-fine approaches, e.g. multi-scale approaches; using context analysis; Selection of dictionaries
    • G06V10/757Matching configurations of points or features
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16Human faces, e.g. facial parts, sketches or expressions
    • G06V40/172Classification, e.g. identification
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/21Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/22Matching criteria, e.g. proximity measures
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00Arrangements for image or video recognition or understanding
    • G06V10/70Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77Processing 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
    • G06V10/771Feature selection, e.g. selecting representative features from a multi-dimensional feature space
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16Human faces, e.g. facial parts, sketches or expressions
    • G06V40/161Detection; Localisation; Normalisation
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16Human faces, e.g. facial parts, sketches or expressions
    • G06V40/168Feature extraction; Face representation
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/12Fingerprints or palmprints
    • YGENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02TCLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00Road transport of goods or passengers
    • Y02T10/10Internal combustion engine [ICE] based vehicles
    • Y02T10/40Engine management systems

Definitions

  • the present disclosure relates to the field of computer technology, and in particular to a method and device for object recognition.
  • object recognition has been deployed in many fields. For example, face recognition, fingerprint recognition, voiceprint recognition, etc.
  • an extraction model is generally used to extract the features of the object to be recognized, and then the extracted features are compared with the features of each object in the matching library to determine the feature with the highest similarity.
  • the object to which the feature with the highest similarity belongs is determined as the matching object of the object to be identified, and the identity information of the matching object is determined as the identity information of the object to be identified.
  • the extraction model will be updated, and the features in the matching library are still extracted using the low version extraction model. It may happen that the features extracted by the high version extraction model cannot match the features of the matching library. At this time , It is necessary to use a higher version of the extraction model to re-extract the features of each object from the samples to form a matching library. However, when the number of samples (such as pictures, fingerprints, audio) is large, the amount of data involved in the extraction process is It will be larger and will take a lot of time.
  • the embodiments of the present disclosure provide a method and device for object recognition, which can at least solve the problem that it takes a long time to re-extract features in a matching library after upgrading a model.
  • the technical solution is as follows:
  • a method for object recognition includes:
  • the object to which the target feature belongs is determined as a matching object of the target object.
  • the method further includes:
  • the first initial feature conversion model determines the first Feature conversion model.
  • the method further includes:
  • the first feature conversion model includes:
  • the first loss value the second loss value, the first initial feature conversion model, the features extracted by the first extraction model of the multiple sample objects, and the features extracted by the second extraction model Features, determining the first feature conversion model.
  • the method further includes:
  • the extracted features and determining the first feature conversion model include:
  • the first feature conversion model is determined for each sample object with features extracted through the first extraction model and features extracted through the second extraction model.
  • the first loss value, the second loss value, the third loss value, and the fourth loss value are all loss values obtained from L1 loss; or,
  • the first loss value, the second loss value, the third loss value, and the fourth loss value are all loss values obtained from L2 loss; or,
  • the first loss value and the third loss value are loss values obtained by L1 loss
  • the second loss value and the fourth loss value are loss values obtained by L2 loss.
  • a method for object recognition includes:
  • the object to which the target feature belongs is determined as a matching object of the target object.
  • the method further includes:
  • the second initial feature conversion model, the features extracted by the first extraction model of the multiple sample objects, and the features extracted by the second extraction model determine the second Feature conversion model.
  • the method further includes:
  • the second feature conversion model includes:
  • the fifth loss value, the sixth loss value, the second initial feature conversion model, the features extracted by the first extraction model of the multiple sample objects, and the features extracted by the second extraction model Features determining the second feature conversion model.
  • the method further includes:
  • the extracted features and determining the second feature conversion model include:
  • the fifth loss value, the sixth loss value, the seventh loss value, the eighth loss value, the first initial feature conversion model, the second initial feature conversion model, and the multiple sample objects The features extracted by the first extraction model and the features extracted by the second extraction model are used to determine the second feature conversion model.
  • the fifth loss value, the sixth loss value, the seventh loss value, and the eighth loss value are all loss values obtained from L1 loss; or,
  • the fifth loss value, the sixth loss value, the seventh loss value, and the eighth loss value are all loss values obtained from L2 loss; or,
  • the fifth loss value and the seventh loss value are loss values obtained by L1 loss
  • the sixth loss value and the eighth loss value are loss values obtained by L2 loss.
  • a device for object recognition includes:
  • the conversion module is used to convert the first feature extracted by the first extraction model to the feature space of the second extraction model through the first feature conversion model to obtain the second feature of the target object in the feature space ;
  • the object to which the target feature belongs is determined as a matching object of the target object.
  • the device further includes a training module for:
  • the first initial feature conversion model determines the first Feature conversion model.
  • the training module is also used for:
  • the training module is used for:
  • the first loss value the second loss value, the first initial feature conversion model, the features extracted by the first extraction model of the multiple sample objects, and the features extracted by the second extraction model Features, determining the first feature conversion model.
  • the training module is also used for:
  • the training module is used for:
  • the first feature conversion model is determined for each sample object with features extracted through the first extraction model and features extracted through the second extraction model.
  • the first loss value, the second loss value, the third loss value, and the fourth loss value are all loss values obtained from L1 loss; or,
  • the first loss value, the second loss value, the third loss value, and the fourth loss value are all loss values obtained from L2 loss; or,
  • the first loss value and the third loss value are loss values obtained by L1 loss
  • the second loss value and the fourth loss value are loss values obtained by L2 loss.
  • a device for object recognition includes:
  • the conversion module is used to convert each object in the matching library using the third feature extracted by the second extraction model to the feature space of the first extraction model through the second feature conversion model to obtain that each object is in the feature space
  • the fourth feature
  • the object to which the target feature belongs is determined as a matching object of the target object.
  • the device further includes a training module for:
  • the second feature conversion model is determined according to the fifth loss value, the second initial feature conversion model, the features extracted by the first extraction model of the multiple sample objects, and the features extracted by the second extraction model.
  • the training module is also used for:
  • the training module is used for:
  • the fifth loss value, the sixth loss value, the second initial feature conversion model, the features extracted by the first extraction model of the multiple sample objects, and the features extracted by the second extraction model determine the The second feature conversion model.
  • the training module is also used for:
  • the training module is used for:
  • the fifth loss value, the sixth loss value, the seventh loss value, the eighth loss value, the first initial feature conversion model, the second initial feature conversion model, and the multiple sample objects The features extracted by the first extraction model and the features extracted by the second extraction model are used to determine the second feature conversion model.
  • the fifth loss value, the sixth loss value, the seventh loss value, and the eighth loss value are all loss values obtained from L1 loss; or,
  • the fifth loss value, the sixth loss value, the seventh loss value, and the eighth loss value are all loss values obtained from L2 loss; or,
  • the fifth loss value and the seventh loss value are loss values obtained by L1 loss
  • the sixth loss value and the eighth loss value are loss values obtained by L2 loss.
  • a computer-readable storage medium stores a computer program.
  • the computer program When executed by a processor, it implements the method for object recognition in the first and second aspects .
  • a server for object recognition includes a processor and a memory.
  • the memory is used for storing computer programs; the processor is used for executing the programs stored on the memory. , To implement the method for object recognition described in the first aspect and the second aspect.
  • the first feature extracted by the first extraction model of the target object is converted to the feature space of the second extraction model through the first feature conversion model to obtain the second feature of the target object in the feature space, and the target The second feature of the object is matched with the feature of the object in the matching library to obtain the target feature that matches the second feature of the target object, and the object to which the target feature belongs is determined as the matching object of the target object.
  • the features extracted by the first extraction model cannot directly match the features extracted by the second extraction model
  • the features extracted by the first extraction model can be converted to the feature space of the second extraction model, and the features extracted by the first extraction model
  • the features of can be matched with the features extracted by the second extraction model. Since the processing amount of feature conversion is much less than re-extracting the features in the matching library, the time for object recognition can be saved.
  • FIG. 1 is a schematic diagram of object recognition provided by an embodiment of the present disclosure
  • FIG. 2 is a flowchart of a method for object recognition provided by an embodiment of the present disclosure
  • Figure 3 is a schematic diagram of a training provided by an embodiment of the present disclosure.
  • FIG. 4 is a schematic diagram of a training provided by an embodiment of the present disclosure.
  • FIG. 5 is a flowchart of a method for object recognition provided by an embodiment of the present disclosure.
  • FIG. 6 is a flowchart of a method for object recognition provided by an embodiment of the present disclosure.
  • FIG. 7 is a schematic structural diagram of a device for object recognition provided by an embodiment of the present disclosure.
  • FIG. 8 is a schematic structural diagram of a device for object recognition provided by an embodiment of the present disclosure.
  • FIG. 9 is a schematic structural diagram of a device for object recognition provided by an embodiment of the present disclosure.
  • FIG. 10 is a schematic structural diagram of a device for object recognition provided by an embodiment of the present disclosure.
  • Fig. 11 is a schematic structural diagram of a server provided by an embodiment of the present disclosure.
  • the embodiments of the present disclosure may be applied to biometric recognition scenarios such as face recognition, fingerprint recognition, behavior feature recognition, and voiceprint recognition.
  • biometric recognition scenarios such as face recognition, fingerprint recognition, behavior feature recognition, and voiceprint recognition.
  • the features of each face picture are pre-stored in the matching library.
  • the person to be recognized can be extracted The feature of the face image, and then compare the feature with the feature of each face image in the matching library to determine the face image with the highest similarity and the similarity is higher than a certain value, you can determine the face image and the face to be recognized
  • Biometric identification refers to the identification of personal identity through the use of the inherent physiological characteristics of the human body (such as fingerprints, iris, face equality) or behavioral characteristics through a computer.
  • Feature extraction refers to the dimensionality reduction process of extracting the most statistically significant features by modeling samples.
  • the object to be identified is extracted first, then the target search is performed in the matching library, and the best matching object is found according to the similarity of the features after sequential comparison.
  • the embodiments of the present disclosure provide a method for object recognition, and the execution subject of the method may be a server or a terminal device.
  • the server may include a processor, a memory, a transceiver, etc.
  • the processor may be used to process the object recognition process, for example, the processor is used to convert the feature extracted by the first extraction model of the target object to the second extraction In the feature space of the model, the target object is the object to be recognized, and the processor can also be used to determine the matching object of the target object.
  • the memory may be used to store data required and generated during the process of object recognition. For example, the memory is used to store a matching library, and the matching library includes features of multiple objects.
  • the transceiver can be used to receive and send data.
  • the terminal device may include a processor, a memory, a transceiver, etc., and the processor may be used to process the object recognition process.
  • the memory can be used to store the data needed and generated during the object recognition process.
  • the transceiver can be used to receive and send data.
  • the embodiment of the present disclosure provides a method for object recognition, which can convert the feature of the object to be recognized into the feature that is compatible with the extraction of the low version extraction model.
  • the execution process of the method can be as follows:
  • Step 201 Convert the features extracted by the target object using the first extraction model into the features extracted by the second extraction model through the first feature conversion model.
  • the target object is an object to be identified, that is, an object whose identity has not yet been determined, and the target object may be a face picture, voiceprint, fingerprint, etc.
  • the server is taken as an example of the execution subject of the method for object recognition.
  • the server needs to recognize the face picture, and the face picture is the target object.
  • the server receives the fingerprint identification request, the fingerprint carried in the fingerprint identification request is the target object.
  • the target object is a face picture as an example for description.
  • the first feature conversion model can be a fully connected neural network.
  • the features extracted by the high version model can be converted to the features extracted by the low version extraction model.
  • the extraction model is a high version extraction model
  • the second extraction model is a low version extraction model.
  • a first feature conversion model can be trained in advance (the training process will be described later), which is used to convert the features extracted by a certain object using the first extraction model to the feature space of the second extraction model, that is, to realize the use of the first feature
  • the features extracted by the first extraction model are converted into the features extracted by the second extraction model.
  • the server can input the target object into the first extraction model to obtain the first feature of the target object.
  • the device connected to the server stores the first extraction model, and the device uses the first extraction model to extract the first feature of the target object, and then sends the first feature to the server. In this way, the server can obtain the features extracted by the target object using the first extraction model.
  • the server obtains the stored first feature conversion model, inputs the first feature of the target object into the first feature conversion model, converts the first feature of the target object into the feature space of the second extraction model, and obtains that the target object is in the feature space
  • the feature space of the second extraction model may refer to the feature space corresponding to the feature extracted by the second extraction model, so that the second feature of the target object in the feature space of the second extraction model is obtained.
  • the first extraction model is a high version extraction model
  • the second extraction model is a low version extraction model.
  • the performance of the high version extraction model is higher than that of the low version extraction model.
  • the above conversion of the first feature to the feature space of the second extraction model may be through a nonlinear mapping function or other conversion relationship, converting the first feature into a feature with the same dimension as the feature space in the feature space of the second extraction model For example, if the dimension of the feature extracted by the target object directly using the second extraction model is 5 dimensions, after the first feature is converted to the feature space of the second extraction model, the dimension of the second feature is also 5 dimensions.
  • Step 202 Match the second feature of the target object with the feature of the object in the matching library to obtain a target feature that matches the second feature of the target object.
  • the features of each object in the matching library are extracted using the second extraction model.
  • the object can be input to the second extraction model, and the output is the feature extracted by the second extraction model for the object, and then the feature of the object is stored in the matching In the library.
  • the second feature of the target object can be compared with the features of each object in the matching library to obtain the second feature of the target object.
  • the similarity between the feature and the feature of each object in the matching library, the feature with the greatest similarity and greater than the target value is determined as the target feature matching the second feature of the target object.
  • the process of calculating the similarity may be to calculate the cosine distance, and determine the cosine distance of the features as the similarity between the features.
  • the above target value can be preset and stored in the server.
  • the target value can be 0.6.
  • the features of the target object extracted by the first extraction model have been converted into the feature space of the second extraction model, it can be directly compared with the features of each object in the matching library.
  • Step 203 Determine the object to which the target feature belongs as a matching object of the target object.
  • the object to which the target feature belongs can be determined as the matching object of the target object.
  • the identity information of the matching object of the target object can be determined as the identity information of the target object. Is the object to be identified.
  • a process of training the first feature conversion model is also provided, and the processing may be as follows:
  • multiple sample objects can be obtained (for example, the server can obtain multiple sample objects from a device connected to itself, or receive multiple sample objects input by the user), and input multiple sample objects into the first extraction model, Obtain features extracted by multiple sample objects through the first extraction model, and input multiple sample objects into the second extraction model to obtain features extracted by multiple samples through the second extraction model.
  • the server directly obtains the features extracted by the first extraction model and the features extracted by the second extraction model of multiple sample objects from the device connected to the server.
  • the first extraction model of the target number of sample objects among the multiple sample objects into the first initial feature conversion model to obtain the first output result, and calculate the first output result and the target number of sample objects after the first extraction model.
  • extract the first loss value of the feature extracted by the model Use the first loss value as a constraint, and use the gradient descent algorithm to adjust the parameter values of the parameters in the first initial feature transformation model, and then select the target number of sample objects from multiple sample objects again, and perform the above process, you can also get one For the first loss value, continue to adjust the parameter values of the parameters in the first initial feature conversion model until the first loss value is minimized.
  • the parameter values of the parameters when the first loss value is minimized are selected and substituted into the first initial feature conversion model to obtain the first A feature conversion model.
  • the feature extracted by the sample object through the first extraction model is represented by fs
  • the feature extracted by the sample object through the second extraction model is represented by ft
  • the first output result of fs through the first initial feature conversion model is f-
  • the loss between t1, f-t1 and ft is the first loss value.
  • the features converted by the second initial feature conversion model will be converted from the feature space of the second extraction model to the feature space of the first extraction model.
  • the first output result can be input to the second initial feature conversion model to obtain the second output result. Then determine the second output result and the second loss value of the features extracted by the first extraction model for the target number of sample objects. Then you can use the gradient descent algorithm to adjust the parameter values of the parameters in the first initial feature transformation model and the second initial feature transformation model until a certain set of parameter values can be used to minimize the first loss value and the second loss value, that is, The first feature conversion model and the second feature conversion model, and only the first feature conversion model may be used in subsequent object recognition.
  • the server inputs the features extracted by the second extraction model of the target number of sample objects among the multiple sample objects into the second initial feature conversion model to obtain the third output result, determines the third output result and the target number of sample objects that pass the first Extract the third loss value of the feature extracted by the model, and input the third output result into the first initial feature conversion model to obtain the fourth output result; determine the fourth output result and the target number of sample objects extracted by the second extraction model
  • the fourth loss value of the feature is based on the first loss value, the second loss value, the third loss value, the fourth loss value, the first initial feature transformation model, the second initial feature transformation model, and the first extraction of multiple sample objects
  • the features extracted by the model and the features extracted by the second extraction model determine the first feature conversion model.
  • the server may input the features extracted by the second extraction model of the target number of sample objects among the multiple sample objects into the second initial feature conversion model to obtain the third output result. Calculate the third output result and the third loss value of the features extracted by the first extraction model for the target number of sample objects. Input the third output result into the first initial feature conversion model to obtain the fourth output result. Determine the fourth output result and the fourth loss value of the feature extracted by the second extraction model for the target number of sample objects. In this way, there will be four loss values, namely the first loss value, the second loss value, the third loss value and the fourth loss value.
  • the features extracted by the sample object through the first extraction model are represented by f-s
  • the features extracted by the sample object through the second extraction model are represented by f-t.
  • the first output result of fs through the first initial feature conversion model is f-t1
  • the loss between f-t1 and ft is the first loss value
  • the second output result of f-t1 through the second initial feature conversion model is f -s2
  • the loss between f-s2 and fs is the second loss value.
  • the third output result of ft passing through the second initial feature conversion model is f-s1
  • the loss between f-s1 and fs is the third loss value
  • the third output result f-s1 passing through the fourth of the first initial feature conversion model The output result is f-t2
  • the loss between f-t2 and ft is the fourth loss value.
  • first loss value second loss value
  • third loss value third loss value
  • fourth loss value first, second, third loss value, and fourth loss value
  • first loss value, second loss value, third loss value, and fourth loss value are all loss values obtained from L1 loss.
  • the L1 loss can be expressed as:
  • first loss value is equal to the above-mentioned first loss value, second loss value, third loss value, and fourth loss value are all loss values obtained by L2 loss.
  • L2 loss can be expressed by the formula:
  • the above-mentioned second loss value and fourth loss value are loss values obtained from L2 loss
  • the first loss value and third loss value are loss values obtained from L1 loss.
  • the expressions of L1 loss and L2 loss can be seen in the above description, because the first loss value is the loss value between the actual value and the estimated value after one conversion, and the third loss value is the actual value and the one converted
  • the loss value between the estimated values, the first loss value and the third loss value use the loss value obtained by the L1 loss. Since the L1 loss is more robust, the L1 loss is used as a constraint, and the first feature conversion model trained And the performance of the second feature conversion model is better.
  • the first feature extracted by the first extraction model of the target object is converted to the feature space of the second extraction model through the first feature conversion model to obtain the second feature of the target object in the feature space, and the target The second feature of the object is matched with the feature of the object in the matching library to obtain the target feature that matches the second feature of the target object, and the object to which the target feature belongs is determined as the matching object of the target object.
  • the features extracted by the second extraction model cannot directly match the features extracted by the first extraction model
  • the features extracted by the first extraction model can be converted to the feature space of the second extraction model, and the features extracted by the first extraction model
  • the features of can be matched with the features extracted by the second extraction model. Since the processing amount of feature conversion is much less than re-extracting the features in the matching library, the time for object recognition can be saved.
  • a block diagram of the processing flow in which the matching library is the ID photo matching library is also provided.
  • the ID photo imported or registered in the matching library is input to the first
  • the second extraction model is to obtain the features of the ID photo in the feature space of the second extraction model.
  • the face image captured in real time ie, the target object
  • the first extraction model is input to the first extraction model to obtain the first feature of the face image
  • the first feature is input to the first feature conversion model to obtain the face image in the second extraction model.
  • Features in feature space In this way, the features of the ID photo and the face image in the matching library belong to the feature space of the second extraction model, so the matching can be directly performed.
  • the embodiments of the present disclosure provide a method for object recognition, which can convert the features of the matching library into features that can be extracted by a higher version of the extraction model.
  • the execution subject of the method can be a server, as shown in FIG. 6, the method
  • the execution process can be as follows:
  • Step 601 Convert each object in the matching library using the third feature extracted by the second extraction model to the feature space of the first extraction model through the second feature conversion model to obtain the fourth feature of each object in the feature space.
  • the features extracted by the low version extraction model can be converted to the feature space of the high version extraction model.
  • the first extraction model is the high version extraction model.
  • the second extraction model is a low version extraction model.
  • a feature conversion model i.e., the second feature conversion model
  • the first extraction model is a high version extraction model
  • the second extraction model is a low version extraction model.
  • the performance of the high version extraction model is higher than that of the low version extraction model.
  • the above-mentioned transforming the third feature into the feature space of the first extraction model may be through a nonlinear mapping function, transforming the third feature into features with the same dimension as the feature dimension in the feature space of the first extraction model, for example, if The dimension of the feature extracted by an object in the matching library directly using the second extraction model is 5 dimensions. After the third feature is converted to the feature space of the first extraction model, the dimension of the fourth feature is also 5 dimensions.
  • Step 602 Match the feature extracted by the target object using the first extraction model with the fourth feature of each object in the feature space to obtain a target feature that matches the feature extracted by the target object using the first extraction model.
  • the first extraction model when recognizing the target object, can be used to extract the characteristics of the target object to obtain the characteristics of the target object, and the characteristics of the target object are compared with the fourth of each object in the feature space of the first extraction model.
  • the feature is matched to obtain the similarity between the feature of the target object and the fourth feature of each object in the feature space, and the target feature with the largest similarity and greater than the target value is selected.
  • Step 603 Determine the object to which the target feature belongs as a matching object of the target object.
  • the object to which the target feature belongs can be determined as the matching object of the target object.
  • a process of training the second feature conversion model is also provided, and the processing may be as follows:
  • the server directly obtains the features extracted by the first extraction model and the features extracted by the second extraction model of multiple sample objects from the device connected to the server.
  • the features extracted by the second extraction model of the target number of sample objects among the multiple sample objects into the second initial feature conversion model to obtain the fifth output result calculate the fifth output result and the target number of sample objects after the first A fifth loss value of the feature extracted by the extraction model.
  • Use the fifth loss value as a constraint use the gradient descent algorithm to adjust the parameter values of the parameters in the second initial feature transformation model, and then select the target number of sample objects from multiple sample objects again, and perform the above process, you can also get a first Five loss values, continue to adjust the parameter values of the parameters in the second initial feature conversion model until the fifth loss value is minimized, and select the parameter values that minimize the fifth loss value into the second initial feature conversion model to obtain the second Feature conversion model.
  • the features converted by the first initial feature conversion model will be converted from the feature space of the first extraction model to the feature space of the second extraction model.
  • the fifth output result can be input to the first initial feature conversion model to obtain the sixth output result. Then determine the sixth output result and the sixth loss value of the features extracted by the second extraction model for the target number of sample objects, and then use the gradient descent algorithm to adjust the parameters in the first initial feature transformation model and the second initial feature transformation model Parameter values, up to a certain set of parameter values, can minimize the fifth loss value and the sixth loss value, that is, obtain the first feature conversion model and the second feature conversion model, and then only the second feature conversion model can be used.
  • the features extracted by the first extraction model of the target number of sample objects among the multiple sample objects can be input into the first initial feature conversion model to obtain the seventh output result.
  • four loss values are obtained, namely the fifth loss value, the sixth loss value, the seventh loss value, and the eighth loss value.
  • the foregoing fifth loss value, sixth loss value, seventh loss value, and eighth loss value are all loss values obtained from L1 loss.
  • the L1 loss can be expressed by a formula, which can be referred to the description above.
  • the above-mentioned fifth loss value, sixth loss value, seventh loss value, and eighth loss value are all loss values obtained by L2 loss.
  • the formula for L2 loss can be referred to the previous description.
  • the above-mentioned fifth loss value and seventh loss value are loss values obtained from L1 loss
  • the sixth loss value and eighth loss value are loss values obtained from L2 loss.
  • the expressions of L1 loss and L2 loss can be seen in the above description, because the fifth loss value is the loss value between the actual value and the estimated value after one conversion, and the seventh loss value is the actual value and the one converted The loss value between the estimated value, so the fifth loss value and the seventh loss value use the loss value obtained by the L1 loss.
  • the L1 loss is more robust, so the L1 loss is used as the constraint, and the first training is The performance of the feature conversion model and the second feature conversion model is better.
  • the server obtains the second feature conversion model, and uses the third feature extracted by the second extraction model to convert each object in the matching library into the feature space of the first extraction model through the second feature conversion model to obtain each The fourth feature of the object in the feature space, the feature extracted by the target object using the first extraction model is matched with the fourth feature of each object in the feature space, and the feature matching the feature extracted by the target object using the first extraction model is obtained.
  • the target feature, the object to which the target feature belongs is determined as the matching object of the target object.
  • the features extracted by the second extraction model cannot directly match the features extracted by the first extraction model
  • the features extracted by the second extraction model can be converted to the feature space of the first extraction model, and the features extracted by the first extraction model The features of can be matched with the features extracted by the second extraction model, without reusing the first extraction model to extract the features in the matching library, so the time for object recognition can be saved.
  • the second feature conversion model performs feature conversion on the feature of an object
  • the calculation amount is relatively small, which is much lower than that of re-extracting the feature of the object using the first extraction model.
  • the idea of the two embodiments of the present disclosure is also provided. Since the recognition algorithm itself has a similar discrimination ability for a specific object test set, the task objectives completed by each feature are also consistent. Therefore, although there is no direct linear correlation between the features of different versions of the extracted model, it can be assumed that there is a non-linear mapping function that can map the features from the feature space of the high version extraction model to the low version extraction model. Feature space.
  • the above-mentioned first feature conversion model and the second feature conversion model may be a neural network model.
  • the first feature conversion model and the second feature conversion model can be a multi-layer fully connected neural network, and the structure can include a normalization layer (Norm), a fully connected layer (Fully connected layer), and a batch normalization layer (batch normalization layer). ), hyperbolic tangent function processing layer.
  • the normalization layer can be used to make the modulus length of the vector to 1.
  • the function of the batch normalization layer is to converge better when the gradient drops, and the trained first feature conversion model makes the performance better.
  • the hyperbolic tangent function is a nonlinear function and can be used to fit a nonlinear mapping function.
  • the first feature conversion model includes a 4-layer fully connected neural network.
  • Each fully connected neural network includes a normalization layer, a fully connected layer, a batch normalization layer, and a hyperbolic tangent function layer.
  • the input of a fully connected layer is 256 channels
  • the output is 512 channels
  • the input of the second fully connected layer is 512 channels
  • the output is 768 channels
  • the input of the third fully connected layer There are 768 channels
  • the output is 1024 channels
  • the input in the fourth fully connected layer is 1024 and the number of channels
  • the output is 256 channels.
  • the first extraction model and the second extraction model are any models that can be used to extract features, as long as the version of the first extraction model is higher than the version of the second extraction model.
  • the embodiments of the present disclosure are not limited.
  • an embodiment of the present disclosure also provides a device for object recognition. As shown in FIG. 7, the device includes:
  • the conversion module 710 is configured to convert the first feature extracted by the target object using the first extraction model to the feature space of the second extraction model through the first feature conversion model to obtain the second feature space of the target object in the feature space. feature;
  • the matching module 720 is used for:
  • the object to which the target feature belongs is determined as a matching object of the target object.
  • the device further includes a training module 730 for:
  • the first initial feature conversion model determines the first Feature conversion model.
  • the training module 730 is also used to:
  • the training module 730 is used to:
  • the first loss value the second loss value, the first initial feature conversion model, the features extracted by the first extraction model of the multiple sample objects, and the features extracted by the second extraction model Features, determining the first feature conversion model.
  • the training module 730 is also used to:
  • the training module 730 is used to:
  • the first feature conversion model is determined for each sample object with features extracted through the first extraction model and features extracted through the second extraction model.
  • the first loss value, the second loss value, the third loss value, and the fourth loss value are all loss values obtained from L1 loss; or,
  • the first loss value, the second loss value, the third loss value, and the fourth loss value are all loss values obtained from L2 loss; or,
  • the first loss value and the third loss value are loss values obtained by L1 loss
  • the second loss value and the fourth loss value are loss values obtained by L2 loss.
  • the first feature extracted by the first extraction model of the target object is converted to the feature space of the second extraction model through the first feature conversion model to obtain the second feature of the target object in the feature space, and the target The second feature of the object is matched with the feature of the object in the matching library to obtain the target feature that matches the second feature of the target object, and the object to which the target feature belongs is determined as the matching object of the target object.
  • the features extracted by the first extraction model cannot directly match the features extracted by the second extraction model
  • the features extracted by the first extraction model can be converted to the feature space of the second extraction model, and the features extracted by the first extraction model
  • the features of can be matched with the features extracted by the second extraction model. Since the processing amount of feature conversion is much less than re-extracting features in the matching library, time can be saved.
  • the device for object recognition provided in the above embodiment only uses the division of the above functional modules for example when performing object recognition.
  • the above functions can be allocated to different functional modules as needed.
  • Complete that is, divide the internal structure of the device into different functional modules to complete all or part of the functions described above.
  • the device for performing object recognition provided in the above-mentioned embodiment belongs to the same concept as the embodiment of the method for performing object recognition. For the specific implementation process, please refer to the method embodiment, which will not be repeated here.
  • the embodiments of the present disclosure also provide a device for object recognition. As shown in FIG. 9, the device includes:
  • the conversion module 910 is configured to convert each object in the matching library using the third feature extracted by the second extraction model to the feature space of the first extraction model through the second feature conversion model, to obtain that each object is in the feature space
  • the matching module 920 is used for:
  • the object to which the target feature belongs is determined as a matching object of the target object.
  • the device further includes a training module 930 for:
  • the second initial feature conversion model, the features extracted by the first extraction model of the multiple sample objects, and the features extracted by the second extraction model determine the second Feature conversion model.
  • the training module 930 is also used to:
  • the training module 930 is used to:
  • the fifth loss value, the sixth loss value, the second initial feature conversion model, the features extracted by the first extraction model of the multiple sample objects, and the features extracted by the second extraction model Features determining the second feature conversion model.
  • the training module 930 is also used to:
  • the training module 930 is used to:
  • the fifth loss value, the sixth loss value, the seventh loss value, the eighth loss value, the first initial feature conversion model, the second initial feature conversion model, and the multiple sample objects The features extracted by the first extraction model and the features extracted by the second extraction model are used to determine the second feature conversion model.
  • the fifth loss value, the sixth loss value, the seventh loss value, and the eighth loss value are all loss values obtained from L1 loss; or,
  • the fifth loss value, the sixth loss value, the seventh loss value, and the eighth loss value are all loss values obtained from L2 loss; or,
  • the fifth loss value and the seventh loss value are loss values obtained by L1 loss
  • the sixth loss value and the eighth loss value are loss values obtained by L2 loss.
  • the second feature conversion model is acquired, and each object in the matching library uses the third feature extracted by the second extraction model to convert the second feature conversion model to the feature space of the first extraction model to obtain each object
  • the feature extracted by the target object using the first extraction model is matched with the fourth feature of each object in the feature space to obtain a target that matches the feature extracted by the target object using the first extraction model Feature, the object to which the target feature belongs is determined as the matching object of the target object.
  • the features extracted by the second extraction model cannot directly match the features extracted by the first extraction model
  • the features extracted by the second extraction model can be converted to the feature space of the first extraction model, and the features extracted by the first extraction model The features of can be matched with the features extracted by the second extraction model, without the need to reuse the first extraction model to extract the features in the matching library, so time can be saved.
  • the device for object recognition provided in the above embodiment only uses the division of the above functional modules for example when performing object recognition.
  • the above functions can be allocated to different functional modules as needed.
  • Complete that is, divide the internal structure of the device into different functional modules to complete all or part of the functions described above.
  • the device for performing object recognition provided in the above-mentioned embodiment belongs to the same concept as the embodiment of the method for performing object recognition. For the specific implementation process, please refer to the method embodiment, which will not be repeated here.
  • FIG. 11 is a schematic structural diagram of a server provided by an embodiment of the present disclosure.
  • the server 1100 may have relatively large differences due to different configurations or performance, and may include one or more processors 1101 and one or more memories 1102. Wherein, at least one instruction is stored in the memory 1102, and the at least one instruction is loaded and executed by the processor 1101 to implement each step in the foregoing method for object recognition.
  • a computer-readable storage medium is also provided, and a computer program is stored in the storage medium.
  • the computer program is executed by a processor, the foregoing method for object recognition is realized.
  • a server for object recognition includes a processor and a memory, wherein the memory is used for storing computer programs; the processor is used for executing all data on the memory.
  • the stored program implements the above method of object recognition.

Landscapes

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

Abstract

本公开提供了一种进行对象识别的方法和装置,属于计算机技术领域。所述方法包括:将目标对象使用第一提取模型提取的第一特征,通过第一特征转换模型转换到第二提取模型的特征空间,得到目标对象在特征空间中的第二特征,将目标对象的第二特征,与匹配库中的对象的特征进行匹配,得到与目标对象的第二特征相匹配的目标特征,将目标特征所属的对象,确定为目标对象的匹配对象。采用本公开,可以节约时长。

Description

进行对象识别的方法和装置
本公开要求于2019年07月30日提交的申请号为201910696479.X、发明名称为“进行对象识别的方法和装置”的中国专利申请的优先权,其全部内容通过引用结合在本公开中。
技术领域
本公开涉及计算机技术领域,特别涉及一种进行对象识别的方法和装置。
背景技术
随着计算机技术的发展,对象识别已经在多个领域中部署应用。例如,人脸识别、指纹识别、声纹识别等。
相关技术中,在进行对象识别时,一般是使用提取模型,提取待识别对象的特征,然后将提取的特征与匹配库中各对象的特征进行比对,确定相似度最高的特征。将相似度最高的特征所属的对象确定为待识别对象的匹配对象,将匹配对象的身份信息确定为待识别对象的身份信息。
随着时间推移,提取模型会进行更新,而匹配库中的特征还是使用低版本提取模型提取的,有可能会出现高版本提取模型提取的特征无法与匹配库的特征进行匹配的情况,这时,就需要使用高版本的提取模型,从样本中重新提取一遍每个对象的特征形成匹配库,然而样本(如,图片、指纹、音频)的数量较大时,提取过程中涉及的数据量就会更加庞大,会花费大量的时间。
发明内容
本公开实施例提供了一种进行对象识别的方法和装置,至少能够解决升级模型后,重新提取匹配库中的特征,花费时间长的问题。所述技术方案如下:
第一方面,提供了一种进行对象识别的方法,所述方法包括:
将目标对象使用第一提取模型提取的第一特征,通过第一特征转换模型转换到第二提取模型的特征空间,得到所述目标对象在所述特征空间中的第二特征;
将所述目标对象的第二特征,与匹配库中的对象的特征进行匹配,得到与所述目标对象的第二特征相匹配的目标特征;
将所述目标特征所属的对象,确定为所述目标对象的匹配对象。
可选的,所述方法还包括:
获取多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征;
将所述多个样本对象中目标数目个样本对象经过所述第一提取模型提取的特征,输入到第一初始特征转换模型,得到第一输出结果;
确定所述第一输出结果与所述目标数目个样本对象经过所述第二提取模型提取的特征的第一损失值;
根据所述第一损失值、所述第一初始特征转换模型、所述多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征,确定所述第一特征转换模型。
可选的,所述方法还包括:
将所述第一输出结果输入到第二初始特征转换模型,得到第二输出结果;
确定所述第二输出结果与所述目标数目个样本对象经过所述第一提取模型提取的特征的第二损失值;
所述根据所述第一损失值、所述第一初始特征转换模型、所述多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征,确定所述第一特征转换模型,包括:
根据所述第一损失值、所述第二损失值、所述第一初始特征转换模型、所述多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征,确定所述第一特征转换模型。
可选的,所述方法还包括:
将所述多个样本对象中目标数目个样本对象经过第二提取模型提取的特征,输入到所述第二初始特征转换模型,得到第三输出结果;
确定所述第三输出结果与所述目标数目个样本对象经过第一提取模型提取的特征的第三损失值;
将所述第三输出结果输入到所述第一初始特征转换模型中,得到第四输出结果;
确定所述第四输出结果与所述目标数目个样本对象经过第二提取模型提取的特征的第四损失值;
所述根据所述第一损失值、所述第二损失值、所述第一初始特征转换模型、所述多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征,确定所述第一特征转换模型,包括:
根据所述第一损失值、所述第二损失值、所述第三损失值、所述第四损失值、所述第一初始特征转换模型、所述第二初始特征转换模型、所述多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征,确定所述第一特征转换模型。
可选的,所述第一损失值、所述第二损失值、所述第三损失值和所述第四损失值均为L1损失得到的损失值;或者,
所述第一损失值、所述第二损失值、所述第三损失值和所述第四损失值均为L2损失得到的损失值;或者,
所述第一损失值和所述第三损失值为L1损失得到的损失值,所述第二损失值和所述第四损失值为L2损失得到的损失值。
第二方面,提供了一种进行对象识别的方法,所述方法包括:
将匹配库中各对象使用第二提取模型提取的第三特征,分别通过第二特征转换模型转换到第一提取模型的特征空间,得到所述各对象在所述特征空间中的第四特征;
将目标对象使用所述第一提取模型提取的特征与所述各对象在所述特征空间中的第四特征进行匹配,得到与所述目标对象使用所述第一提取模型提取的特征相匹配的目标特征;
将所述目标特征所属的对象,确定为所述目标对象的匹配对象。
可选的,所述方法还包括:
获取多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征;
将所述多个样本对象中目标数目个样本对象经过第二提取模型提取的特征,输入到第二初始特征转换模型,得到第五输出结果;
确定所述第五输出结果与所述目标数目个样本对象经过第一提取模型提取的特征的第五损失值;
根据所述第五损失值、所述第二初始特征转换模型、所述多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征,确定所述第二特征转换模型。
可选的,所述方法还包括:
将所述第五输出结果输入到第一初始特征转换模型,得到第六输出结果;
确定所述第六输出结果与所述目标数目个样本对象经过第二提取模型提取的特征的第六损失值;
所述根据所述第五损失值、所述第二初始特征转换模型、所述多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征,确定所述第二特征转换模型,包括:
根据所述第五损失值、所述第六损失值、所述第二初始特征转换模型、所述多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征,确定所述第二特征转换模型。
可选的,所述方法还包括:
将所述多个样本对象中目标数目个样本对象经过第一提取模型提取的特征,输入到所述第一初始特征转换模型,得到第七输出结果;
确定所述第七输出结果与所述目标数目个样本对象经过第二提取模型提取的特征的第七损失值;
将所述第七输出结果输入到所述第二初始特征转换模型中,得到第八输出结果;
确定所述第八输出结果与所述目标数目个样本对象经过第一提取模型提取的特征的第八损失值;
所述根据所述第五损失值、所述第六损失值、所述第二初始特征转换模型、所述多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征,确定所述第二特征转换模型,包括:
根据所述第五损失值、所述第六损失值、第七损失值、第八损失值、所述第一初始特征转换模型、所述第二初始特征转换模型、所述多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征,确定所述第二特征转换模型。
可选的,所述第五损失值、所述第六损失值、所述第七损失值和所述第八 损失值均为L1损失得到的损失值;或者,
所述第五损失值、所述第六损失值、所述第七损失值和所述第八损失值均为L2损失得到的损失值;或者,
所述第五损失值和所述第七损失值为L1损失得到的损失值,所述第六损失值和所述第八损失值为L2损失得到的损失值。
第三方面,提供了一种进行对象识别的装置,该装置包括:
转换模块,用于将目标对象使用第一提取模型提取的第一特征,通过第一特征转换模型转换到第二提取模型的特征空间,得到所述目标对象在所述特征空间中的第二特征;
匹配模块,用于:
将所述目标对象的第二特征,与匹配库中的对象的特征进行匹配,得到与所述目标对象的第二特征相匹配的目标特征;
将所述目标特征所属的对象,确定为所述目标对象的匹配对象。
可选的,所述装置还包括训练模块,用于:
获取多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征;
将所述多个样本对象中目标数目个样本对象经过所述第一提取模型提取的特征,输入到第一初始特征转换模型,得到第一输出结果;
确定所述第一输出结果与所述目标数目个样本对象经过所述第二提取模型提取的特征的第一损失值;
根据所述第一损失值、所述第一初始特征转换模型、所述多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征,确定所述第一特征转换模型。
可选的,所述训练模块,还用于:
将所述第一输出结果输入到第二初始特征转换模型,得到第二输出结果;
确定所述第二输出结果与所述目标数目个样本对象经过所述第一提取模型提取的特征的第二损失值;
所述训练模块,用于:
根据所述第一损失值、所述第二损失值、所述第一初始特征转换模型、所述多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提 取的特征,确定所述第一特征转换模型。
可选的,所述训练模块,还用于:
将所述多个样本对象中目标数目个样本对象经过第二提取模型提取的特征,输入到所述第二初始特征转换模型,得到第三输出结果;
确定所述第三输出结果与所述目标数目个样本对象经过第一提取模型提取的特征的第三损失值;
将所述第三输出结果输入到所述第一初始特征转换模型中,得到第四输出结果;
确定所述第四输出结果与所述目标数目个样本对象经过第二提取模型提取的特征的第四损失值;
所述训练模块,用于:
根据所述第一损失值、所述第二损失值、所述第三损失值、所述第四损失值、所述第一初始特征转换模型、所述第二初始特征转换模型、所述多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征,确定所述第一特征转换模型。
可选的,所述第一损失值、所述第二损失值、所述第三损失值和所述第四损失值均为L1损失得到的损失值;或者,
所述第一损失值、所述第二损失值、所述第三损失值和所述第四损失值均为L2损失得到的损失值;或者,
所述第一损失值和所述第三损失值为L1损失得到的损失值,所述第二损失值和所述第四损失值为L2损失得到的损失值。
第四方面,提供了一种进行对象识别的装置,该装置包括:
转换模块,用于将匹配库中各对象使用第二提取模型提取的第三特征,分别通过第二特征转换模型转换到第一提取模型的特征空间,得到所述各对象在所述特征空间中的第四特征;
匹配模块,用于:
将目标对象使用所述第一提取模型提取的特征与所述各对象在所述特征空间中的第四特征进行匹配,得到与所述目标对象使用所述第一提取模型提取的特征相匹配的目标特征;
将所述目标特征所属的对象,确定为所述目标对象的匹配对象。
可选的,所述装置还包括训练模块,用于:
获取多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征;
将所述多个样本对象中目标数目个样本对象经过第二提取模型提取的特征,输入到第二初始特征转换模型,得到第五输出结果;
确定所述第五输出结果与所述目标数目个样本对象经过第一提取模型提取的特征的第五损失值;
根据所述第五损失值、所述第二初始特征转换模型、所述多个样本对象经过第一提取模型提取的特征和经过第二提取模型提取的特征,确定所述第二特征转换模型。
可选的,所述训练模块,还用于:
将所述第五输出结果输入到第一初始特征转换模型,得到第六输出结果;
确定所述第六输出结果与所述目标数目个样本对象经过第二提取模型提取的特征的第六损失值;
所述训练模块,用于:
根据所述第五损失值、所述第六损失值、所述第二初始特征转换模型、所述多个样本对象经过第一提取模型提取的特征和经过第二提取模型提取的特征,确定所述第二特征转换模型。
可选的,所述训练模块,还用于:
将所述多个样本对象中目标数目个样本对象经过第一提取模型提取的特征,输入到所述第一初始特征转换模型,得到第七输出结果;
确定所述第七输出结果与所述目标数目个样本对象经过第二提取模型提取的特征的第七损失值;
将所述第七输出结果输入到所述第二初始特征转换模型中,得到第八输出结果;
确定所述第八输出结果与所述目标数目个样本对象经过第一提取模型提取的特征的第八损失值;
所述训练模块,用于:
根据所述第五损失值、所述第六损失值、第七损失值、第八损失值、所述第一初始特征转换模型、所述第二初始特征转换模型、所述多个样本对象经过 第一提取模型提取的特征和经过第二提取模型提取的特征,确定所述第二特征转换模型。
可选的,所述第五损失值、所述第六损失值、所述第七损失值和所述第八损失值均为L1损失得到的损失值;或者,
所述第五损失值、所述第六损失值、所述第七损失值和所述第八损失值均为L2损失得到的损失值;或者,
所述第五损失值和所述第七损失值为L1损失得到的损失值,所述第六损失值和所述第八损失值为L2损失得到的损失值。
第三方面,提供了一种计算机可读存储介质,该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,实现上述第一方面和第二方面所述的进行对象识别的方法。
第四方面,提供了一种进行对象识别的服务器,该服务器包括处理器和存储器,其中,所述存储器,用于存放计算机程序;所述处理器,用于执行所述存储器上所存放的程序,实现上述第一方面和第二方面所述的进行对象识别的方法。
本公开实施例提供的技术方案带来的有益效果至少包括:
本公开实施例中,将目标对象使用第一提取模型提取的第一特征,通过第一特征转换模型转换到第二提取模型的特征空间,得到目标对象在特征空间中的第二特征,将目标对象的第二特征,与匹配库中的对象的特征进行匹配,得到与目标对象的第二特征相匹配的目标特征,将目标特征所属的对象,确定为目标对象的匹配对象。这样,在经过第一提取模型提取的特征与经过第二提取模型提取的特征不能直接匹配时,可以将第一提取模型提取的特征转换到第二提取模型的特征空间,经过第一提取模型提取的特征就能与经过第二提取模型提取的特征进行匹配,由于特征转换的处理量远远小于重新提取匹配库中的特征,所以可以节约对象识别的时长。
附图说明
为了更清楚地说明本公开实施例中的技术方案,下面将对实施例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公开的一些实施例,对于本领域普通技术人员来讲,在不付出创造性劳动的前提下, 还可以根据这些附图获得其他的附图。
图1是本公开实施例提供的一种对象识别的示意图;
图2是本公开实施例提供的一种进行对象识别的方法流程图;
图3是本公开实施例提供的一种训练的示意图;
图4是本公开实施例提供的一种训练的示意图;
图5是本公开实施例提供的一种进行对象识别的方法流程图;
图6是本公开实施例提供的一种进行对象识别的方法流程图;
图7是本公开实施例提供的一种进行对象识别的装置的结构示意图;
图8是本公开实施例提供的一种进行对象识别的装置的结构示意图;
图9是本公开实施例提供的一种进行对象识别的装置的结构示意图;
图10是本公开实施例提供的一种进行对象识别的装置的结构示意图;
图11是本公开实施例提供的一种服务器的结构示意图。
具体实施方式
为使本公开的目的、技术方案和优点更加清楚,下面将结合附图对本公开实施方式作进一步地详细描述。
在进行实施前,首先介绍一下本公开实施例的应用场景,以及本公开实施例可能涉及的名词概念:
本公开实施例可以是应用于人脸识别、指纹识别、行为特征识别、声纹识别等生物特征识别场景中。如图1所示,在人脸识别场景中,匹配库中预先存储有各人脸图片的特征,在确定待识别人脸图片与匹配库中的某个人脸图片匹配时,可以提取待识别人脸图片的特征,然后将该特征与匹配库中各人脸图片的特征进行比对,确定相似度最高的人脸图片且相似度高于一定数值,可以确定该人脸图片与待识别人脸图片相匹配,即该人脸图片与待识别人脸图片属于同一身份。
生物特征识别,指通过计算机利用人体固有的生理特征(如指纹、虹膜、面相等)或行为特征进行个人身份的鉴定。
特征提取,指通过对样本进行建模,抽取出最具统计意义特征的降维过程。
特征比对,对待识别对象先进行特征提取,再在匹配库中进行目标搜索, 依次比较后根据特征相似程度找出最佳匹配对象。
本公开实施例提供了一种进行对象识别的方法,该方法的执行主体可以是服务器或终端设备等。其中,服务器中可以包括处理器、存储器和收发器等,处理器可以用于进行对象识别的过程的处理,例如,处理器用于将目标对象的使用第一提取模型提取的特征转换到第二提取模型的特征空间,目标对象为待识别对象,并且处理器还可以用于确定目标对象的匹配对象。存储器可以用于存储进行对象识别过程中需要的数据以及产生的数据,例如,存储器用于存储匹配库,匹配库中包括多个对象的特征。收发器可以用于接收以及发送数据。
终端设备中可以包括处理器、存储器和收发器等,处理器可以用于进行对象识别的过程的处理。存储器可以用于存储进行对象识别过程中需要的数据以及产生的数据。收发器可以用于接收以及发送数据。
本公开实施例提供了一种进行对象识别的方法,可以将待识别对象的特征转换为可以兼容低版本提取模型提取的特征,如图2所示,该方法的执行流程可以如下:
步骤201,将目标对象使用第一提取模型提取的特征,通过第一特征转换模型转换为使用第二提取模型提取的特征。
其中,目标对象为待识别对象,也即还未确定身份的对象,目标对象可以是人脸图片、声纹、指纹等。例如,在图2的流程中以服务器为进行对象识别的方法的执行主体为例进行说明。在抓拍到人脸图片时,服务器要对人脸图片进行识别,人脸图片即为目标对象。再例如,服务器接收到指纹识别请求时,指纹识别请求中携带的指纹即为目标对象。本公开实施例以目标对象为人脸图片为例进行说明。第一特征转换模型可以是全连接的神经网络,在进行对象识别时,为了兼容低版本提取模型提取的特征,可以将高版本模型提取的特征转换到使用低版本提取模型提取的特征,第一提取模型为高版本提取模型,第二提取模型为低版本提取模型。可以预先训练一个第一特征转换模型(训练过程在后面进行说明),用于将某个对象使用第一提取模型提取的特征转换到第二提取模型的特征空间,即实现将某个对象使用第一提取模型提取的特征转换为使用第二提取模型提取的特征。
在实施中,服务器在获取到目标对象后,可以将目标对象输入到第一提取模型,得到目标对象的第一特征。或者,与服务器连接设备中存储有第一提取模型,该设备使用第一提取模型提取到目标对象的第一特征后,将该第一特征发送给服务器。这样,服务器即可获得目标对象使用第一提取模型提取的特征。
然后服务器获取存储的第一特征转换模型,将目标对象的第一特征输入到第一特征转换模型,将目标对象的第一特征转换到第二提取模型的特征空间,得到目标对象在该特征空间中的第二特征,第二提取模型的特征空间可以指利用第二提取模型提取的特征对应的特征空间,这样,就得到目标对象在第二提取模型的特征空间中的第二特征。
需要说明的是,第一提取模型为高版本提取模型,第二提取模型为低版本提取模型,高版本提取模型的性能要高于低版本提取模型的性能。上述将第一特征转换到第二提取模型的特征空间,可以是通过一个非线性的映射函数或者其它转换关系,将第一特征转换为与第二提取模型的特征空间中的特征维度一样的特征,例如,若目标对像直接使用第二提取模型的提取的特征的维度为5维,将第一特征转换到第二提取模型的特征空间后,第二特征的维度也为5维。
步骤202,将目标对象的第二特征,与匹配库中的对象的特征进行匹配,得到与目标对象的第二特征相匹配的目标特征。
其中,匹配库中各对象的特征是使用第二提取模型提取的。在某个对象入库(即添加至匹配库)时,可以将该对象输入到第二提取模型中,输出则为该对象使用第二提取模型提取的特征,然后将该对象的特征存储在匹配库中。
在实施中,在获取到目标对象在第二提取模型的特征空间中的第二特征后,可以将目标对象的第二特征与匹配库中各对象的特征进行比对,得到目标对象的第二特征与匹配库中各对象的特征的相似度,将相似度最大,且大于目标数值的特征,确定为与目标对象的第二特征相匹配的目标特征。
需要说明的是,计算相似度的过程可以是计算余弦距离,将特征的余弦距离,确定为特征之间的相似度。上述目标数值可以预设,并且存储至服务器中,目标数值可以为0.6。另外,由于目标对象经过第一提取模型提取的特征已经转换到第二提取模型的特征空间中,所以可以与匹配库中的各对象的特征进行直接比较。
步骤203,将目标特征所属的对象,确定为目标对象的匹配对象。
在实施中,在确定目标特征之后,可以将目标特征所属的对象,确定为目标对象的匹配对象,这样,可以将目标对象的匹配对象的身份信息确定为目标对象的身份信息,目标对象不再是待识别对象。
可选的,本公开实施例中,还提供了训练第一特征转换模型的过程,处理可以如下:
获取多个样本对象经过第一提取模型提取的特征和经过第二提取模型提取的特征,将多个样本对象中目标数目个样本对象经过第一提取模型提取的特征,输入到第一初始特征转换模型,得到第一输出结果,确定第一输出结果与目标数目个样本对象经过第二提取模型提取的特征的第一损失值,根据第一损失值、第一初始特征转换模型、多个样本对象经过第一提取模型提取的特征和经过第二提取模型提取的特征,确定第一特征转换模型。
在实施中,可以获取多个样本对象(如服务器可以从与自身连接的设备中获取多个样本对象,或者接收用户输入的多个样本对象),将多个样本对象输入到第一提取模型,得到多个样本对象经过第一提取模型提取的特征,并将多个样本对象输入到第二提取模型,得到多个样本经过第二提取模型提取的特征。或者,服务器直接从与自身连接的设备获取多个样本对象经过第一提取模型提取的特征和经过第二提取模型提取的特征。
然后将多个样本对象中目标数目个样本对象经过第一提取模型提取的特征,输入到第一初始特征转换模型,得到第一输出结果,计算第一输出结果与该目标数目个样本对象经过第二提取模型提取的特征的第一损失值。使用第一损失值作为约束,并使用梯度下降算法,调整第一初始特征转换模型中参数的参数值,然后重新在多个样本对象中选取目标数目个样本对象,执行上述过程,也能得到一个第一损失值,继续调整第一初始特征转换模型中参数的参数值,直到使第一损失值最小,选择使第一损失值最小时的参数的参数值代入第一初始特征转换模型,得到第一特征转换模型。如图3所示,样本对象经过第一提取模型提取的特征使用f-s表示,样本对象经过第二提取模型提取的特征使用f-t表示,f-s经过第一初始特征转换模型的第一输出结果为f-t1,f-t1与f-t之间的损失为第一损失值。
可选的,为了使第一特征转换模型性能更好,可以使用更多的约束,进行训练,相应的处理可以如下:
将第一输出结果输入到第二初始特征转换模型,得到第二输出结果,确定第二输出结果与目标数目个样本对象经过第一提取模型提取的特征的第二损失值,根据第一损失值、第二损失值、第一初始特征转换模型、多个样本对象经过第一提取模型提取的特征和经过第二提取模型提取的特征,确定第一特征转换模型。
其中,经过第二初始特征转换模型转换后的特征,会从第二提取模型的特征空间转换到第一提取模型的特征空间。
在实施中,在得到第一输出结果之后,可以将第一输出结果输入到第二初始特征转换模型,得到第二输出结果。然后确定第二输出结果与目标数目个样本对象经过第一提取模型提取的特征的第二损失值。然后可以使用梯度下降算法调整第一初始特征转换模型和第二初始特征转换模型中的参数的参数值,直到某一组参数值,可以使第一损失值和第二损失值均最小,即得到第一特征转换模型和第二特征转换模型,后续在进行对象识别时可以仅使用第一特征转换模型。
这样,由于使用了两个损失进行约束,所以可以使训练出的第一特征转换模型性能更好。
可选的,为了使第一特征转换模型的性能更好,可以使用更多的约束来训练第一特征转换模型,处理可以如下:
服务器将多个样本对象中目标数目个样本对象经过第二提取模型提取的特征,输入到第二初始特征转换模型,得到第三输出结果,确定第三输出结果与目标数目个样本对象经过第一提取模型提取的特征的第三损失值,将第三输出结果输入到第一初始特征转换模型中,得到第四输出结果;确定第四输出结果与目标数目个样本对象经过第二提取模型提取的特征的第四损失值,根据第一损失值、第二损失值、第三损失值、第四损失值、第一初始特征转换模型、第二初始特征转换模型、多个样本对象经过第一提取模型提取的特征和经过第二提取模型提取的特征,确定第一特征转换模型。
在实施中,服务器可以将多个样本对象中目标数目个样本对象经过第二提取模型提取的特征,输入到第二初始特征转换模型中,得到第三输出结果。计算第三输出结果与目标数目个样本对象经过第一提取模型提取的特征的第三损失值。将第三输出结果再输入到第一初始特征转换模型中,得到第四输出结果。 确定第四输出结果与目标数目个样本对象经过第二提取模型提取的特征的第四损失值。这样,会有四个损失值,即第一损失值、第二损失值、第三损失值和第四损失值。如图4所示,样本对象经过第一提取模型提取的特征使用f-s表示,样本对象经过第二提取模型提取的特征使用f-t表示。f-s经过第一初始特征转换模型的第一输出结果为f-t1,f-t1与f-t之间的损失为第一损失值,f-t1经过第二初始特征转换模型的第二输出结果为f-s2,f-s2与f-s之间的损失为第二损失值。f-t经过第二初始特征转换模型的第三输出结果为f-s1,f-s1与f-s之间的损失为第三损失值,第三输出结果f-s1经过第一初始特征转换模型的第四输出结果为f-t2,f-t2与f-t之间的损失为第四损失值。
然后可以使用梯度下降算法调整第一初始特征转换模型和第二初始特征转换模型中的参数的参数值,直到某一组参数值,可以使第一损失值、第二损失值、第三损失值和第四损失值均最小,即得到第一特征转换模型和第二特征转换模型,后续在进行对象识别时可以仅使用第一特征转换模型。
需要说明的是,上述第一损失值、第二损失值、第三损失值和第四损失值中“第一”、“第二”、“第三”和“第四”用于区别类似的对象,而不必用于描述特定的顺序或先后次序。
可选的,上述第一损失值、第二损失值、第三损失值和第四损失值均为L1损失得到的损失值。L1损失用公式表示可以为:
Figure PCTCN2020105500-appb-000001
表示m个特征的实际值和估计值之间的差值的绝对值之和。
或者,上述第一损失值、第二损失值、第三损失值和第四损失值均为L2损失得到的损失值。L2损失用公式表示可以为:
Figure PCTCN2020105500-appb-000002
表示m个特征的实际值和估计值之间的差值的平方之和。此处的m即为上述提到的目标数目。
或者,上述第二损失值和第四损失值为L2损失得到的损失值,第一损失值和第三损失值为L1损失得到的损失值。
在实施中,L1损失和L2损失的表达式可以见上述描述,由于第一损失值是实际值与经过一次转换的估计值之间的损失值,第三损失值是实际值与经过一次转换的估计值之间的损失值,第一损失值和第三损失值采用L1损失得到的损失值,由于L1损失的鲁棒性更好,所以使用L1损失作为约束,训练出的第 一特征转换模型和第二特征转换模型的性能更好。
本公开实施例中,将目标对象使用第一提取模型提取的第一特征,通过第一特征转换模型转换到第二提取模型的特征空间,得到目标对象在特征空间中的第二特征,将目标对象的第二特征,与匹配库中的对象的特征进行匹配,得到与目标对象的第二特征相匹配的目标特征,将目标特征所属的对象,确定为目标对象的匹配对象。这样,在经过第二提取模型提取的特征与经过第一提取模型提取的特征不能直接匹配时,可以将第一提取模型提取的特征转换到第二提取模型的特征空间,经过第一提取模型提取的特征就能与经过第二提取模型提取的特征进行匹配,由于特征转换的处理量远远小于重新提取匹配库中的特征,所以可以节约对象识别的时长。
另外,在特征转换后,进行匹配,虽然性能要低于直接使用第一提取模型提取特征时的匹配,但是要高于第二提取模型提取特征时的匹配。
对应图2所示流程,为了更好的理解本公开实施例,还提供了匹配库为证件照匹配库的处理流程框图,如图5所示,匹配库中导入或注册的证件照输入至第二提取模型,获得证件照在第二提取模型的特征空间中的特征。实时抓拍的人脸图片(即为目标对象)输入至第一提取模型获得人脸图片的第一特征,将该第一特征输入至第一特征转换模型,获得人脸图片在第二提取模型的特征空间中的特征。这样,匹配库中的证件照的特征和人脸图像的特征都属于第二提取模型的特征空间,所以可以直接进行匹配。
本公开实施例提供了一种进行对象识别的方法,可以将匹配库的特征转换为可以兼容高版本的提取模型提取的特征,该方法的执行主体可以是服务器,如图6所示,该方法的执行流程可以如下:
步骤601,将匹配库中各对象使用第二提取模型提取的第三特征,分别通过第二特征转换模型转换到第一提取模型的特征空间,得到各对象在特征空间中的第四特征。
在实施中,在进行对象识别时,为了兼容高版本提取模型提取的特征,可以将低版本提取模型提取的特征转换到高版本提取模型的特征空间,第一提取模型为高版本提取模型,第二提取模型为低版本提取模型。可以预先训练一个特征转换模型(即第二特征转换模型),用于将第二提取模型提取的特征转换到 第一提取模型的特征空间。
在有高版本的提取模型出现后,可以使用第二特征转换模型,将匹配库中各对象经过第二提取模型提取的第三特征,转换到第一提取模型的特征空间,这样,就可以得到匹配库中各对象在第一提取模型的特征空间中的第四特征。
需要说明的是,第一提取模型为高版本提取模型,第二提取模型为低版本提取模型,高版本提取模型的性能要高于低版本提取模型的性能。上述将第三特征转换到第一提取模型的特征空间,可以是通过一个非线性的映射函数,将第三特征转换为与第一提取模型的特征空间中的特征维度一样的特征,例如,若匹配库中某个对象直接使用第二提取模型的提取的特征的维度为5维,将第三特征转换到第一提取模型的特征空间后,第四特征的维度也为5维。
步骤602,将目标对象使用第一提取模型提取的特征与各对象在特征空间中的第四特征进行匹配,得到与目标对象使用第一提取模型提取的特征相匹配的目标特征。
在实施中,在对目标对象进行识别时,可以使用第一提取模型提取目标对象的特征,得到目标对象的特征,将目标对象的特征与各对象在第一提取模型的特征空间中的第四特征进行匹配,得到目标对象的特征与各对象在该特征空间中的第四特征的相似度,选取相似度最大,且大于目标数值的目标特征。
步骤603,将目标特征所属的对象,确定为目标对象的匹配对象。
在实施中,在选取出目标特征之后,可以将目标特征所属的对象,确定为目标对象的匹配对象。
这样,由于匹配库中的各对象的特征与目标对象的特征均属于第一提取模型所属的特征空间,所以可以直接进行匹配。
可选的,本公开实施例中,还提供了训练第二特征转换模型的过程,处理可以如下:
获取多个样本对象经过第一提取模型提取的特征和经过第二提取模型提取的特征,将多个样本对象中目标数目个样本对象经过第二提取模型提取的特征,输入到第二初始特征转换模型,得到第五输出结果,确定第五输出结果与目标数目个样本对象经过第一提取模型提取的特征的第五损失值,根据第五损失值、第二初始特征转换模型、多个样本对象经过第一提取模型提取的特征和经过第二提取模型提取的特征,确定第二特征转换模型。
在实施中,可以获取多个样本对象,将多个样本对象输入到第一提取模型,得到多个样本对象经过第一提取模型提取的特征,并将多个样本对象输入到第二提取模型,得到多个样本经过第二提取模型提取的特征。或者,服务器直接从与自身连接的设备获取多个样本对象经过第一提取模型提取的特征和经过第二提取模型提取的特征。
然后将多个样本对象中目标数目个样本对象经过第二提取模型提取的特征,输入到第二初始特征转换模型,得到第五输出结果,计算第五输出结果与该目标数目个样本对象经过第一提取模型提取的特征的第五损失值。使用第五损失值作为约束,使用梯度下降算法,调整第二初始特征转换模型中参数的参数值,然后重新在多个样本对象中选取目标数目个样本对象,执行上述过程,也能得到一个第五损失值,继续调整第二初始特征转换模型中参数的参数值,直到使第五损失值最小,选择使第五损失值最小时的参数的参数值代入第二初始特征转换模型,得到第二特征转换模型。
可选的,为了使第二特征转换模型性能更好,可以使用更多的约束,进行训练,相应的处理可以如下:
将第五输出结果输入到第一初始特征转换模型,得到第六输出结果,确定第六输出结果与目标数目个样本对象经过第二提取模型提取的特征的第六损失值,根据第五损失值、第六损失值、第二初始特征转换模型、多个样本对象经过第一提取模型提取的特征和经过第二提取模型提取的特征,确定第二特征转换模型。
其中,经过第一初始特征转换模型转换后的特征,会从第一提取模型的特征空间转换到第二提取模型的特征空间。
在实施中,在得到第五输出结果之后,可以将第五输出结果输入到第一初始特征转换模型,得到第六输出结果。然后确定第六输出结果与目标数目个样本对象经过第二提取模型提取的特征的第六损失值,然后可以使用梯度下降算法调整第一初始特征转换模型和第二初始特征转换模型中的参数的参数值,直到某一组参数值,可以使第五损失值和第六损失值均最小,即得到第一特征转换模型和第二特征转换模型,后续可以仅使用第二特征转换模型。
这样,由于使用了两个损失进行约束,所以可以使训练出的第二特征转换模型性能更好。
可选的,为了使第二特征转换模型的性能更好,可以使用更多的约束来训练第二特征转换模型,处理可以如下:
将多个样本对象中目标数目个样本对象经过第一提取模型提取的特征,输入到第一初始特征转换模型,得到第七输出结果,确定第七输出结果与目标数目个样本对象经过第二提取模型提取的特征的第七损失值,将第七输出结果输入到第二初始特征转换模型中,得到第八输出结果,确定第八输出结果与目标数目个样本对象经过第一提取模型提取的特征的第八损失值,根据第五损失值、第六损失值、第七损失值、第八损失值、第一初始特征转换模型、第二初始特征转换模型、多个样本对象经过第一提取模型提取的特征和经过第二提取模型提取的特征,确定第二特征转换模型。
在实施中,可以将多个样本对象中目标数目个样本对象经过第一提取模型提取的特征,输入到第一初始特征转换模型中,得到第七输出结果。计算第七输出结果与目标数目个样本对象经过第二提取模型提取的特征的第七损失值。将第七输出结果再输入到第二初始特征转换模型中,得到第八输出结果。确定第八输出结果与目标数目个样本对象经过第一提取模型提取的特征的第八损失值。这样,会获得四个损失值,即第五损失值、第六损失值、第七损失值和第八损失值。
然后可以使用梯度下降算法调整第一初始特征转换模型和第二初始特征转换模型中的参数的参数值,直到某一组参数值,可以使第五损失值、第六损失值、第七损失值和第八损失值均最小,即得到第一特征转换模型和第二特征转换模型,后续可以仅使用第二特征转换模型。
需要说明的是,上述第五损失值、第六损失值、第七损失值和第八损失值中“第五”、“第六”、“第七”和“第八”用于区别类似的对象,而不必用于描述特定的顺序或先后次序。
可选的,上述第五损失值、第六损失值、第七损失值和第八损失值均为L1损失得到的损失值。L1损失用公式表示可以参见前文中的描述。
或者,上述第五损失值、第六损失值、第七损失值和第八损失值均为L2损失得到的损失值。L2损失用公式表示可以参见前文中的描述。
或者,上述第五损失值和第七损失值为L1损失得到的损失值,第六损失值和第八损失值为L2损失得到的损失值。
在实施中,L1损失和L2损失的表达式可以见上述描述,由于第五损失值是实际值与经过一次转换的估计值之间的损失值,第七损失值是实际值与经过一次转换的估计值之间的损失值,所以第五损失值和第七损失值采用L1损失得到的损失值,这是由于L1损失的鲁棒性更好,所以使用L1损失作为约束,训练出的第一特征转换模型和第二特征转换模型的性能更好。
本公开实施例中,服务器获取第二特征转换模型,将匹配库中各对象使用第二提取模型提取的第三特征,分别通过第二特征转换模型转换到第一提取模型的特征空间,得到各对象在特征空间中的第四特征,将目标对象使用第一提取模型提取的特征与各对象在特征空间中的第四特征进行匹配,得到与目标对象使用第一提取模型提取的特征相匹配的目标特征,将目标特征所属的对象,确定为目标对象的匹配对象。这样,在经过第二提取模型提取的特征与经过第一提取模型提取的特征不能直接匹配时,可以将第二提取模型提取的特征转换到第一提取模型的特征空间,经过第一提取模型提取的特征就能与经过第二提取模型提取的特征进行匹配,而不需要重新使用第一提取模型提取匹配库中的特征,所以可以节约对象识别的时长。
另外需要说明的是,上述第二特征转换模型在对某个对象的特征,进行特征换换时,计算量比较小,远低于使用第一提取模型重新提取该对象的特征。
另外,为了更方便理解本公开两个实施例,还提供了本公开两个实施例的思想,由于识别算法本身对特定的对象测试集具有相似的辨别能力,各特征所完成的任务目标也一致,因此,不同版本的提取模型的特征之间虽然不具备直接的线性相关性,但是可以假设存在一个非线性的映射函数,能够将特征从高版本提取模型的特征空间映射到低版本提取模型的特征空间。
由于神经网络是一种优秀的非线性映射器,所以上述第一特征转换模型和第二特征转换模型可以是一种神经网络模型。第一特征转换模型和第二特征转换模型可以是多层的全连接神经网络,结构可以包括归一化层(Norm)、全连接层(Fully connected layer)、批归一化层(batch normalization layer)、双曲正切函数处理层。归一化层可以用于使向量的模长为1,批归一化层的作用是在梯度下降时,收敛更好,训练出的第一特征转换模型使性能更好。双曲正切函数是一个非线性函数,可以用于拟合非线性的映射函数。
例如,第一特征转换模型包括4层全连接神经网络,每层全连接神经网络 包括一个归一化层、一个全连接层、一个批归一化层和一个双曲正切函数层,在第一个全连接层的输入为256个通道数,输出为512个通道数,在第二个全连接层的输入为512个通道数,输出为768个通道数,在第三个全连接层的输入为768个通道数,输出为1024个通道数,在第四个全连接层的输入为1024和通道数,输出为256个通道数。
还需要说明的是,在上述两个实施例中,训练特征转换模型时,每次调整参数的参数值后,会在多个样本对象中,重新选择一批样本对象进行训练。
还需要说明的是,上述两个实施例中,第一提取模型和第二提取模型,是任意的可以用于提取特征的模型,只要第一提取模型的版本高于第二提取模型的版本即可,本公开实施例不做限定。
基于相同的技术构思,本公开实施例还提供了一种进行对象识别的装置,如图7所示,该装置包括:
转换模块710,用于将目标对象使用第一提取模型提取的第一特征,通过第一特征转换模型转换到第二提取模型的特征空间,得到所述目标对象在所述特征空间中的第二特征;
匹配模块720,用于:
将所述目标对象的第二特征,与匹配库中的对象的特征进行匹配,得到与所述目标对象的第二特征相匹配的目标特征;
将所述目标特征所属的对象,确定为所述目标对象的匹配对象。
可选的,如图8所示,所述装置还包括训练模块730,用于:
获取多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征;
将所述多个样本对象中目标数目个样本对象经过所述第一提取模型提取的特征,输入到第一初始特征转换模型,得到第一输出结果;
确定所述第一输出结果与所述目标数目个样本对象经过所述第二提取模型提取的特征的第一损失值;
根据所述第一损失值、所述第一初始特征转换模型、所述多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征,确定所述第一特征转换模型。
可选的,所述训练模块730,还用于:
将所述第一输出结果输入到第二初始特征转换模型,得到第二输出结果;
确定所述第二输出结果与所述目标数目个样本对象经过所述第一提取模型提取的特征的第二损失值;
所述训练模块730,用于:
根据所述第一损失值、所述第二损失值、所述第一初始特征转换模型、所述多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征,确定所述第一特征转换模型。
可选的,所述训练模块730,还用于:
将所述多个样本对象中目标数目个样本对象经过第二提取模型提取的特征,输入到所述第二初始特征转换模型,得到第三输出结果;
确定所述第三输出结果与所述目标数目个样本对象经过第一提取模型提取的特征的第三损失值;
将所述第三输出结果输入到所述第一初始特征转换模型中,得到第四输出结果;
确定所述第四输出结果与所述目标数目个样本对象经过第二提取模型提取的特征的第四损失值;
所述训练模块730,用于:
根据所述第一损失值、所述第二损失值、所述第三损失值、所述第四损失值、所述第一初始特征转换模型、所述第二初始特征转换模型、所述多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征,确定所述第一特征转换模型。
可选的,所述第一损失值、所述第二损失值、所述第三损失值和所述第四损失值均为L1损失得到的损失值;或者,
所述第一损失值、所述第二损失值、所述第三损失值和所述第四损失值均为L2损失得到的损失值;或者,
所述第一损失值和所述第三损失值为L1损失得到的损失值,所述第二损失值和所述第四损失值为L2损失得到的损失值。
本公开实施例中,将目标对象使用第一提取模型提取的第一特征,通过第一特征转换模型转换到第二提取模型的特征空间,得到目标对象在特征空间中 的第二特征,将目标对象的第二特征,与匹配库中的对象的特征进行匹配,得到与目标对象的第二特征相匹配的目标特征,将目标特征所属的对象,确定为目标对象的匹配对象。这样,在经过第一提取模型提取的特征与经过第二提取模型提取的特征不能直接匹配时,可以将第一提取模型提取的特征转换到第二提取模型的特征空间,经过第一提取模型提取的特征就能与经过第二提取模型提取的特征进行匹配,由于特征转换的处理量远远小于重新提取匹配库中的特征,所以可以节约时长。
需要说明的是:上述实施例提供的进行对象识别的装置在进行对象识别时,仅以上述各功能模块的划分进行举例说明,实际应用中,可以根据需要而将上述功能分配由不同的功能模块完成,即将装置的内部结构划分成不同的功能模块,以完成以上描述的全部或者部分功能。另外,上述实施例提供的进行对象识别的装置与进行对象识别的方法实施例属于同一构思,其具体实现过程详见方法实施例,这里不再赘述。
基于相同的技术构思,本公开实施例还提供了一种进行对象识别的装置,如图9所示,该装置包括:
转换模块910,用于将匹配库中各对象使用第二提取模型提取的第三特征,分别通过第二特征转换模型转换到第一提取模型的特征空间,得到所述各对象在所述特征空间中的第四特征;
匹配模块920,用于:
将目标对象使用所述第一提取模型提取的特征与所述各对象在所述特征空间中的第四特征进行匹配,得到与所述目标对象使用所述第一提取模型提取的特征相匹配的目标特征;
将所述目标特征所属的对象,确定为所述目标对象的匹配对象。
可选的,如图10所示,所述装置还包括训练模块930,用于:
获取多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征;
将所述多个样本对象中目标数目个样本对象经过第二提取模型提取的特征,输入到第二初始特征转换模型,得到第五输出结果;
确定所述第五输出结果与所述目标数目个样本对象经过第一提取模型提取 的特征的第五损失值;
根据所述第五损失值、所述第二初始特征转换模型、所述多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征,确定所述第二特征转换模型。
可选的,所述训练模块930,还用于:
将所述第五输出结果输入到第一初始特征转换模型,得到第六输出结果;
确定所述第六输出结果与所述目标数目个样本对象经过第二提取模型提取的特征的第六损失值;
所述训练模块930,用于:
根据所述第五损失值、所述第六损失值、所述第二初始特征转换模型、所述多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征,确定所述第二特征转换模型。
可选的,所述训练模块930,还用于:
将所述多个样本对象中目标数目个样本对象经过第一提取模型提取的特征,输入到所述第一初始特征转换模型,得到第七输出结果;
确定所述第七输出结果与所述目标数目个样本对象经过第二提取模型提取的特征的第七损失值;
将所述第七输出结果输入到所述第二初始特征转换模型中,得到第八输出结果;
确定所述第八输出结果与所述目标数目个样本对象经过第一提取模型提取的特征的第八损失值;
所述训练模块930,用于:
根据所述第五损失值、所述第六损失值、第七损失值、第八损失值、所述第一初始特征转换模型、所述第二初始特征转换模型、所述多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征,确定所述第二特征转换模型。
可选的,所述第五损失值、所述第六损失值、所述第七损失值和所述第八损失值均为L1损失得到的损失值;或者,
所述第五损失值、所述第六损失值、所述第七损失值和所述第八损失值均为L2损失得到的损失值;或者,
所述第五损失值和所述第七损失值为L1损失得到的损失值,所述第六损失值和所述第八损失值为L2损失得到的损失值。
本公开实施例中,获取第二特征转换模型,将匹配库中各对象使用第二提取模型提取的第三特征,分别通过第二特征转换模型转换到第一提取模型的特征空间,得到各对象在特征空间中的第四特征,将目标对象使用第一提取模型提取的特征与各对象在特征空间中的第四特征进行匹配,得到与目标对象使用第一提取模型提取的特征相匹配的目标特征,将目标特征所属的对象,确定为目标对象的匹配对象。这样,在经过第二提取模型提取的特征与经过第一提取模型提取的特征不能直接匹配时,可以将第二提取模型提取的特征转换到第一提取模型的特征空间,经过第一提取模型提取的特征就能与经过第二提取模型提取的特征进行匹配,而不需要重新使用第一提取模型提取匹配库中的特征,所以可以节约时长。
需要说明的是:上述实施例提供的进行对象识别的装置在进行对象识别时,仅以上述各功能模块的划分进行举例说明,实际应用中,可以根据需要而将上述功能分配由不同的功能模块完成,即将装置的内部结构划分成不同的功能模块,以完成以上描述的全部或者部分功能。另外,上述实施例提供的进行对象识别的装置与进行对象识别的方法实施例属于同一构思,其具体实现过程详见方法实施例,这里不再赘述。
图11是本公开实施例提供的一种服务器的结构示意图,该服务器1100可因配置或性能不同而产生比较大的差异,可以包括一个或一个以上处理器1101和一个或一个以上的存储器1102,其中,所述存储器1102中存储有至少一条指令,所述至少一条指令由所述处理器1101加载并执行以实现上述进行对象识别的方法中的各个步骤。
本公开实施例中,还提供了一种计算机可读存储介质,该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,实现上述进行对象识别的方法。
本公开实施例中,还提供了一种进行对象识别的服务器,该服务器包括处理器和存储器,其中,所述存储器,用于存放计算机程序;所述处理器,用于 执行所述存储器上所存放的程序,实现上述进行对象识别的方法。
本领域普通技术人员可以理解实现上述实施例的全部或部分步骤可以通过硬件来完成,也可以通过程序来指令相关的硬件完成,所述的程序可以存储于一种计算机可读存储介质中,上述提到的存储介质可以是只读存储器,磁盘或光盘等。
以上所述仅为本公开的较佳实施例,并不用以限制本公开,凡在本公开的精神和原则之内,所作的任何修改、等同替换、改进等,均应包含在本公开的保护范围之内。

Claims (14)

  1. 一种进行对象识别的方法,其特征在于,所述方法包括:
    将目标对象使用第一提取模型提取的第一特征,通过第一特征转换模型转换到第二提取模型的特征空间,得到所述目标对象在所述特征空间中的第二特征;
    将所述目标对象的第二特征,与匹配库中的对象的特征进行匹配,得到与所述目标对象的第二特征相匹配的目标特征;
    将所述目标特征所属的对象,确定为所述目标对象的匹配对象。
  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:
    获取多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征;
    将所述多个样本对象中目标数目个样本对象经过所述第一提取模型提取的特征,输入到第一初始特征转换模型,得到第一输出结果;
    确定所述第一输出结果与所述目标数目个样本对象经过所述第二提取模型提取的特征的第一损失值;
    根据所述第一损失值、所述第一初始特征转换模型、所述多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征,确定所述第一特征转换模型。
  3. 根据权利要求2所述的方法,其特征在于,所述方法还包括:
    将所述第一输出结果输入到第二初始特征转换模型,得到第二输出结果;
    确定所述第二输出结果与所述目标数目个样本对象经过所述第一提取模型提取的特征的第二损失值;
    所述根据所述第一损失值、所述第一初始特征转换模型、所述多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征,确定所述第一特征转换模型,包括:
    根据所述第一损失值、所述第二损失值、所述第一初始特征转换模型、所述多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征,确定所述第一特征转换模型。
  4. 根据权利要求3所述的方法,其特征在于,所述方法还包括:
    将所述多个样本对象中目标数目个样本对象经过第二提取模型提取的特征,输入到所述第二初始特征转换模型,得到第三输出结果;
    确定所述第三输出结果与所述目标数目个样本对象经过第一提取模型提取的特征的第三损失值;
    将所述第三输出结果输入到所述第一初始特征转换模型中,得到第四输出结果;
    确定所述第四输出结果与所述目标数目个样本对象经过第二提取模型提取的特征的第四损失值;
    所述根据所述第一损失值、所述第二损失值、所述第一初始特征转换模型、所述多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征,确定所述第一特征转换模型,包括:
    根据所述第一损失值、所述第二损失值、所述第三损失值、所述第四损失值、所述第一初始特征转换模型、所述第二初始特征转换模型、所述多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征,确定所述第一特征转换模型。
  5. 根据权利要求4所述的方法,其特征在于,所述第一损失值、所述第二损失值、所述第三损失值和所述第四损失值均为L1损失得到的损失值;或者,
    所述第一损失值、所述第二损失值、所述第三损失值和所述第四损失值均为L2损失得到的损失值;或者,
    所述第一损失值和所述第三损失值为L1损失得到的损失值,所述第二损失值和所述第四损失值为L2损失得到的损失值。
  6. 一种进行对象识别的方法,其特征在于,所述方法包括:
    将匹配库中各对象使用第二提取模型提取的第三特征,分别通过第二特征转换模型转换到第一提取模型的特征空间,得到所述各对象在所述特征空间中的第四特征;
    将目标对象使用所述第一提取模型提取的特征与所述各对象在所述特征空间中的第四特征进行匹配,得到与所述目标对象使用所述第一提取模型提取的特征相匹配的目标特征;
    将所述目标特征所属的对象,确定为所述目标对象的匹配对象。
  7. 根据权利要求6所述的方法,其特征在于,所述方法还包括:
    获取多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征;
    将所述多个样本对象中目标数目个样本对象经过第二提取模型提取的特征,输入到第二初始特征转换模型,得到第五输出结果;
    确定所述第五输出结果与所述目标数目个样本对象经过第一提取模型提取的特征的第五损失值;
    根据所述第五损失值、所述第二初始特征转换模型、所述多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征,确定所述第二特征转换模型。
  8. 根据权利要求7所述的方法,其特征在于,所述方法还包括:
    将所述第五输出结果输入到第一初始特征转换模型,得到第六输出结果;
    确定所述第六输出结果与所述目标数目个样本对象经过第二提取模型提取的特征的第六损失值;
    所述根据所述第五损失值、所述第二初始特征转换模型、所述多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征,确定所述第二特征转换模型,包括:
    根据所述第五损失值、所述第六损失值、所述第二初始特征转换模型、所述多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征,确定所述第二特征转换模型。
  9. 根据权利要求8所述的方法,其特征在于,所述方法还包括:
    将所述多个样本对象中目标数目个样本对象经过第一提取模型提取的特征,输入到所述第一初始特征转换模型,得到第七输出结果;
    确定所述第七输出结果与所述目标数目个样本对象经过第二提取模型提取的特征的第七损失值;
    将所述第七输出结果输入到所述第二初始特征转换模型中,得到第八输出结果;
    确定所述第八输出结果与所述目标数目个样本对象经过第一提取模型提取的特征的第八损失值;
    所述根据所述第五损失值、所述第六损失值、所述第二初始特征转换模型、 所述多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征,确定所述第二特征转换模型,包括:
    根据所述第五损失值、所述第六损失值、第七损失值、第八损失值、所述第一初始特征转换模型、所述第二初始特征转换模型、所述多个样本对象经过所述第一提取模型提取的特征和经过所述第二提取模型提取的特征,确定所述第二特征转换模型。
  10. 根据权利要求9所述的方法,其特征在于,所述第五损失值、所述第六损失值、所述第七损失值和所述第八损失值均为L1损失得到的损失值;或者,
    所述第五损失值、所述第六损失值、所述第七损失值和所述第八损失值均为L2损失得到的损失值;或者,
    所述第五损失值和所述第七损失值为L1损失得到的损失值,所述第六损失值和所述第八损失值为L2损失得到的损失值。
  11. 一种进行对象识别的装置,其特征在于,所述装置包括:
    转换模块,用于将目标对象使用第一提取模型提取的第一特征,通过第一特征转换模型转换到第二提取模型的特征空间,得到所述目标对象在所述特征空间中的第二特征;
    匹配模块,用于:
    将所述目标对象的第二特征,与匹配库中的对象的特征进行匹配,得到与所述目标对象的第二特征相匹配的目标特征;
    将所述目标特征所属的对象,确定为所述目标对象的匹配对象。
  12. 一种进行对象识别的装置,其特征在于,所述装置包括:
    转换模块,用于将匹配库中各对象使用第二提取模型提取的第三特征,分别通过第二特征转换模型转换到第一提取模型的特征空间,得到所述各对象在所述特征空间中的第四特征;
    匹配模块,用于:
    将目标对象使用所述第一提取模型提取的特征与所述各对象在所述特征空间中的第四特征进行匹配,得到与所述目标对象使用所述第一提取模型提取的特征相匹配的目标特征;
    将所述目标特征所属的对象,确定为所述目标对象的匹配对象。
  13. 一种进行对象识别的服务器,其特征在于,包括处理器和存储器;所述存储器,用于存放至少一条指令;所述处理器,用于执行所述存储器上所存放的至少一条指令,实现权利要求1-10任一项所述的方法步骤。
  14. 一种计算机可读存储介质,其特征在于,所述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1-10任一项所述的方法步骤。
PCT/CN2020/105500 2019-07-30 2020-07-29 进行对象识别的方法和装置 WO2021018189A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
EP20846719.1A EP4006775A4 (en) 2019-07-30 2020-07-29 METHOD AND DEVICE FOR OBJECT RECOGNITION
US17/631,446 US20220277588A1 (en) 2019-07-30 2020-07-29 Methods and apparatuses for performing object recognition

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910696479.XA CN112308101B (zh) 2019-07-30 2019-07-30 进行对象识别的方法和装置
CN201910696479.X 2019-07-30

Publications (1)

Publication Number Publication Date
WO2021018189A1 true WO2021018189A1 (zh) 2021-02-04

Family

ID=74228391

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/105500 WO2021018189A1 (zh) 2019-07-30 2020-07-29 进行对象识别的方法和装置

Country Status (4)

Country Link
US (1) US20220277588A1 (zh)
EP (1) EP4006775A4 (zh)
CN (1) CN112308101B (zh)
WO (1) WO2021018189A1 (zh)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105975861A (zh) * 2016-05-27 2016-09-28 百度在线网络技术(北京)有限公司 应用检测方法和装置
CN108806716A (zh) * 2018-06-15 2018-11-13 想象科技(北京)有限公司 用于基于情感框架的计算机化匹配的方法与装置
US10140713B1 (en) * 2017-07-19 2018-11-27 Schwalb Consulting, LLC Morphology identification in tissue samples based on comparison to named feature vectors

Family Cites Families (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5031225A (en) * 1987-12-09 1991-07-09 Ricoh Company, Ltd. Character recognition method for recognizing character in an arbitrary rotation position
CN101165682A (zh) * 2006-10-20 2008-04-23 北京北大方正电子有限公司 一种文档文件的处理方法及系统
KR20190041790A (ko) * 2017-10-13 2019-04-23 한국전자통신연구원 신경망 번역 모델 구축 장치 및 방법
CN107832058B (zh) * 2017-10-31 2022-01-18 北京小米移动软件有限公司 数据处理方法及装置
CN108108357B (zh) * 2018-01-12 2022-08-09 京东方科技集团股份有限公司 口音转换方法及装置、电子设备
CN108509539B (zh) * 2018-03-16 2021-08-17 联想(北京)有限公司 信息处理方法电子设备
CN108595927B (zh) * 2018-04-04 2023-09-19 北京市商汤科技开发有限公司 身份认证、解锁及支付方法、装置、存储介质、产品和设备
CN108765295B (zh) * 2018-06-12 2019-11-26 腾讯科技(深圳)有限公司 图像处理方法、图像处理装置及存储介质
CN110009018B (zh) * 2019-03-25 2023-04-18 腾讯科技(深圳)有限公司 一种图像生成方法、装置以及相关设备

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105975861A (zh) * 2016-05-27 2016-09-28 百度在线网络技术(北京)有限公司 应用检测方法和装置
US10140713B1 (en) * 2017-07-19 2018-11-27 Schwalb Consulting, LLC Morphology identification in tissue samples based on comparison to named feature vectors
CN108806716A (zh) * 2018-06-15 2018-11-13 想象科技(北京)有限公司 用于基于情感框架的计算机化匹配的方法与装置

Also Published As

Publication number Publication date
EP4006775A4 (en) 2022-08-31
CN112308101B (zh) 2023-08-22
CN112308101A (zh) 2021-02-02
US20220277588A1 (en) 2022-09-01
EP4006775A1 (en) 2022-06-01

Similar Documents

Publication Publication Date Title
WO2010004958A1 (ja) 個人認証システム、個人認証方法
WO2020019591A1 (zh) 用于生成信息的方法和装置
US11126827B2 (en) Method and system for image identification
CN108549848B (zh) 用于输出信息的方法和装置
CN112801054B (zh) 人脸识别模型的处理方法、人脸识别方法及装置
CN111832581B (zh) 肺部特征识别方法、装置、计算机设备及存储介质
CN113191461B (zh) 一种图片识别方法、装置、设备及可读存储介质
KR20200083119A (ko) 사용자 인증 장치 및 방법
CN113298158A (zh) 数据检测方法、装置、设备及存储介质
CN111108508A (zh) 脸部情感识别方法、智能装置和计算机可读存储介质
KR20200110064A (ko) 변환 모델을 이용한 인증 방법 및 장치
CN115240280A (zh) 人脸活体检测分类模型的构建方法、检测分类方法及装置
Rai et al. Software development framework for real-time face detection and recognition in mobile devices
Abate et al. Smartphone enabled person authentication based on ear biometrics and arm gesture
CN116152938A (zh) 身份识别模型训练和电子资源转移方法、装置及设备
WO2021018189A1 (zh) 进行对象识别的方法和装置
Geetha et al. 3D face recognition using Hadoop
CN110633647A (zh) 活体检测方法及装置
KR20160042646A (ko) 얼굴 인식 방법
Gupta et al. Integration of biometric security system to improve the protection of digital wallet
Kumar et al. Cross domain descriptor for face sketch-photo image recognition
KR101494309B1 (ko) 디지털 컨텐츠를 위한 비대칭형 핑거프린트 정합 시스템 및 그 제공방법
KR102027786B1 (ko) 다수의 얼굴 이미지를 동시에 사용하는 얼굴 인식 및 인증 방법 및 시스템
Soviany et al. A multimodal biometric identification method for mobile applications security
AbdulAbaus et al. Person identification based on facial biometrics in different lighting conditions

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: 20846719

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: 2020846719

Country of ref document: EP

Effective date: 20220228