WO2019228089A1 - 人体属性识别方法、装置、设备及介质 - Google Patents

人体属性识别方法、装置、设备及介质 Download PDF

Info

Publication number
WO2019228089A1
WO2019228089A1 PCT/CN2019/082977 CN2019082977W WO2019228089A1 WO 2019228089 A1 WO2019228089 A1 WO 2019228089A1 CN 2019082977 W CN2019082977 W CN 2019082977W WO 2019228089 A1 WO2019228089 A1 WO 2019228089A1
Authority
WO
WIPO (PCT)
Prior art keywords
attribute
human
neural network
convolutional neural
image
Prior art date
Application number
PCT/CN2019/082977
Other languages
English (en)
French (fr)
Inventor
杨思骞
李季檩
吴永坚
晏轶超
贺珂珂
葛彦昊
黄飞跃
汪铖杰
Original Assignee
腾讯科技（深圳）有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 腾讯科技（深圳）有限公司 filed Critical 腾讯科技（深圳）有限公司
Publication of WO2019228089A1 publication Critical patent/WO2019228089A1/zh
Priority to US16/938,858 priority Critical patent/US11275932B2/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
    • G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
    • G06V10/449—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters
    • G06V10/451—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters with interaction between the filter responses, e.g. cortical complex cells
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/50—Context or environment of the image
    • G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/103—Static body considered as a whole, e.g. static pedestrian or occupant recognition
    • G—PHYSICS
    • G07—CHECKING-DEVICES
    • G07F—COIN-FREED OR LIKE APPARATUS
    • G07F19/00—Complete banking systems; Coded card-freed arrangements adapted for dispensing or receiving monies or the like and posting such transactions to existing accounts, e.g. automatic teller machines
    • G07F19/20—Automatic teller machines [ATMs]
    • G07F19/209—Monitoring, auditing or diagnose of functioning of ATMs

Definitions

  • the present application relates to the field of image recognition technology, and in particular, to a method, a device, a device, and a medium for identifying human attributes.
  • Human attribute recognition is the identification of people's gender, age, clothing type and color, etc., and is widely used in dangerous behavior warning, traffic violation monitoring, industrial security and vending machines, automatic teller machines (ATM), shopping malls And target areas such as public places such as stations.
  • ATM automatic teller machines
  • Examples of the present application provide a method, a device, a device, and a medium for identifying human attributes.
  • the example of this application provides a method for identifying human attributes, which is executed by a computing device and includes:
  • the human body area image is input into a multi-attribute convolutional neural network model, and the probability that each human body attribute in the human body area image corresponds to each of the predefined attribute values is obtained; among them, the multi-attribute convolutional neural network model uses a multi-attribute convolutional neural network
  • the network is obtained by performing multi-attribute recognition training on the training images obtained in advance;
  • the attribute value of each human attribute in the human area image is determined.
  • the example of the present application also provides a device for identifying human attributes, including:
  • a determining unit configured to determine an image of a human body region in the image
  • An obtaining unit is configured to input a human area image into a multi-attribute convolutional neural network model, and obtain a probability that each human property in the human area image corresponds to each attribute value that is predefined; wherein the multi-attribute convolutional neural network model uses The multi-attribute convolutional neural network is obtained by performing multi-attribute recognition training on the training image obtained in advance;
  • the recognition unit is configured to determine the attribute value of each human attribute in the human body area image based on the probability that each human attribute in the human body area image corresponds to each attribute value defined in advance.
  • the example of the present application also provides a device for identifying human attributes, which includes a memory and a processor.
  • the memory stores computer-readable instructions, and the instructions can enable the processor to:
  • the human body area image is input into a multi-attribute convolutional neural network model, and the probability that each human body property in the human body area image corresponds to each attribute value defined in advance is obtained, wherein the multi-attribute convolutional neural network model is The multi-attribute convolutional neural network is used to perform multi-attribute recognition training on the training image obtained in advance;
  • An attribute value of each human attribute in the human body region image is determined based on a probability that each human attribute in the human body region image corresponds to each attribute value defined in advance.
  • the example of the present application also provides a storage medium that stores processor-executable instructions. When the instructions are executed by one or more processors, the steps of the foregoing human body attribute recognition method are implemented.
  • FIG. 1A is a schematic flowchart of a training method for a multi-attribute convolutional neural network model provided by an example of this application;
  • FIG. 1A is a schematic flowchart of a training method for a multi-attribute convolutional neural network model provided by an example of this application;
  • FIG. 1B is a comparison diagram of input and output when using a multi-attribute convolutional neural network model provided in the example of the present application to recognize human attributes and input and output when using a multi-attribute and multi-model approach to recognize human attributes;
  • FIG. 2 is a schematic flowchart of a human attribute recognition method provided by an example of this application.
  • FIG. 3 is a schematic flowchart of a human attribute recognition method provided by an example of the present application when “traffic violation monitoring” is used as a specific application scenario;
  • FIG. 4 is a functional structure diagram of a human attribute recognition device provided by an example of the present application.
  • FIG. 5 is a schematic diagram of a hardware structure of a human attribute recognition device provided by an example of the present application.
  • the target person is usually locked according to multiple human attributes. It can be seen that the attribute values of multiple human attributes are accurately identified, which is particularly important for the target person's locking.
  • multiple attributes and multiple models are mainly used to identify individual human attributes. Specifically, an attribute recognition model is established for each human attribute, that is, one human attribute corresponds to one attribute recognition model. In this way, during the video surveillance process, real-time The monitoring image is collected and the human body area image in the monitoring image is detected, so that each human body property in the human body area image is identified through the established attribute recognition model, thereby achieving the purpose of multi-attribute recognition.
  • the inventor of the present application thought of using a multi-attribute convolutional neural network to perform multi-attribute recognition training on a training image obtained in advance, so as to establish a multi-attribute convolutional neural network model.
  • the monitoring image can be collected in real time, and when the human body area image in the monitoring image is detected, the multi-attribute convolutional neural network model is used to identify each human body property in the human area image at the same time, thereby achieving the simultaneous recognition of multiple human body properties
  • the function effectively improves the recognition efficiency of multiple human attributes.
  • the human attribute recognition method provided in the examples of this application can be applied to a variety of video surveillance scenarios, such as dangerous behavior monitoring, traffic violation monitoring, industrial security monitoring and vending machines, ATMs, shopping malls and stations in public places.
  • video surveillance scenarios such as dangerous behavior monitoring, traffic violation monitoring, industrial security monitoring and vending machines, ATMs, shopping malls and stations in public places.
  • the above-mentioned application scenarios are shown only to facilitate understanding of the spirit and principle of the application, and the examples of the application are not limited in this regard. Instead, the examples of this application can be applied to any scenario where applicable.
  • it can be applied to offline retail scenarios such as shopping malls and stores, to identify attribute information of customers entering and passing customers, and to collect consumer portraits, thereby assisting precision marketing, personalized recommendations, store location, and trend analysis.
  • it can be applied to the intelligent upgrade of buildings, outdoor advertising, etc., collecting human body information, analyzing crowd attributes, and targeted advertising materials, thereby improving user experience and business efficiency.
  • Human attributes are human visual characteristics that can be perceived by both computers and people, such as gender, age, top texture, top color, sleeve type, bottom texture, bottom color, bottom length, bottom type, etc.
  • the attribute value of the human attribute is the value assigned to the human attribute.
  • the attribute value of the human attribute gender is male and female
  • the attribute value of the human attribute age is infant, child, youth, middle-aged, elderly, etc.
  • Attribute The attribute value of the coat texture is solid color, horizontal stripes, vertical stripes, plaids, large color blocks, etc.
  • the human attribute the attribute value of the coat color is red, yellow, blue, black, mixed colors, etc.
  • the attribute recognition model is a model capable of identifying attribute values of multiple human attributes at the same time.
  • the example in this application uses the attribute recognition model as a multi-attribute convolutional neural network model as an example for illustration.
  • the multi-attribute convolutional neural network model is an attribute recognition model that can be used to recognize multiple human attributes at the same time by using multi-attribute convolutional neural networks to perform multi-attribute recognition training on previously obtained training images.
  • Multi-attribute convolutional neural network a feedforward neural network for deep learning, including data input layer, convolution calculation layer, incentive layer and fully connected layer, or, including data input layer, convolution calculation layer, incentive layer and Global average pooling layer.
  • the human attribute recognition method provided by the example of the present application uses a multi-attribute convolutional neural network model, and the multi-attribute convolutional neural network model uses multi-attribute convolutional neural network to perform multi-attribute recognition on the training image obtained in advance.
  • the training is obtained, so, next, a training method of a multi-attribute convolutional neural network model according to an exemplary embodiment of the present application will be described. Referring to FIG. 1A, the process of training a multi-attribute convolutional neural network model is as follows:
  • Step 101 Collect training images. Specifically, a monitoring image can be directly intercepted from the stored monitoring video as a training image.
  • Step 102 Define each human attribute and attribute values corresponding to each human attribute.
  • the human attributes are defined as gender, age, face orientation, shirt texture, shirt color, sleeve type, bag type, shirt color, shirt length, shirt type.
  • Step 103 Determine a human attribute vector according to the attribute value of each human attribute.
  • the attribute value of each of the above attributes can also be dichotomized and assigned, for example, whether it is male, 1 if it is male, 0 if it is not male, and 99 if it is unrecognizable; similarly, whether it is Green is assigned 1 if it is green, 0 if it is not green, and 99 if it is not recognized; and so on, you can get the assignment of other attributes, and then get a 1 ⁇ 30 human attribute vector. For example, by performing the above processing on the training image or the test image, the following vectors can be obtained respectively:
  • Imgae_i represents the relative path of picture i
  • 0 and 1 represent the corresponding binary attribute values
  • 99 represents unrecognizable.
  • Step 104 Input the training image and the human attribute vector into the data input layer of the multi-attribute convolutional neural network.
  • Step 105 The data input layer of the multi-attribute convolutional neural network performs pre-processing such as de-averaging, normalization, principal component analysis (PCV), and whitening on the training image.
  • pre-processing such as de-averaging, normalization, principal component analysis (PCV), and whitening on the training image.
  • Step 106 The convolution calculation layer of the multi-attribute convolutional neural network performs feature extraction on the training image output from the data input layer to obtain each feature matrix corresponding to the training image.
  • Step 107 The excitation layer of the multi-attribute convolutional neural network performs non-linear mapping processing on each feature matrix output from the convolution calculation layer, thereby mapping the feature values in each feature matrix to a certain range.
  • a Sigmoid function a Tanh function (Hyperbolic Tangent Function), a ReLU (Rectified Linear Unit) function, etc. may be adopted as the excitation function.
  • the ReLU function can be used as the excitation function to perform non-linear mapping processing on each feature matrix output by the convolution calculation layer.
  • Step 108 The fully connected layer or the global average pooling layer of the multi-attribute convolutional neural network obtains the probability that each human attribute in the training image corresponds to each attribute value that is predefined according to each feature matrix output from the excitation layer.
  • the following uses only the fully connected layer of the multi-attribute convolutional neural network to determine the probability that each human attribute in the training image corresponds to each predefined attribute value as an example. Specifically, it can be adopted but not limited to the following: A bias amount and a weight matrix for each body attribute, and according to each feature matrix output from the excitation layer and a preset bias amount and each weight matrix, it is determined that each human attribute in the training image corresponds to each of the predefined The probability of an attribute value.
  • Step 109 Determine the predicted attribute value of each human attribute in the training image based on the probability that each human attribute in the training image corresponds to each predefined attribute value, and according to the predicted attribute value of each human attribute in the training image, The vector of prediction attribute values that make up the training image.
  • the following manner may be adopted: For each human attribute of, from the probability that the human attribute corresponds to each of the predefined attribute values, the attribute value with the highest probability is selected as the predicted attribute value of the human attribute.
  • Step 110 Determine the degree of difference between the predicted attribute value of each human attribute in the predicted attribute value vector output by the fully connected layer or the global average pooling layer and the real attribute value of each predefined human attribute.
  • a cross-entropy loss function as shown in the following formula (1) may be used to determine:
  • L represents the value of the cross-entropy loss function, that is, the degree of difference
  • n represents the number of training images
  • x represents the x-th training image
  • m represents the number of predefined human attributes.
  • y i and a i respectively represent the true attribute value and the predicted attribute value of the ith personal attribute.
  • Step 111 Adjust the network parameters of each layer used in the training process according to the determined difference degree.
  • the network parameters of each layer include, but are not limited to, kernel parameters and initial bias matrices of each convolutional calculation layer, parameters of each excitation layer, parameters of each fully connected layer or global average pooling layer, and the like.
  • Step 112 Use the multi-attribute convolutional neural network after adjusting network parameters to perform multi-attribute recognition training on subsequent training pictures, and so on until the difference between the predicted attribute value of each human attribute and the real attribute value is not greater than the pre- Set the threshold.
  • the network parameters of each layer corresponding to the multi-attribute convolutional neural network are the optimal values.
  • the training process of the multi-attribute convolutional neural network model is ended, and the multi-attributes with the optimal network parameters at each layer are obtained. Convolutional neural network model.
  • the new training image and the original training image can be selected as training images according to a set ratio, and these training images are processed using a multi-attribute convolutional neural network.
  • Multi-attribute recognition training a multi-attribute convolutional neural network model with new human attributes and / or new attribute values is obtained.
  • the method of adding the original training image to the new training image to increase the number of training images of new human attributes and / or new attribute values can effectively avoid inaccurate training results due to the small number of new training images.
  • This problem indirectly improves the accuracy of the multi-attribute convolutional neural network model in identifying human attributes.
  • a comparison diagram of input and output when using a multi-attribute convolutional neural network model provided in the example of the present application to identify human attributes and input and output when using a multi-attribute multi-model method to identify human attributes is compared with multi-attributes.
  • the human attribute recognition method of the model The multi-attribute convolutional neural network model provided by the example of this application can identify multiple human attributes at the same time, which not only solves the problem that the human attribute recognition method of the multi-attribute multi-model cannot identify multiple human attributes at the same time. The problem also improves the recognition efficiency of multiple human attributes as much as possible.
  • the multi-attribute convolutional neural network model can be used to identify multiple human attributes at the same time.
  • the human attributes of the exemplary embodiment of the present application The process of the identification method is as follows. This method can be performed by a human attribute recognition device:
  • Step 201 Determine a human body region image in the monitoring image.
  • Step 202 The human body area image is input into a multi-attribute convolutional neural network model, and the probability that each human body property in the human body area image corresponds to each attribute value defined in advance is obtained.
  • the multi-attribute convolutional neural network model is obtained by using multi-attribute convolutional neural network to perform multi-attribute recognition training on the training image obtained in advance. Specifically, the multi-attribute convolutional neural network model training method and the training method described above are used. The same is not repeated here.
  • Step 203 Determine the attribute value of each human attribute in the human body area image based on the probability that each human attribute in the human body area image corresponds to each of the predefined attribute values.
  • an attribute value with the highest probability is selected from the probability that the human attribute corresponds to each of the predefined attribute values, as the attribute value of the human attribute.
  • each human body property in the human body area image can be identified simultaneously through a multi-attribute convolutional neural network model, thereby achieving multiple human body properties.
  • the simultaneous recognition function effectively improves the recognition efficiency of multiple human attributes.
  • Step 301 intercept a surveillance image from a surveillance video collected by a camera.
  • Step 302 Determine a body region image in the monitoring image.
  • Step 303 Input the human body area image into a multi-attribute convolutional neural network model, and obtain the probability that each human body property in the human body area image corresponds to each attribute value defined in advance, and based on each human body property in the human body area image corresponding to the The probability of each attribute value defined defines the attribute value of each human attribute in the human area image.
  • the attribute value with the highest probability is selected as the attribute value of the human attribute, thereby obtaining the human region image. Attribute value of each human attribute.
  • Step 304 The target person in the monitoring image is locked according to the attribute value of each human attribute in the human body area image, and information such as the attribute value of each human attribute of the target person and traffic violation behavior is marked on the monitored image.
  • the target person when it is determined that the target person does indeed have traffic violations, select the violation violation in the violation handling drop-down box, and generate traffic according to the identified attribute values of various human attributes of the target person and traffic violations. Records of violations, as well as information about the attributes of individual human attributes and traffic violations on the surveillance image.
  • the relevant person is dispatched to the police for traffic violations of the target person, you can also handle the violations drop-down box.
  • the police have been selected to end the handling of this traffic violation; when it is determined that the target person does not have a traffic violation, you can choose to ignore the violation in the violation handling drop-down box and end the handling of this traffic violation. At that time, no traffic violation record is generated.
  • the human attribute recognition device includes at least:
  • a determining unit 401 configured to determine a body region image in a monitoring image
  • An obtaining unit 402 is configured to input a body region image into a multi-attribute convolutional neural network model, and obtain a probability that each human attribute in the body region image corresponds to each attribute value defined in advance; wherein the multi-attribute convolutional neural network model is The multi-attribute convolutional neural network is used to perform multi-attribute recognition training on the training image obtained in advance;
  • the recognition unit 403 is configured to determine the attribute values of the individual human attributes in the human body area image based on the probability that each human attribute in the human body area image corresponds to each of the predefined attribute values.
  • the human attribute recognition device further includes:
  • a training unit 404 is configured to perform a pre-preparation on the training image according to the predefined attribute values of various human attributes and sequentially pass through the data input layer, convolution calculation layer, incentive layer, and fully connected layer of the multi-attribute convolutional neural network.
  • Multi-attribute recognition training to obtain a multi-attribute convolutional neural network model; or, based on the pre-defined attribute values of each human attribute, and sequentially through the data input layer, convolution calculation layer, incentive layer, and global
  • the average pooling layer is used to perform multi-attribute recognition training on the training image obtained in advance to obtain a multi-attribute convolutional neural network model.
  • the human attribute recognition device further includes:
  • the adjusting unit 405 is configured to determine a degree of difference between the predicted attribute value of each human attribute of the training image output by the multi-attribute convolutional neural network model and a real attribute value of each human attribute defined in advance, and according to the degree of difference, The predicted attribute value is determined based on the probability that each human attribute in the training image corresponds to each of the predefined attribute values; adjusting the network parameters of each layer corresponding to the multi-attribute convolutional neural network model.
  • the adjustment unit 405 when determining the degree of difference between the predicted attribute value of each human attribute of the training image output by the multi-attribute convolutional neural network model and the real attribute value of each predefined human attribute, the adjustment unit 405 is specifically configured to: :
  • a cross-entropy loss function is used to determine the degree of difference between the predicted attribute value of each human attribute and the real attribute value of each predefined human attribute.
  • the training unit 404 is further configured to:
  • a new training image and the original training image are selected as training images according to a set ratio
  • the multi-attribute convolutional neural network is used for multi-attribute recognition training on the training image to obtain a multi-attribute convolutional neural network model with new human attributes and / or new attribute values.
  • the identifying unit 403 when determining the attribute values of the individual human attributes in the human area image based on the probability that each human attribute in the human area image corresponds to each of the predefined attribute values, is specifically configured to:
  • the attribute value with the highest probability is selected as the attribute value of the human attribute.
  • the implementation of the human attribute recognition device can refer to the implementation of the human attribute recognition method, and the repeated description is omitted.
  • the multi-attribute convolutional neural network model is used to simultaneously analyze each human attribute in the human region image The recognition is performed, thereby realizing the function of simultaneously identifying multiple human attributes, and effectively improving the recognition efficiency of multiple human attributes.
  • the example of the present application also provides a human attribute recognition device.
  • the human attribute recognition device includes at least: a memory 501, a processor 502, and a computer program stored on the memory 502, and the processor 502 executes the The computer program implements the steps of the above-mentioned human attribute recognition method.
  • the human attribute recognition device may further include an input device 503, an output device 504, and the like.
  • the input device 503 may include a stylus, a keyboard, a mouse, a touch screen, and the like;
  • the output device 504 may include a display device, such as a liquid crystal display (LCD), a cathode ray tube (CRT), a touch screen, and the like.
  • LCD liquid crystal display
  • CRT cathode ray tube
  • the examples in this application are not limited to a specific connection medium between the memory 501, the processor 502, the input device 503, and the output device 504.
  • a memory 501, a processor 502, an input device 503, and an output device 504 are connected by a bus 505 in FIG. 5, and the bus 505 is indicated by a thick line in FIG. 5.
  • the description is schematic and is not limited.
  • the bus 505 may be divided into an address bus, a data bus, a control bus, and the like. For ease of representation, only a thick line is used in FIG. 5, but it does not mean that there is only one bus or one type of bus.
  • the multi-attribute convolutional neural network model is used to simultaneously analyze each human attribute in the human region image. The recognition is performed, thereby realizing the function of simultaneously identifying multiple human attributes, and effectively improving the recognition efficiency of multiple human attributes.
  • non-volatile computer-readable storage medium provides a non-volatile computer-readable storage medium, where the non-volatile computer-readable storage medium stores computer-executable instructions, and the executable program is executed by a processor to implement the steps of the human body attribute recognition method described above.
  • the executable program can be built into the human attribute recognition device.
  • the human attribute recognition device can implement the steps of the human attribute recognition method by executing the built-in executable program.
  • the executable program can also be used as a The application software is downloaded and installed on the human attribute recognition device, so that the human attribute recognition device can implement the steps of the above-mentioned human attribute recognition method by downloading and installing the executable program.
  • the human body attribute recognition method provided in the example of the present application can also be implemented as a program product.
  • the program product includes a program code.
  • the program code is used to enable the human body attribute recognition device to execute. Steps of the above-mentioned human attribute recognition method.
  • the program products provided in the examples of this application may adopt any combination of one or more readable media, where the readable media may be a readable signal medium or a readable storage medium, and the readable storage medium may be A system, apparatus, or device that is not limited to being electrical, magnetic, optical, electromagnetic, infrared, or semiconductor, or any combination thereof. More specifically, non-exhaustive lists of readable storage media include: Electrical connection of one or more wires, portable disk, hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only Memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination of the foregoing.
  • RAM random access memory
  • ROM read-only memory
  • EPROM or flash memory erasable programmable read-only memory
  • CD-ROM portable compact disk read-only Memory
  • the program product provided in the examples of the present application may adopt a portable compact disk read-only memory (CD-ROM) and include program code, and may also be run on a computing device.
  • CD-ROM portable compact disk read-only memory
  • the program products provided in the examples of the present application are not limited thereto.
  • the readable storage medium may be any tangible medium containing or storing a program, and the program may be used by or in combination with an instruction execution system, apparatus, or device. .
  • a readable signal medium may include a data signal that is carried in baseband or propagated as part of a carrier wave, and which carries readable program code. Such a propagated data signal may take many forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the foregoing.
  • the readable signal medium may also be any readable medium other than a readable storage medium, and the readable medium may send, propagate, or transmit a program for use by or in combination with an instruction execution system, apparatus, or device.
  • the program code contained on the readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wired, optical fiber cable, etc., or any suitable combination of the foregoing.
  • the program code for performing the operations of the present application may be written in any combination of one or more programming languages, which includes an object-oriented programming language, such as Java, C ++, etc., and also includes conventional procedural A programming language, such as "C" or a similar programming language.
  • the program code can be executed entirely on the user computing device, partly on the user device, as an independent software package, partly on the user computing device, partly on the remote computing device, or entirely on the remote computing device or server On.
  • the remote computing device may be connected to the user computing device through any kind of network, such as a user computing device through a local area network (LAN) or wide area network (WAN); or, it may be connected to an external computing device ( (E.g. using an Internet service provider to connect via the Internet).
  • LAN local area network
  • WAN wide area network
  • this application may be provided as methods, systems, or computer program products. Therefore, this application may take the form of an entirely hardware example, an entirely software example, or an example combining software and hardware aspects. Moreover, this application may take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
  • computer-usable storage media including, but not limited to, disk storage, CD-ROM, optical storage, etc.
  • These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to work in a specific manner such that the instructions stored in the computer-readable memory produce a manufactured article including an instruction device, the instructions
  • the device implements the functions specified in one or more flowcharts and / or one or more blocks of the block diagram.
  • These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of steps can be performed on the computer or other programmable device to produce a computer-implemented process, which can be executed on the computer or other programmable device.
  • the instructions provide steps for implementing the functions specified in one or more flowcharts and / or one or more blocks of the block diagrams.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Multimedia (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Mathematical Physics (AREA)
  • Biophysics (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Human Computer Interaction (AREA)
  • Databases & Information Systems (AREA)
  • Medical Informatics (AREA)
  • Business, Economics & Management (AREA)
  • Accounting & Taxation (AREA)
  • Biodiversity & Conservation Biology (AREA)
  • Image Analysis (AREA)
  • Finance (AREA)

Abstract

一种人体属性识别方法、装置、设备及介质，应用于图像识别技术领域。该方法由计算设备执行，包括：确定监控图像中的人体区域图像（201）；将人体区域图像输入多属性卷积神经网络模型，得到人体区域图像中的各个人体属性对应于预先定义的每一个属性值的概率（202）；其中，多属性卷积神经网络模型是利用多属性卷积神经网络对预先获得的训练图像进行多属性识别训练得到的；基于人体区域图像中的各个人体属性对应于预先定义的每一个属性值的概率，确定人体区域图像中的各个人体属性的属性值（203）。

Description

人体属性识别方法、装置、设备及介质
本申请要求于2018年05月30日提交中国专利局、申请号为201810541546.6、发明名称为“一种人体属性识别方法、装置、设备及介质”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及图像识别技术领域，尤其涉及一种人体属性识别方法、装置、设备及介质。
背景
人体属性识别是对人物的性别、年龄、衣物的类型和颜色等的识别，被广泛应用于危险行为预警、交通违章监控、工业安防和自动售货机、自动柜员机(Automatic Teller Machine，ATM)、商场和车站等公共场所的目标人物锁定等领域。
虽然，目前的人体属性识别技术已经取得了很大的成就，但是，大多数的人体属性识别技术都是针对单一的人体属性进行识别，无法同时识别多个人体属性。
技术内容
本申请实例提供了一种人体属性识别方法、装置、设备及介质。
本申请实例提供的具体技术方案如下：
本申请实例提供了一种人体属性识别方法，由计算设备执行，包括：
确定图像中的人体区域图像；
将人体区域图像输入多属性卷积神经网络模型，得到人体区域图像中的各个人体属性对应于预先定义的每一个属性值的概率；其中，多属性卷积神经网络模型是利用多属性卷积神经网络对预先获得的训练图像进行多属性识别训练得到的；
基于人体区域图像中的各个人体属性对应于预先定义的每一个属性值的概率，确定人体区域图像中的各个人体属性的属性值。
本申请实例还提供了一种人体属性识别装置，包括：
确定单元，用于确定图像中的人体区域图像；
获取单元，用于将人体区域图像输入多属性卷积神经网络模型，得到人体区域图像中的各个人体属性对应于预先定义的每一个属性值的概率；其中，多属性卷积神经网络模型是利用多属性卷积神经网络对预先获得的训练图像进行多属性识别训练得到的；
识别单元，用于基于人体区域图像中的各个人体属性对应于预先定义的每一个属性值的概率，确定人体区域图像中的各个人体属性的属性值。
本申请实例还提供了一种人体属性识别设备，包括存储器和处理器，所述存储器中存储有计算机可读指令，所述指令可以使所述处理器：
确定图像中的人体区域图像；
将所述人体区域图像输入多属性卷积神经网络模型，得到所述人体区域图像中的各个人体属性对应于预先定义的每一个属性值的概率；其中，所述多属性卷积神经网络模型是利用多属性卷积神经网络对预先获得的训练图像进行多属性识别训练得到的；
基于所述人体区域图像中的各个人体属性对应于预先定义的每一个属性值的概率，确定所述人体区域图像中的各个人体属性的属性值。
本申请实例还提供了一种存储介质，存有处理器可执行指令，所述指令由一个或一个以上处理器执行时，实现上述人体属性识别方法的步骤。
附图简要说明
图1A为本申请实例提供的多属性卷积神经网络模型的训练方法的流程示意图；
图1B为采用本申请实例提供的多属性卷积神经网络模型识别人体属性时 的输入输出与采用多属性多模型方式识别人体属性时的输入输出的对比图；
图2为本申请实例提供的人体属性识别方法的流程示意图；
图3为以“交通违章监控”为具体应用场景时，本申请实例提供的人体属性识别方法的流程示意图；
图4为本申请实例提供的人体属性识别装置的功能结构示意图；
图5为本申请实例提供的人体属性识别设备的硬件结构示意图。
实施方式
在实际应用中，通常是根据多个人体属性来锁定目标人物，可见，精确地识别出多个人体属性的属性值，对于目标人物的锁定来说尤为重要。目前，主要采用多属性多模型的方式来识别各个人体属性，具体的，针对每一个人体属性分别建立属性识别模型，即一个人体属性对应一个属性识别模型，这样，在视频监控过程中，可以实时采集监控图像并检测监控图像中的人体区域图像，从而通过建立的各个属性识别模型，分别对人体区域图像中的各个人体属性进行识别，进而达到多属性识别的目的。然而，在具体实践过程中，本申请的发明人发现，虽然这种多属性多模型的方式在一定程度上能够实现多属性识别，但是，需要针对每一个人体属性分别建立一个属性识别模型，预处理操作较为繁琐，而且，在多属性识别过程中，需要利用预先建立的各个属性识别模型分别对相应的人体属性进行识别，无法实现对多个人体属性的同时识别，识别效率较低。
为此，本申请的发明人想到，利用多属性卷积神经网络对预先获得的训练图像进行多属性识别训练，以此来建立一个多属性卷积神经网络模型，这样，在视频监控过程中，就可以实时采集监控图像，并在检测到监控图像中的人体区域图像时，通过多属性卷积神经网络模型同时对人体区域图像中的各个人体属性进行识别，从而实现了多个人体属性同时识别的功能，有效地提高了多个人体属性的识别效率。
值得说的是，本申请实例提供的人体属性识别方法可以应用于多种视频监 控场景，比如，危险行为监控、交通违章监控、工业安防监控和自动售货机、ATM、商场和车站等公共场所的监控等，需要注意的是，上述提及的应用场景仅是为了便于理解本申请的精神和原理而示出，本申请实例在此方面不受任何限制。相反，本申请实例可以应用于适用的任何场景。例如，可以应用于商场、门店等线下零售场景，识别入店及路过客群的属性信息，收集消费者画像，从而辅助精准营销、个性化推荐、门店选址、流行趋势分析等应用。又例如，可以应用于楼宇、户外等广告的智能化升级，采集人体信息，分析人群属性，定向投放广告物料，从而提升用户体验和商业效率。
在简单的介绍了本申请实例提供的人体属性识别方法和该人体属性识别方法的应用场景之后，接下来，结合本申请实例中的附图，对本申请实例提供的人体属性识别方法进行清楚、完整地描述，显然，所描述的实例仅仅是本申请一部分实例，并不是全部的实例。基于本申请中的实例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实例，都属于本申请保护的范围。
为了便于理解本申请，首先对本申请实例中涉及的部分技术用语进行说明。
1、人体属性，为能够同时被计算机和人感知的人体视觉特征，比如，性别、年龄、上衣纹理、上衣颜色、袖子类型、下衣纹理、下衣颜色、下衣长短、下衣类型等。
2、人体属性的属性值，为赋予人体属性的值，比如，人体属性—性别的属性值为男、女，人体属性—年龄的属性值为婴儿、儿童、青年、中年、老年等，人体属性—上衣纹理的属性值为纯色、横条纹、竖条纹、格子、大色块等，人体属性—上衣颜色的属性值为红、黄、蓝、黑、混色等。
3、属性识别模型，为能够同时识别出多个人体属性的属性值的模型，本申请实例以属性识别模型是多属性卷积神经网络模型为例进行说明。
4、多属性卷积神经网络模型，为利用多属性卷积神经网络对预先获得的训练图像进行多属性识别训练得到的、能够同时识别出多个人体属性的属性识别 模型。
5、多属性卷积神经网络，为深度学习的前馈神经网络，包括数据输入层、卷积计算层、激励层和全连接层，或者，包括数据输入层、卷积计算层、激励层和全局平均池化层。
其次，基于本申请实例提供的人体属性识别方法采用的是多属性卷积神经网络模型，而该多属性卷积神经网络模型是利用多属性卷积神经网络对预先获得的训练图像进行多属性识别训练得到的，所以，接下来，对本申请示例性实施方式的多属性卷积神经网络模型的训练方法进行说明。参阅图1A所示，多属性卷积神经网络模型的训练方法的流程如下：
步骤101：采集训练图像。具体的，可以直接从存储的监控视频中直接截取监控图像作为训练图像。
步骤102：定义各个人体属性以及各个人体属性对应的属性值。
例如：将人体属性定义为性别、年龄、人脸朝向、上衣纹理、上衣颜色、袖子类型、包型、下衣颜色、下衣长短、下衣类型。
将人体属性—性别的属性值定义为男、女，人体属性—年龄的属性值定义为儿童、青年、中年、老年；将人体属性—人脸朝向的属性值定义为正向、背向；将人体属性—上衣纹理的属性值定义为纯色、格子、大色块；将人体属性—上衣颜色的属性值定义为红、黄、绿、蓝、黑、灰白；将人体属性—袖子类型的属性值定义为长袖、短袖；将人体属性—包型的属性值定义为双肩包、斜挎包、手拎包、无包；将人体属性—下衣颜色的属性值定义为纯色、格子、大色块；将人体属性—上衣颜色的属性值定义为黑、灰白、彩色；将人体属性—下衣长短的属性值定义为长、短；将人体属性—上衣颜色的属性值定义为黑、灰白、彩色；将人体属性—下衣类型的属性值定义为裤子、裙子，等等，具体如表1所示：
表1.
Figure PCTCN2019082977-appb-000001
Figure PCTCN2019082977-appb-000002
步骤103：根据各个人体属性的属性值，确定人体属性向量。
例如：根据上述表1中的各个人体属性的属性值，可以确定出一个1×30的人体属性向量，即人体属性向量＝(性别男，性别女，儿童，青年，中年，老年，人脸正向，人脸背向，上衣纹理纯色，上衣纹理格子，上衣纹理大色块，上衣红色，上衣黄色，上衣绿色，上衣颜色，上衣黑色，上衣灰白色，上衣长袖，上衣短袖，双肩包，斜挎包，手拎包，无包，下衣黑色，下衣灰白色，下衣彩色，下衣长，下衣短，下衣裤子，下衣裙子)。
在一些实例中，还可以对上述每一个属性的属性值进行二分并赋值，例如，是否是男，是男则赋1，不是男则赋0，无法识别，则赋99；同理，是否是绿，是绿则赋1，不是绿则赋0，无法识别，则赋99；依次类推，可以得到其他属性的赋值，进而得到一个1×30的人体属性向量。例如，对训练图像或测试图像进行上述处理，可以分别得到如下向量：
Image_1 1 0 1…0 1
Image_2 0 1 99...1 0
...
Image_i 1 0 99...0 1
...
Image_n 1 0 0...0 1
其中，Imgae_i表示图片i的相对路径，0和1表示对应二分属性值，99 表示无法识别。
步骤104：将训练图像和人体属性向量输入多属性卷积神经网络的数据输入层。
步骤105：多属性卷积神经网络的数据输入层对训练图像进行去均值、归一化、主成分分析(Principal Component Analysis，PCV)和白化等预处理。
步骤106：多属性卷积神经网络的卷积计算层对数据输入层输出的训练图像进行特征提取，得到训练图像对应的各个特征矩阵。
步骤107：多属性卷积神经网络的激励层对卷积计算层输出的各个特征矩阵进行非线性映射处理，从而将各个特征矩阵中的特征值映射到一定范围内。具体的，在进行非线性映射时，可以采用但不限于Sigmoid函数、Tanh函数(双曲正切函数)、ReLU(Rectified Linear Unit，线性整流单元)函数等作为激励函数。在一些实例中，可以采用ReLU函数作为激励函数，对卷积计算层输出的各个特征矩阵进行非线性映射处理。
步骤108：多属性卷积神经网络的全连接层或者全局平均池化层根据激励层输出的各个特征矩阵，得到训练图像中的各个人体属性对应于预先定义的每一个属性值的概率。
下面仅以通过多属性卷积神经网络的全连接层确定训练图像中的各个人体属性对应于预先定义的每一个属性值的概率为例进行说明，具体的，可以采用但不限于以下方式：设置一个偏置量以及为各个人体属性分别设置一个权重矩阵，并根据激励层输出的各个特征矩阵以及预先设置的偏置量和各个权重矩阵，确定训练图像中的各个人体属性对应于预先定义的每一个属性值的概率。
步骤109：基于训练图像中的各个人体属性对应于预先定义的每一个属性值的概率，确定训练图像中的各个人体属性的预测属性值，并根据训练图像中的各个人体属性的预测属性值，组成训练图像的预测属性值向量。
具体的，在基于训练图像中的各个人体属性对应于预先定义的每一个属性值的概率，确定训练图像中的各个人体属性的预测属性值时，可以采用但不限 于以下方式：针对训练图像中的每一个人体属性，从该人体属性对应于预先定义的每一个属性值的概率中，选取对应的概率最大的属性值作为该人体属性的预测属性值。
步骤110：确定全连接层或者全局平均池化层输出的预测属性值向量中的各个人体属性的预测属性值与预先定义的各个人体属性的真实属性值之间的差异度。
在一些实例中，在确定各个人体属性的预测属性值与真实属性值之间的差异度时，可以采用但不限于如下述公式(1)所示的交叉熵损失函数来确定：
Figure PCTCN2019082977-appb-000003
其中，在上述公式(1)中，L表示交叉熵损失函数的值，即差异度，n表示训练图像的个数，x表示第x个训练图像，m表示预先定义的人体属性的个数，y i和a i分别表示第i个人体属性的真实属性值和预测属性值。
步骤111：根据确定出的差异度，调整训练过程中使用的各层网络参数。其中，各层网络参数包括但不限于：各卷积计算层的内核参数和初始偏置矩阵、各激励层的参数、各全连接层或者全局平均池化层的参数等。
步骤112：利用调整网络参数后的多属性卷积神经网络对后续的训练图片进行多属性识别训练，如此循环往复，直至各个人体属性的预测属性值与真实属性值之间的差异度不大于预设阈值为止。
此时，多属性卷积神经网络对应的各层网络参数均为最优值，至此，多属性卷积神经网络模型的训练过程结束，并得到了各层网络参数均为最优值的多属性卷积神经网络模型。
值得说的是，在多属性卷积神经网络模型的整个训练过程中，没有使用池化层对特征数据进行压缩，尽可能地避免了由于使用池化层对特征数据进行压缩导致整个多属性卷积神经网络模型的表达能力较差的问题，有效地提高了人体属性识别的精确度。
此外，当增加新人体属性和/或为原人体属性增加新属性值时，可以按照设 定比例选取新训练图像和原训练图像为训练图像，并利用多属性卷积神经网络对这些训练图像进行多属性识别训练，得到增加有新人体属性和/或新属性值的多属性卷积神经网络模型。这样，采用在新训练图像中增加原训练图像的方法，来增加新人体属性和/或新属性值的训练图像的数量，可以有效地避免由于新训练图像的数量较少导致训练结果不准确的问题，间接地提高了多属性卷积神经网络模型识别人体属性的精确度。
参阅图1B所示，为采用本申请实例提供的多属性卷积神经网络模型识别人体属性时的输入输出与采用多属性多模型方式识别人体属性时的输入输出的对比图，相对于多属性多模型的人体属性识别方式，本申请实例提供的多属性卷积神经网络模型能够同时识别出多个人体属性，不仅解决了多属性多模型的人体属性识别方式存在的无法同时识别多个人体属性的问题，还尽可能地提高了多个人体属性的识别效率。
在建立了多属性卷积神经网络模型之后，即可利用该多属性卷积神经网络模型同时对多个人体属性进行识别，具体的，参阅图2所示，本申请示例性实施方式的人体属性识别方法的流程如下，该方法可以由人体属性识别设备执行：
步骤201：确定监控图像中的人体区域图像。
步骤202：将人体区域图像输入多属性卷积神经网络模型，得到人体区域图像中的各个人体属性对应于预先定义的每一个属性值的概率。
其中，多属性卷积神经网络模型是利用多属性卷积神经网络对预先获得的训练图像进行多属性识别训练得到的，具体的，多属性卷积神经网络模型的训练方法与上述描述的训练方法相同，在此不再赘述。
步骤203：基于人体区域图像中的各个人体属性对应于预先定义的每一个属性值的概率，确定人体区域图像中的各个人体属性的属性值。
具体的，可以针对人体区域图像中的每一个人体属性，从该人体属性对应于预先定义的每一个属性值的概率中，选取对应的概率最大的属性值作为该人体属性的属性值。
通过上述图2所示的方案，可以在检测到监控图像中的人体区域图像时，通过多属性卷积神经网络模型同时对人体区域图像中的各个人体属性进行识别，从而实现了多个人体属性同时识别的功能，有效地提高了多个人体属性的识别效率。
下面以“交通违章监控”为具体应用场景，对本申请示例性实施方式的人体属性识别方法进行说明，具体的，参阅图3所示，本申请示例性实施方式的人体属性识别方法的流程如下：
步骤301：从摄像头采集到的监控视频中截取监控图像。
步骤302：确定监控图像中的人体区域图像。
步骤303：将人体区域图像输入多属性卷积神经网络模型，得到人体区域图像中的各个人体属性对应于预先定义的每一个属性值的概率，并基于人体区域图像中的各个人体属性对应于预先定义的每一个属性值的概率，确定人体区域图像中的各个人体属性的属性值。
针对人体区域图像中的每一个人体属性，从该人体属性对应于预先定义的每一个属性值的概率中，选取对应的概率最大的属性值作为该人体属性的属性值，从而获得人体区域图像中的各个人体属性的属性值。
步骤304：根据人体区域图像中的各个人体属性的属性值，锁定监控图像中的目标人物，并在监控图像上标注该目标人物的各个人体属性的属性值和交通违章行为等信息。
在一些实例中，可以在确定该目标人物确实有交通违章行为时，在违章处理下拉框中选取确认违章，并根据识别出的该目标人物的各个人体属性的属性值和交通违章行为，生成交通违章记录，以及在监控图像上标注该目标人物的各个人体属性的属性值和交通违章行为等信息，当然，在针对该目标人物的交通违章行为调派相关人出警后，还可以在违章处理下拉框中选取已出警，从而结束此次交通违章行为的处理；而在确定该目标人物确实没有交通违章行为时，可以在违章处理下拉框中选取忽略违章，并结束此次交通违章行为的处理，此 时，不会生成交通违章记录。
基于上述实例，本申请实例提供了一种人体属性识别装置，参阅图4所示，该人体属性识别装置至少包括：
确定单元401，用于确定监控图像中的人体区域图像；
获取单元402，用于将人体区域图像输入多属性卷积神经网络模型，得到人体区域图像中的各个人体属性对应于预先定义的每一个属性值的概率；其中，多属性卷积神经网络模型是利用多属性卷积神经网络对预先获得的训练图像进行多属性识别训练得到的；
识别单元403，用于基于人体区域图像中的各个人体属性对应于预先定义的每一个属性值的概率，确定人体区域图像中的各个人体属性的属性值。
在一些实例中，该人体属性识别装置还包括：
训练单元404，用于根据预先定义的各个人体属性的属性值，并依次通过多属性卷积神经网络的数据输入层、卷积计算层、激励层和全连接层，对预先获得的训练图像进行多属性识别训练，得到多属性卷积神经网络模型；或者，根据预先定义的各个人体属性的属性值，并依次通过多属性卷积神经网络的数据输入层、卷积计算层、激励层和全局平均池化层，对预先获得的训练图像进行多属性识别训练，得到多属性卷积神经网络模型。
在一些实例中，该人体属性识别装置还包括：
调节单元405，用于确定多属性卷积神经网络模型输出的训练图像的各个人体属性的预测属性值与预先定义的各个人体属性的真实属性值之间的差异度，并根据差异度，其中，所述预测属性值是基于训练图像中的各个人体属性对应于预先定义的每一个属性值的概率确定的；调整多属性卷积神经网络模型对应的各层网络参数。
在一些实例中，在确定多属性卷积神经网络模型输出的训练图像的各个人体属性的预测属性值与预先定义的各个人体属性的真实属性值之间的差异度时，调节单元405具体用于：
利用交叉熵损失函数，确定各个人体属性的预测属性值与预先定义的各个人体属性的真实属性值之间的差异度。
在一些实例中，训练单元404还用于：
当增加新人体属性和/或为原人体属性增加新属性值时，按照设定比例选取新训练图像和原训练图像为训练图像；
利用多属性卷积神经网络对训练图像进行多属性识别训练，得到增加有新人体属性和/或新属性值的多属性卷积神经网络模型。
在一些实例中，在基于人体区域图像中的各个人体属性对应于预先定义的每一个属性值的概率，确定人体区域图像中的各个人体属性的属性值时，识别单元403具体用于：
针对人体区域图像中的每一个人体属性，从人体属性对应于预先定义的每一个属性值的概率中，选取对应的概率最大的属性值作为人体属性的属性值。
需要说明的是，由于上述人体属性识别装置解决技术问题的原理与上述人体属性识别方法相似，因此，上述人体属性识别装置的实施可以参见上述人体属性识别方法的实施，重复之处不再赘述。
综上所述，本申请实例所示的方案，所述人体属性识别装置可以在检测到监控图像中的人体区域图像时，通过多属性卷积神经网络模型同时对人体区域图像中的各个人体属性进行识别，从而实现了多个人体属性同时识别的功能，有效地提高了多个人体属性的识别效率。
此外，本申请实例还提供了一种人体属性识别设备，参阅图5所示，该人体属性识别设备至少包括：存储器501、处理器502和存储在存储器502上的计算机程序，处理器502执行该计算机程序时实现上述人体属性识别方法的步骤。
在一些实例中，该人体属性识别设备还可以包括输入装置503和输出装置504等。输入装置503可以包括触控笔、键盘、鼠标、触摸屏等；输出装置504可以包括显示设备，如液晶显示器(Liquid Crystal Display，LCD)、阴极射线 管(Cathode Ray Tube，CRT)，触摸屏等。
本申请实例中不限定存储器501，处理器502、输入装置503和输出装置504之间的具体连接介质。本申请实例在图5中以存储器501，处理器502、输入装置503和输出装置504之间通过总线505连接，总线505在图5中以粗线表示，其它部件之间的连接方式，仅是进行示意性说明，并不引以为限。其中，总线505可以分为地址总线、数据总线、控制总线等。为便于表示，图5中仅用一条粗线表示，但并不表示仅有一根总线或一种类型的总线。
综上所述，本申请实例所示的方案，所述人体属性识别设备可以在检测到监控图像中的人体区域图像时，通过多属性卷积神经网络模型同时对人体区域图像中的各个人体属性进行识别，从而实现了多个人体属性同时识别的功能，有效地提高了多个人体属性的识别效率。
接下来，对本申请示例性实施方式的非易失性计算机可读存储介质进行介绍。本申请实例提供了一种非易失性计算机可读存储介质，该非易失性计算机可读存储介质存储有计算机可执行指令，该可执行程序被处理器执行实现上述人体属性识别方法的步骤。具体地，该可执行程序可以内置在人体属性识别设备中，这样，人体属性识别设备就可以通过执行内置的可执行程序实现上述人体属性识别方法的步骤，当然，该可执行程序也可以作为一个应用软件下载并安装到人体属性识别设备上，这样，人体属性识别设备就可以通过下载并安装的可执行程序实现上述人体属性识别方法的步骤。
此外，本申请实例提供的人体属性识别方法还可以实现为一种程序产品，该程序产品包括程序代码，当该程序产品可以在移动终端上运行时，该程序代码用于使人体属性识别设备执行上述人体属性识别方法的步骤。
在一些实例中，本申请实例提供的程序产品可以采用一个或多个可读介质的任意组合，其中，可读介质可以是可读信号介质或者可读存储介质，而可读存储介质可以是但不限于是电、磁、光、电磁、红外线、或半导体的系统、装置或器件，或者任意以上的组合，具体地，可读存储介质的更具体的例子(非 穷举的列表)包括：具有一个或多个导线的电连接、便携式盘、硬盘、随机存取存储器(RAM)、只读存储器(ROM)、可擦式可编程只读存储器(EPROM或闪存)、光纤、便携式紧凑盘只读存储器(CD-ROM)、光存储器件、磁存储器件、或者上述的任意合适的组合。
在一些实例中，本申请实例提供的程序产品可以采用便携式紧凑盘只读存储器(CD-ROM)并包括程序代码，还可以在计算设备上运行。然而，本申请实例提供的程序产品不限于此，在本申请实例中，可读存储介质可以是任何包含或存储程序的有形介质，该程序可以被指令执行系统、装置或者器件使用或者与其结合使用。
可读信号介质可以包括在基带中或者作为载波一部分传播的数据信号，其中承载了可读程序代码。这种传播的数据信号可以采用多种形式，包括但不限于电磁信号、光信号或上述的任意合适的组合。可读信号介质还可以是可读存储介质以外的任何可读介质，该可读介质可以发送、传播或者传输用于由指令执行系统、装置或者器件使用或者与其结合使用的程序。
可读介质上包含的程序代码可以用任何适当的介质传输，包括但不限于无线、有线、光缆等等，或者上述的任意合适的组合。
可以以一种或多种程序设计语言的任意组合来编写用于执行本申请操作的程序代码，所述程序设计语言包括面向对象的程序设计语言，诸如Java、C++等，还包括常规的过程式程序设计语言，诸如“C”语言或类似的程序设计语言。程序代码可以完全地在用户计算设备上执行、部分地在用户设备上执行、作为一个独立的软件包执行、部分在用户计算设备上部分在远程计算设备上执行、或者完全在远程计算设备或服务器上执行。在涉及远程计算设备的情形中，远程计算设备可以通过任意种类的网络连接到用户计算设备，诸如通过局域网(LAN)或广域网(WAN)连接到用户计算设备；或者，可以连接到外部计算设备(例如利用因特网服务提供商来通过因特网连接)。
应当注意，尽管在上文详细描述中提及了装置的若干单元或子单元，但是 这种划分仅仅是示例性的并非强制性的。实际上，根据本申请的实施方式，上文描述的两个或更多单元的特征和功能可以在一个单元中具体化。反之，上文描述的一个单元的特征和功能可以进一步划分为由多个单元来具体化。
此外，尽管在附图中以特定顺序描述了本申请方法的操作，但是，这并非要求或者暗示必须按照该特定顺序来执行这些操作，或是必须执行全部所示的操作才能实现期望的结果。附加地或备选地，可以省略某些步骤，将多个步骤合并为一个步骤执行，和/或将一个步骤分解为多个步骤执行。
本领域内的技术人员应明白，本申请的实例可提供为方法、系统、或计算机程序产品。因此，本申请可采用完全硬件实例、完全软件实例、或结合软件和硬件方面的实例的形式。而且，本申请可采用在一个或多个其中包含有计算机可用程序代码的计算机可用存储介质(包括但不限于磁盘存储器、CD-ROM、光学存储器等)上实施的计算机程序产品的形式。
本申请是参照根据本申请实例的方法、设备(系统)、和计算机程序产品的流程图和/或方框图来描述的。应理解可由计算机程序指令实现流程图和/或方框图中的每一流程和/或方框、以及流程图和/或方框图中的流程和/或方框的结合。可提供这些计算机程序指令到通用计算机、专用计算机、嵌入式处理机或其他可编程数据处理设备的处理器以产生一个机器，使得通过计算机或其他可编程数据处理设备的处理器执行的指令产生用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的装置。
这些计算机程序指令也可存储在能引导计算机或其他可编程数据处理设备以特定方式工作的计算机可读存储器中，使得存储在该计算机可读存储器中的指令产生包括指令装置的制造品，该指令装置实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能。
这些计算机程序指令也可装载到计算机或其他可编程数据处理设备上，使得在计算机或其他可编程设备上执行一系列操作步骤以产生计算机实现的处理，从而在计算机或其他可编程设备上执行的指令提供用于实现在流程图一个流程 或多个流程和/或方框图一个方框或多个方框中指定的功能的步骤。
尽管已描述了本申请的优选实例，但本领域内的技术人员一旦得知了基本创造性概念，则可对这些实例作出另外的变更和修改。所以，所附权利要求意欲解释为包括优选实例以及落入本申请范围的所有变更和修改。
显然，本领域的技术人员可以对本申请实例进行各种改动和变型而不脱离本申请实例的精神和范围。这样，倘若本申请实例的这些修改和变型属于本申请权利要求及其等同技术的范围之内，则本申请也意图包含这些改动和变型在内。

Claims (22)

  1. 一种人体属性识别方法，由人体属性识别设备执行，包括：
    确定图像中的人体区域图像；
    将所述人体区域图像输入多属性卷积神经网络模型，得到所述人体区域图像中的各个人体属性对应于预先定义的每一个属性值的概率；其中，所述多属性卷积神经网络模型是利用多属性卷积神经网络对预先获得的训练图像进行多属性识别训练得到的；
    基于所述人体区域图像中的各个人体属性对应于预先定义的每一个属性值的概率，确定所述人体区域图像中的各个人体属性的属性值。
  2. 如权利要求1所述的人体属性识别方法，其中，所述确定图像中的人体区域图像，包括：
    确定监控图像中的人体区域图像。
  3. 如权利要求2所述的人体属性识别方法，其中，利用多属性卷积神经网络对预先获得的训练图像进行多属性识别训练得到所述多属性卷积神经网络模型，包括：
    根据预先定义的各个人体属性的属性值，并依次通过所述多属性卷积神经网络的数据输入层、卷积计算层、激励层和全连接层，对预先获得的训练图像进行多属性识别训练，得到所述多属性卷积神经网络模型；或者，
    根据预先定义的各个人体属性的属性值，并依次通过所述多属性卷积神经网络的数据输入层、卷积计算层、激励层和全局平均池化层，对预先获得的训练图像进行多属性识别训练，得到所述多属性卷积神经网络模型。
  4. 如权利要求3所述的人体属性识别方法，还包括：
    确定所述多属性卷积神经网络模型输出的所述训练图像的各个人体属性的预测属性值与预先定义的各个人体属性的真实属性值之间的差异度；其中，所述预测属性值是基于训练图像中的各个人体属性对应于预先定义的每一个属性值的概率确定的；
    根据所述差异度，调整所述多属性卷积神经网络模型对应的各层网络参数。
  5. 如权利要求4所述的人体属性识别方法，其中，确定所述多属性卷积神经网络模型输出的所述训练图像的各个人体属性的预测属性值与预先定义的各个人体属性的真实属性值之间的差异度，包括：
    利用交叉熵损失函数，确定各个人体属性的预测属性值与预先定义的各个人体属性的真实属性值之间的差异度。
  6. 如权利要求3-5任一项所述的人体属性识别方法，还包括：
    当增加新人体属性和/或为原人体属性增加新属性值时，按照设定比例选取新训练图像和原训练图像为训练图像；
    利用所述多属性卷积神经网络对所述训练图像进行多属性识别训练，得到增加有新人体属性和/或新属性值的多属性卷积神经网络模型。
  7. 如权利要求1所述的人体属性识别方法，其中，基于所述人体区域图像中的各个人体属性对应于预先定义的每一个属性值的概率，确定所述人体区域图像中的各个人体属性的属性值，包括：
    针对所述人体区域图像中的每一个人体属性，从所述人体属性对应于预先定义的每一个属性值的概率中，选取对应的概率最大的属性值作为所述人体属性的属性值。
  8. 一种人体属性识别装置，包括：
    确定单元，用于确定图像中的人体区域图像；
    获取单元，用于将所述人体区域图像输入多属性卷积神经网络模型，得到所述人体区域图像中的各个人体属性对应于预先定义的每一个属性值的概率；其中，所述多属性卷积神经网络模型是利用多属性卷积神经网络对预先获得的训练图像进行多属性识别训练得到的；
    识别单元，用于基于所述人体区域图像中的各个人体属性对应于预先定义的每一个属性值的概率，确定所述人体区域图像中的各个人体属性的属性值。
  9. 如权利要求8所述的人体属性识别装置，其中，所述确定单元，具体用 于：
    确定监控图像中的人体区域图像。
  10. 如权利要求9所述的人体属性识别装置，还包括：
    训练单元，用于根据预先定义的各个人体属性的属性值，并依次通过所述多属性卷积神经网络的数据输入层、卷积计算层、激励层和全连接层，对预先获得的训练图像进行多属性识别训练，得到所述多属性卷积神经网络模型；或者，根据预先定义的各个人体属性的属性值，并依次通过所述多属性卷积神经网络的数据输入层、卷积计算层、激励层和全局平均池化层，对预先获得的训练图像进行多属性识别训练，得到所述多属性卷积神经网络模型。
  11. 如权利要求9所述的人体属性识别装置，还包括：
    调节单元，用于确定所述多属性卷积神经网络模型输出的所述训练图像的各个人体属性的预测属性值与预先定义的各个人体属性的真实属性值之间的差异度，其中，所述预测属性值是基于训练图像中的各个人体属性对应于预先定义的每一个属性值的概率确定的；并根据所述差异度，调整所述多属性卷积神经网络模型对应的各层网络参数。
  12. 如权利要求11所述的人体属性识别装置，其中，在确定所述多属性卷积神经网络模型输出的所述训练图像的各个人体属性的预测属性值与预先定义的各个人体属性的真实属性值之间的差异度时，所述调节单元具体用于：
    利用交叉熵损失函数，确定各个人体属性的预测属性值与预先定义的各个人体属性的真实属性值之间的差异度。
  13. 如权利要求10-12任一项所述的人体属性识别装置，其中，所述训练单元还用于：
    当增加新人体属性和/或为原人体属性增加新属性值时，按照设定比例选取新训练图像和原训练图像为训练图像；
    利用所述多属性卷积神经网络对所述训练图像进行多属性识别训练，得到增加有新人体属性和/或新属性值的多属性卷积神经网络模型。
  14. 如权利要求8所述的人体属性识别装置，其中，在基于所述人体区域图像中的各个人体属性对应于预先定义的每一个属性值的概率，确定所述人体区域图像中的各个人体属性的属性值时，所述识别单元具体用于：
    针对所述人体区域图像中的每一个人体属性，从所述人体属性对应于预先定义的每一个属性值的概率中，选取对应的概率最大的属性值作为所述人体属性的属性值。
  15. 一种人体属性识别设备，包括存储器和处理器，所述存储器中存储有计算机可读指令，所述指令可以使所述处理器：
    确定图像中的人体区域图像；
    将所述人体区域图像输入多属性卷积神经网络模型，得到所述人体区域图像中的各个人体属性对应于预先定义的每一个属性值的概率；其中，所述多属性卷积神经网络模型是利用多属性卷积神经网络对预先获得的训练图像进行多属性识别训练得到的；
    基于所述人体区域图像中的各个人体属性对应于预先定义的每一个属性值的概率，确定所述人体区域图像中的各个人体属性的属性值。
  16. 如权利要求15所述的人体属性识别设备，其中，所述计算机可读指令可以使所述处理器：
    确定监控图像中的人体区域图像。
  17. 如权利要求16所述的人体属性识别设备，其中，所述计算机可读指令可以使所述处理器：
    根据预先定义的各个人体属性的属性值，并依次通过所述多属性卷积神经网络的数据输入层、卷积计算层、激励层和全连接层，对预先获得的训练图像进行多属性识别训练，得到所述多属性卷积神经网络模型；或者，
    根据预先定义的各个人体属性的属性值，并依次通过所述多属性卷积神经网络的数据输入层、卷积计算层、激励层和全局平均池化层，对预先获得的训练图像进行多属性识别训练，得到所述多属性卷积神经网络模型。
  18. 如权利要求17所述的人体属性识别设备，其中，所述计算机可读指令可以使所述处理器：
    确定所述多属性卷积神经网络模型输出的所述训练图像的各个人体属性的预测属性值与预先定义的各个人体属性的真实属性值之间的差异度；其中，所述预测属性值是基于训练图像中的各个人体属性对应于预先定义的每一个属性值的概率确定的；
    根据所述差异度，调整所述多属性卷积神经网络模型对应的各层网络参数。
  19. 如权利要求18所述的人体属性识别设备，其中，所述计算机可读指令可以使所述处理器：
    利用交叉熵损失函数，确定各个人体属性的预测属性值与预先定义的各个人体属性的真实属性值之间的差异度。
  20. 如权利要求17-19任一项所述的人体属性识别设备，其中，所述计算机可读指令可以使所述处理器：
    当增加新人体属性和/或为原人体属性增加新属性值时，按照设定比例选取新训练图像和原训练图像为训练图像；
    利用所述多属性卷积神经网络对所述训练图像进行多属性识别训练，得到增加有新人体属性和/或新属性值的多属性卷积神经网络模型。
  21. 如权利要求15所述的人体属性识别设备，其中，所述计算机可读指令可以使所述处理器：
    针对所述人体区域图像中的每一个人体属性，从所述人体属性对应于预先定义的每一个属性值的概率中，选取对应的概率最大的属性值作为所述人体属性的属性值。
  22. 一种存储介质，存有处理器可执行指令，所述指令由一个或一个以上处理器执行时，实现如权利要求1-7任一项所述的人体属性识别方法。
PCT/CN2019/082977 2018-05-30 2019-04-17 人体属性识别方法、装置、设备及介质 WO2019228089A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US16/938,858 US11275932B2 (en) 2018-05-30 2020-07-24 Human body attribute recognition method, apparatus, and device and medium

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201810541546.6A CN108921022A (zh) 2018-05-30 2018-05-30 一种人体属性识别方法、装置、设备及介质
CN201810541546.6 2018-05-30

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US16/938,858 Continuation US11275932B2 (en) 2018-05-30 2020-07-24 Human body attribute recognition method, apparatus, and device and medium

Publications (1)

Publication Number Publication Date
WO2019228089A1 true WO2019228089A1 (zh) 2019-12-05

Family

ID=64417932

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/082977 WO2019228089A1 (zh) 2018-05-30 2019-04-17 人体属性识别方法、装置、设备及介质

Country Status (3)

Country Link
US (1) US11275932B2 (zh)
CN (1) CN108921022A (zh)
WO (1) WO2019228089A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111414812A (zh) * 2020-03-03 2020-07-14 平安科技（深圳）有限公司 人体属性识别方法、系统、计算机设备及存储介质
CN114529731A (zh) * 2020-10-30 2022-05-24 北京眼神智能科技有限公司 人脸特征点定位及属性分析方法、装置、存储介质及设备

Families Citing this family (16)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108921022A (zh) * 2018-05-30 2018-11-30 腾讯科技（深圳）有限公司 一种人体属性识别方法、装置、设备及介质
CN111291604A (zh) * 2018-12-07 2020-06-16 深圳光启空间技术有限公司 面部属性识别方法、装置、存储介质及处理器
WO2020124390A1 (zh) * 2018-12-18 2020-06-25 华为技术有限公司 一种面部属性的识别方法及电子设备
CN109800678A (zh) * 2018-12-29 2019-05-24 上海依图网络科技有限公司 一种视频中对象的属性确定方法及装置
CN111753596A (zh) * 2019-03-29 2020-10-09 商汤集团有限公司 神经网络的训练方法及装置、电子设备和存储介质
US11151412B2 (en) * 2019-07-01 2021-10-19 Everseen Limited Systems and methods for determining actions performed by objects within images
CN110472611A (zh) * 2019-08-21 2019-11-19 图谱未来(南京)人工智能研究院有限公司 人物属性识别的方法、装置、电子设备及可读存储介质
CN112541374B (zh) * 2019-09-20 2024-04-30 南京行者易智能交通科技有限公司 一种基于深度学习的乘客属性的获取方法、装置及模型训练方法
CN110674756B (zh) * 2019-09-25 2022-07-05 普联技术有限公司 人体属性识别模型训练方法、人体属性识别方法及装置
CN111783619B (zh) * 2020-06-29 2023-08-11 北京百度网讯科技有限公司 人体属性的识别方法、装置、设备及存储介质
CN112633119A (zh) * 2020-12-17 2021-04-09 北京赢识科技有限公司 一种人体属性识别方法、装置、电子设备及介质
CN112906525A (zh) * 2021-02-05 2021-06-04 广州市百果园信息技术有限公司 年龄识别方法、装置及电子设备
US20230057295A1 (en) * 2021-08-23 2023-02-23 Element8 Technology Investment Group Inc. System and method for providing a multi-sided platform for broadband and content delivery networks
CN115019350B (zh) * 2022-08-09 2022-12-06 中科视语(北京)科技有限公司 人体属性识别方法、装置及电子设备
CN115294611B (zh) * 2022-09-29 2023-02-03 深圳爱莫科技有限公司 人体属性识别模型训练方法、识别方法、装置及设备
CN116030500B (zh) * 2023-02-16 2023-07-18 深圳华付技术股份有限公司 一种人员着装规范识别方法及系统

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20130322740A1 (en) * 2012-05-31 2013-12-05 Lihui Chen Method of Automatically Training a Classifier Hierarchy by Dynamic Grouping the Training Samples
CN106250874A (zh) * 2016-08-16 2016-12-21 东方网力科技股份有限公司 一种服饰及随身物品的识别方法和装置
CN107247947A (zh) * 2017-07-07 2017-10-13 北京智慧眼科技股份有限公司 人脸属性识别方法及装置
CN107330396A (zh) * 2017-06-28 2017-11-07 华中科技大学 一种基于多属性和多策略融合学习的行人再识别方法
CN108052894A (zh) * 2017-12-11 2018-05-18 北京飞搜科技有限公司 一种目标对象的多属性识别方法、设备、介质及神经网络
CN108921022A (zh) * 2018-05-30 2018-11-30 腾讯科技（深圳）有限公司 一种人体属性识别方法、装置、设备及介质

Family Cites Families (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104992142B (zh) * 2015-06-03 2018-04-17 江苏大学 一种基于深度学习和属性学习相结合的行人识别方法
KR20170006356A (ko) * 2015-07-08 2017-01-18 주식회사 케이티 이차원 영상 기반 고객 분석 방법 및 장치
CN105957086B (zh) * 2016-05-09 2019-03-26 西北工业大学 一种基于优化神经网络模型的遥感图像变化检测方法
CN108073876B (zh) * 2016-11-14 2023-09-19 北京三星通信技术研究有限公司 面部解析设备和面部解析方法
BR112019019517A8 (pt) * 2017-03-20 2023-04-04 Huawei Tech Co Ltd Método e aparelho para reconhecer atributo descritivo de característica de aparência
CN107316001A (zh) * 2017-05-31 2017-11-03 天津大学 一种自动驾驶场景中小且密集的交通标志检测方法
CN107516316B (zh) * 2017-07-19 2020-11-20 盐城禅图智能科技有限公司 一种在fcn中引入聚焦机制对静态人体图像进行分割的方法
CN107678059B (zh) * 2017-09-05 2019-06-28 中国石油大学(北京) 一种储层含气识别的方法、装置及系统
CN109598176A (zh) * 2017-09-30 2019-04-09 佳能株式会社 识别装置和识别方法
CN108053061B (zh) * 2017-12-08 2021-08-31 天津大学 一种基于改进卷积神经网络的太阳能辐照度预测方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20130322740A1 (en) * 2012-05-31 2013-12-05 Lihui Chen Method of Automatically Training a Classifier Hierarchy by Dynamic Grouping the Training Samples
CN106250874A (zh) * 2016-08-16 2016-12-21 东方网力科技股份有限公司 一种服饰及随身物品的识别方法和装置
CN107330396A (zh) * 2017-06-28 2017-11-07 华中科技大学 一种基于多属性和多策略融合学习的行人再识别方法
CN107247947A (zh) * 2017-07-07 2017-10-13 北京智慧眼科技股份有限公司 人脸属性识别方法及装置
CN108052894A (zh) * 2017-12-11 2018-05-18 北京飞搜科技有限公司 一种目标对象的多属性识别方法、设备、介质及神经网络
CN108921022A (zh) * 2018-05-30 2018-11-30 腾讯科技（深圳）有限公司 一种人体属性识别方法、装置、设备及介质

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111414812A (zh) * 2020-03-03 2020-07-14 平安科技（深圳）有限公司 人体属性识别方法、系统、计算机设备及存储介质
CN114529731A (zh) * 2020-10-30 2022-05-24 北京眼神智能科技有限公司 人脸特征点定位及属性分析方法、装置、存储介质及设备

Also Published As

Publication number Publication date
CN108921022A (zh) 2018-11-30
US11275932B2 (en) 2022-03-15
US20200356767A1 (en) 2020-11-12

Similar Documents

Publication Publication Date Title
WO2019228089A1 (zh) 人体属性识别方法、装置、设备及介质
US11042728B2 (en) Electronic apparatus for recognition of a user and operation method thereof
CN105376255B (zh) 一种基于K-means聚类的Android平台入侵检测方法
WO2018121690A1 (zh) 对象属性检测、神经网络训练、区域检测方法和装置
WO2016127883A1 (zh) 一种图像区域检测方法及装置
WO2018054329A1 (zh) 物体检测方法和装置、电子设备、计算机程序和存储介质
KR102056806B1 (ko) 영상 통화 서비스를 제공하는 단말과 서버
WO2021051917A1 (zh) 人工智能ai模型的评估方法、系统及设备
CN110163076A (zh) 一种图像数据处理方法和相关装置
US11720627B2 (en) Systems and methods for efficiently sending video metadata
CN108205685A (zh) 视频分类方法、视频分类装置及电子设备
CN107301376B (zh) 一种基于深度学习多层刺激的行人检测方法
US9959480B1 (en) Pixel-structural reference image feature extraction
CN110443366A (zh) 神经网络的优化方法及装置、目标检测方法及装置
JP2023536025A (ja) 路車協調における目標検出方法、装置及び路側機器
CN115797736B (zh) 目标检测模型的训练和目标检测方法、装置、设备和介质
US20230267147A1 (en) Systems and methods for searching for events within video content
CN113326773A (zh) 识别模型训练方法、识别方法、装置、设备及存储介质
Wahyono et al. A new computational intelligence for face emotional detection in ubiquitous
CN113901911B (zh) 图像识别、模型训练方法、装置、电子设备及存储介质
CN113469025B (zh) 应用于车路协同的目标检测方法、装置、路侧设备和车辆
CN108537990B (zh) 一体机欺诈判断方法、装置、设备及计算机可读存储介质
US20230205817A1 (en) Systems and methods for identifying events within video content using intelligent search query
US20230072641A1 (en) Image Processing and Automatic Learning on Low Complexity Edge Apparatus and Methods of Operation
CN116824419A (zh) 一种着装特征识别方法、识别模型的训练方法及装置

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: 19810487

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19810487

Country of ref document: EP

Kind code of ref document: A1