WO2020147414A1 - 网络优化方法及装置、图像处理方法及装置、存储介质 - Google Patents

网络优化方法及装置、图像处理方法及装置、存储介质 Download PDF

Info

Publication number
WO2020147414A1
WO2020147414A1 PCT/CN2019/118922 CN2019118922W WO2020147414A1 WO 2020147414 A1 WO2020147414 A1 WO 2020147414A1 CN 2019118922 W CN2019118922 W CN 2019118922W WO 2020147414 A1 WO2020147414 A1 WO 2020147414A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
feature
loss value
new
image pair
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2019/118922
Other languages
English (en)
French (fr)
Inventor
葛艺潇
沈岩涛
陈大鹏
王晓刚
李鸿升
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing Sensetime Technology Development Co Ltd
Original Assignee
Beijing Sensetime Technology Development Co Ltd
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 Beijing Sensetime Technology Development Co Ltd filed Critical Beijing Sensetime Technology Development Co Ltd
Priority to KR1020207033113A priority Critical patent/KR102454515B1/ko
Priority to SG11202009395SA priority patent/SG11202009395SA/en
Priority to JP2020552759A priority patent/JP7074877B2/ja
Publication of WO2020147414A1 publication Critical patent/WO2020147414A1/zh
Priority to US17/037,654 priority patent/US11416703B2/en
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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 OR CALCULATING; 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 OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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 OR CALCULATING; 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
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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
    • G06N3/047—Probabilistic or stochastic networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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
    • G06N3/0475—Generative networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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
    • G06N3/09—Supervised learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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
    • G06N3/094—Adversarial learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/70—Denoising; Smoothing
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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
    • G06V10/454—Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Definitions

  • the present disclosure relates to the field of network optimization, and in particular to a network optimization method and device, image processing method and device, and storage medium.
  • Pedestrian re-identification aims to learn discriminative features for pedestrian retrieval and matching.
  • factors such as pedestrian pose diversity and background diversity in the image data set affect the extraction of identity features.
  • deep neural networks are used to extract and decompose features for identity recognition.
  • the embodiments of the present disclosure provide a network optimization technical solution.
  • a network optimization method for optimizing a neural network which includes:
  • Acquire an image sample group the image sample group including image pairs formed by images of the same object, and image pairs formed by images of different objects; obtain the first feature and the second feature of the images in the image sample group, and The first classification result is obtained by using the first feature of the image, the first feature includes the identity feature, and the second feature includes the attribute feature; performing feature exchange processing on the image pair in the image sample group to obtain a new image pair, so The feature exchange process is to generate a new first image using the first feature of the first image and the second feature of the second image in the image pair, and to generate a new first image using the second feature of the first image and the first feature of the second image.
  • a new second image using a preset method to obtain the first loss value of the first classification result, the second loss value of the new image pair, and the first feature and second feature of the new image pair
  • the characteristic third loss value at least adjust the parameters of the neural network according to the first loss value, the second loss value and the third loss value until the preset requirements are met.
  • the acquiring the first feature and the second feature of the image in the image sample group includes:
  • the first loss value of the first classification result, the second loss value of the new image pair, and the first loss value of the new image pair are obtained in a preset manner.
  • the third loss value of the feature and the second feature includes:
  • the method before the two images of the image pair are input to the identity encoding network module, the method further includes: adding noise to the image area of the object in the two images of the image pair .
  • the performing feature exchange processing on the image pair in the image sample group to obtain a new image pair includes:
  • the first feature and the second feature of the image in the image pair of the image sample group are input to the generating network module of the neural network; the feature exchange is performed on the image pair in the image sample group through the generating network module Processing to obtain the new image pair.
  • performing feature exchange processing on the image pair in the image sample group to obtain a new image pair includes: The image performs a feature exchange process to obtain the new image pair;
  • the performing a feature exchange process on the images in the image pair to obtain the new image pair includes:
  • a new first image is generated using the first feature of the first image and the second feature of the second image in the image pair, and a new first image is generated using the second feature of the first image and the first feature of the second image Two images.
  • performing feature exchange processing on the image pairs in the image sample group to obtain a new image pair includes: The image performs feature exchange processing twice to obtain a new image pair;
  • the performing feature exchange processing twice on the images in the image pair to obtain a new image pair includes:
  • the first loss value of the first classification result, the second loss value of the new image pair, and the first loss value of the new image pair are obtained in a preset manner.
  • the third loss value of the feature and the second feature includes: obtaining the second loss value of the new image pair obtained by the generating network module relative to the original image pair by using the second preset method, the original image pair and The new image pair corresponds.
  • the first loss value of the first classification result, the second loss value of the new image pair, and the first loss value of the new image pair are obtained in a preset manner.
  • the third loss value of the feature and the second feature includes:
  • the first feature and the second feature of the new image pair are obtained based on the first feature and the second feature of the new image pair and the corresponding first feature and second feature of the original image pair.
  • the third loss value of the second feature, the original image pair corresponds to the new image pair.
  • the method further includes: inputting the generated new image pair to the neural network for identification
  • the network module obtains the label feature representing the true degree of the new image pair; and obtains the fourth loss value of the new image pair based on the label feature by using a fourth preset method.
  • the adjusting the parameters of the neural network at least according to the first loss value, the second loss value, and the third loss value until a preset requirement is met includes:
  • the using the first loss value, the second loss value, the third loss value, and the fourth loss value to obtain the loss value of the neural network includes:
  • the fifth preset method is used to obtain the first loss value, the second loss value, the third loss value, and the fourth loss value.
  • the first network loss value of the neural network when the image sample group input to the neural network is an image pair of different objects, a sixth preset method is used based on the first loss value, the second loss value, and the third loss Value and the fourth loss value to obtain the second network loss value of the neural network; the loss value of the neural network is obtained based on the sum result of the first network loss value and the second network loss value.
  • an image processing method which includes:
  • the network model obtained after the optimization of the network optimization method described in the item.
  • a network optimization device which includes:
  • An acquisition module for acquiring an image sample group, the image sample group including image pairs formed by images of the same object, and image pairs formed by images of different objects; a feature encoding network module for acquiring the images The first feature and the second feature of the image in the sample group; the classification module is used to obtain the first classification result according to the first feature of the image; the generation network module is used to perform feature exchange on the image pairs in the image sample group A new image pair is obtained by processing, and the feature exchange processing is to generate a new first image using the first feature of the first image and the second feature of the second image in the image pair, and using the second feature of the first image and The first feature of the second image generates a new second image; the loss value acquisition module is used to obtain the first loss value of the first classification result and the second loss of the new image pair by using a preset method Value, and the third loss value of the first feature and the second feature of the new image pair; an adjustment module for adjusting the neural network at least according to the first loss value, the second loss value, and the third loss value
  • the feature encoding network module includes an identity encoding network module and an attribute encoding network module, wherein:
  • the acquisition module is also used to input the two images of the image pair to the identity coding network module and the attribute coding network module; and the identity coding network module is used to obtain the information of the two images in the image pair.
  • the first feature, and the attribute coding network module are used to obtain the second feature of the two images in the image pair.
  • the loss value acquisition module is also used to obtain the first classification result corresponding to the first feature acquired through the identity coding network module, and uses a first preset method to obtain a first classification result according to the first classification.
  • the result is the actual classification result corresponding to the image in the image sample group, and the first loss value is obtained.
  • the device further includes: a preprocessing module, which is configured to send the two images of the image pair to the two images of the image pair before inputting the two images of the image pair to the identity coding network module. Noise is added to the image area of the middle object.
  • the generating network module is further configured to perform a feature exchange process on the images in the image pair to obtain the new image pair when the input image pair is an image of the same object. ; Specifically used to generate a new first image using the first feature of the first image and the second feature of the second image in the image pair, and using the second feature of the first image and the first feature of the second image Generate a new second image.
  • the generating network module is further configured to perform feature exchange processing on the images in the image pair twice to obtain a new image pair when the input image pair is images of different objects; Specifically used to generate a new first intermediate image using the first feature of the first image and the second feature of the second image in the image pair, and using the second feature of the first image and the first feature of the second image Generate a new second intermediate image; use the first feature of the first intermediate image and the second feature of the second intermediate image to generate a new first image, use the second feature of the first intermediate image and the second intermediate image The first feature of the image generates a new second image.
  • the loss value acquisition module is further configured to obtain a second loss value of the new image pair obtained by the generating network module relative to the original image pair by using a second preset manner, and the The original image pair corresponds to the new image pair.
  • the loss value acquisition module is further configured to, in a third preset manner, based on the first feature and the second feature of the new image pair and the corresponding first feature and the second feature of the original image pair.
  • the second feature obtains the first feature and the third loss value of the second feature of the new image pair, and the original image pair corresponds to the new image pair.
  • the device further includes:
  • the identification network module is used to receive the new image pair and obtain a label feature representing the authenticity of the new image pair; the loss value acquisition module is also used to use a fourth preset method based on the The label feature obtains the fourth loss value of the new image pair.
  • the adjustment module is further configured to use the first loss value, the second loss value, the third loss value, and the fourth loss value to obtain the loss value of the neural network; and use the The loss value of the neural network adjusts the parameters of the neural network until the preset requirements are met.
  • the adjustment module is further configured to use a fifth preset method based on the first loss value and the second loss value when the image sample group input to the neural network is an image pair of the same object.
  • the loss value, the third loss value, and the fourth loss value obtain the first network loss value of the neural network; when the image sample group input to the neural network is an image pair of different objects, the sixth preset method is used based on The first loss value, the second loss value, the third loss value, and the fourth loss value obtain the second network loss value of the neural network; based on the sum of the first network loss value and the second network loss value As a result, the loss value of the neural network is obtained.
  • an image processing device including:
  • the receiving module is used to receive the input image; the recognition module is used to recognize the first feature of the input image through a neural network model; the identity determination module is used to determine the input image based on the first feature The identity of the object; wherein the neural network model is a network model obtained after optimization processing by the network optimization method described in any one of the first aspect.
  • an electronic device including:
  • a processor ; a memory for storing executable instructions of the processor; wherein the processor is configured to execute the method of any one of the first aspect and the second aspect.
  • a computer-readable storage medium on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the computer program instructions in any one of the first and second aspects are implemented.
  • a computer program wherein the computer program includes computer-readable code, and when the computer-readable code runs in an electronic device, a processor in the electronic device executes To achieve the above network optimization method.
  • the embodiments of the present disclosure can effectively extract the first feature (identity feature) in the input image and the second feature other than the first feature, and form a new picture by exchanging the second features of the two images, which can successfully Separate the identity-related features from the non-identity features, and the identity-related features can be effectively used for pedestrian re-identification.
  • the embodiments of the present disclosure provide that any auxiliary information other than the image data set is not required in the training and application stages, and can provide sufficient generation supervision and effectively improve the recognition accuracy.
  • Fig. 1 shows a flowchart of a network optimization method according to an embodiment of the present disclosure
  • Fig. 2 shows a flowchart of step S200 in a network optimization method according to an embodiment of the present disclosure
  • Fig. 3 shows a flowchart of step S300 of a network optimization method according to an embodiment of the present disclosure
  • step S303 of the network optimization method shows a flowchart of step S303 of the network optimization method according to an embodiment of the present disclosure
  • Fig. 5 shows a flowchart of step S400 in an image processing method according to an embodiment of the present disclosure
  • FIG. 6 shows a schematic diagram of a process of performing network optimization processing using the first type of samples according to an embodiment of the present disclosure
  • FIG. 7 shows a schematic diagram of a process of performing network optimization processing using a second type of samples according to an embodiment of the present disclosure
  • FIG. 8 shows a flowchart of an image processing method according to an embodiment of the present disclosure
  • Fig. 9 shows a block diagram of a network optimization device according to an embodiment of the present disclosure.
  • Fig. 10 shows a block diagram of an image processing device according to an embodiment of the present disclosure
  • FIG. 11 shows a block diagram of an electronic device 800 according to an embodiment of the present disclosure
  • FIG. 12 shows a block diagram of an electronic device 1900 according to an embodiment of the present disclosure.
  • the embodiments of the present disclosure provide a network optimization method, which can be used to train neural networks or other machine learning networks. For example, it can be applied to the training process of machine learning networks in scenarios such as face recognition and identity verification of target users. It can also be applied to the training process of networks that require high accuracy such as authenticating images.
  • the present disclosure does not limit specific application scenarios. As long as the process implemented by the network optimization method provided by the present disclosure is within the protection scope of the present disclosure Inside.
  • the embodiment of the present disclosure takes a neural network as an example for description, but does not specifically limit this. After training by the network optimization method of the embodiments of the present disclosure, the recognition accuracy of the network of the human object can be improved, and at the same time, no auxiliary information other than the input image is required, which is simple and convenient.
  • the network optimization solution provided by the embodiments of the present disclosure can be executed by a terminal device, a server, or other types of electronic devices, where the terminal device can be a user equipment (UE), a mobile device, a user terminal, a terminal, a cellular phone, or a cordless Telephone, personal digital assistant (PDA), handheld devices, computing devices, vehicle-mounted devices, wearable devices, etc.
  • the network optimization method can be implemented by a processor invoking computer-readable instructions stored in a memory.
  • Fig. 1 shows a flowchart of a network optimization method according to an embodiment of the present disclosure.
  • the network optimization method of an embodiment of the present disclosure may include:
  • S100 Acquire an image sample group, the image sample group including image pairs formed by images of the same object, and image pairs formed by images of different objects;
  • S200 Obtain the first feature and the second feature of the image in the image sample group, and obtain a first classification result by using the first feature of the image, the first feature includes an identity feature, and the second feature includes an attribute feature;
  • S300 Perform feature exchange processing on the images in the image sample group to obtain a new image pair.
  • the feature exchange processing is to generate a new first image using the first feature of the first image and the second feature of the second image in the image pair.
  • S500 Adjust the parameters of the neural network according to at least the first loss value, the second loss value, and the third loss value until the preset requirements are met.
  • the image sample group when training the neural network through the embodiments of the present disclosure, may be input to the neural network first, and the image sample group is used as the sample image for training the neural network.
  • the image sample group may include two types of image samples.
  • the first type of samples are image pairs composed of different images of the same object
  • the second type of samples are image pairs composed of different images of different objects. That is, in the first type of samples, the images in each image pair are different images of the same object, and in the second type of samples, the images in each image pair are different images of different objects.
  • Each image pair may include two images, such as the first image and the second image described below.
  • the embodiments of the present disclosure may separately use the two types of image samples to train the neural network.
  • At least one image in the image sample group of the embodiment of the present disclosure may have a corresponding identification, and the identification may correspond to the object in the image, and is used to distinguish the identity of the person object in the image.
  • at least one image in the image sample group may have a real classification label corresponding to its corresponding object, which can be represented in the form of a matrix, and the accuracy of the classification results of the neural network model can be compared according to the real classification label. , If you can determine the corresponding loss value.
  • the method of obtaining the image sample group may include: using a communication component to receive the image sample group transmitted by other electronic devices, such as receiving the image sample group from a server, a mobile phone, any computer device, and the like.
  • At least one image in the image sample group may be a video image collected by a camera, and multiple image pairs obtained after encoding processing, but it is not a specific limitation of the present disclosure.
  • the specific optimization process of the neural network can be executed.
  • the first feature can include the identity feature of the object in the image, such as the color, shape, and accessory features of the clothing
  • the second feature can be a feature other than the first feature, such as an attribute feature, which can include the posture feature of the person object , Background characteristics, environmental characteristics, etc.
  • the method of obtaining the first feature and the second feature is described below with examples.
  • FIG. 2 shows a flowchart of step S200 in the network optimization method according to an embodiment of the present disclosure, wherein the first feature and the second feature of the image in the image sample group are acquired, and the first feature is obtained by using the first feature of the image.
  • a classification result including:
  • S202 Use the identity coding network module to obtain the first features of the two images in the image pair, and use the attribute coding network module to obtain the second features of the two images in the image pair;
  • S203 Obtain a first classification result corresponding to the first feature by using the classification module of the neural network.
  • the neural network of the embodiment of the present disclosure may include an identity coding network module and an attribute coding network module.
  • the identity coding network module can be used to identify the identity characteristics of objects in an image
  • the attribute coding network module can be used to recognize the attribute characteristics of objects in the image . Therefore, at least one image pair in the acquired image sample group can be input to the aforementioned identity coding network module and attribute coding network module respectively.
  • the first feature of the two images in the received image pair can be obtained through the identity coding network module, and the second feature of the two images in the received image pair can be obtained using the attribute coding network module.
  • the two images in the input image pair are represented by A and B respectively, then the first feature of A obtained by the identity coding network module is A u , and the first feature of B obtained by the identity coding network module is B u , through the attribute
  • the second feature of A obtained by the coding network module is A v
  • the second feature of B obtained by the attribute coding network module is B v .
  • the identity coding network module can use a preset character feature extraction algorithm to extract the first feature in the image, or it can also include a convolution module, a pooling module and other module units to perform the acquisition of the first feature.
  • the identity coding network The structure of the module is not specifically limited in the embodiment of the present disclosure. As long as the first feature in the image can be extracted, it can be used as the identity encoding network module of the embodiment of the present disclosure.
  • the attribute coding network module can also use a preset pose and background feature algorithm to extract the second feature in the image, or it can also include a module unit such as a convolution module.
  • a module unit such as a convolution module.
  • the embodiments of the present disclosure This is not specifically limited, as long as the second feature in the image can be extracted, it can be used as the attribute coding network module of the embodiment of the present disclosure.
  • the embodiments of the present disclosure can perform classification and recognition operations using the first feature, and can also perform subsequent feature exchange processing.
  • the neural network of the embodiment of the present disclosure may also include a classification module, and the output side of the identity coding network module may be connected to the input side of the classification module, so as to receive the first feature output by the identity coding network module, and the classification module may receive The first feature of get the first classification result.
  • the first classification result is used to indicate the prediction result of the identity identifier corresponding to the first feature, and the prediction result may be embodied in the form of a matrix, and the elements of the matrix are the probability of predicting the object identifier.
  • composition of the classification module in the embodiment of the present disclosure can be set by itself, and it can use the set classification principle to obtain the first classification result corresponding to the first feature, and as long as the classification of the first feature can be performed, it can be used as an embodiment of the present disclosure.
  • the first loss value corresponding to the first classification result can be obtained, and the loss value of the neural network can be further determined according to the first loss value, and the parameters in the network can be adjusted by feedback.
  • feature exchange processing between every two images of the image pair can be performed.
  • the feature exchange processing may be to exchange the second feature of the first image and the second feature of the second image in the image pair, and obtain the result based on the first feature and the exchanged second feature. New image.
  • the first feature of one image can be combined with the second feature of another image to form a new image.
  • the new image is used to perform classification, which can effectively realize the identification of people based on identity features. And reduce the influence of background, posture and other attributes.
  • Fig. 3 shows a flowchart of step S300 in the image processing method according to an embodiment of the present disclosure, wherein the performing feature exchange processing on the image pair in the image sample group to obtain a new image pair may include:
  • S301 Input the image pairs of the image sample group to the generation network module of the neural network;
  • S302 Perform the feature exchange processing on the image pair in the image sample group through the generation network module to obtain the new image pair.
  • the neural network in the embodiment of the present disclosure may further include a generating network module, which may perform feature exchange processing on the first feature and the second feature obtained by the identity coding network module and the attribute coding network module, and according to the exchanged features Get a new image.
  • the image sample group input in the embodiment of the present disclosure may include two types of image sample groups. Among them, the image pairs in the first type of samples are images of the same object. For the image pairs of the first type of sample, the embodiment of the present disclosure may perform feature exchange processing once on the images in each image pair.
  • the performing feature exchange processing on the images in the image sample group to obtain a new image pair may include: performing feature exchange processing once on the images in the image pair to obtain the new image Correct.
  • the process can include:
  • the new image obtained after the feature exchange processing is performed can still be the image of the same object.
  • the difference between the obtained new image and the corresponding original image, and the first feature and second feature of the new image and the first feature and second feature of the corresponding original image can be used.
  • the generated new image pair can be input to the classification module, and the classification can be performed to obtain the second classification result.
  • the image pair in the first type of sample includes image A and image B.
  • the first feature of A can be obtained through the identity coding network module as A u
  • the first feature of B obtained by the identity coding network module is B u
  • the second feature of A obtained by the coding network module is A v
  • the second feature of B obtained by the attribute coding network module is B v .
  • a and B are the first image and the second image of the same object, respectively, and the first image and the second image are different.
  • a first feature and the second feature U A B V B of a first image to obtain a new A '
  • B using a first characteristic and U A B A second feature V A new second image B' is obtained.
  • the neural network of the embodiment of the present disclosure may include a generating network module, and the generating network module may be used to generate a new image according to the received first feature and second feature.
  • the generating network module may include at least one convolution unit, or may also include other processing units, and the image corresponding to the first feature and the second feature can be obtained through the generating network module. That is, the exchange of the above-mentioned second feature and the process of generating an image based on the exchanged feature can be completed through the generation network.
  • a new picture can be formed by exchanging the second features of the two images, so that the identity-related features and the identity-unrelated features can be successfully separated, and the neural network can be trained in this way , Can improve the recognition accuracy of the neural network for identity features.
  • the image sample group of the embodiment of the present disclosure may also include a second type of sample group, in which image pairs are images of different objects.
  • the embodiment of the present disclosure may perform feature exchange processing twice on the images in each image pair.
  • FIG. 4 shows a flowchart of step S303 of the network optimization method according to an embodiment of the present disclosure, where the input image pair is an image of a different object, the image sample group is Performing feature exchange processing on the inner image to obtain a new image pair may include: performing feature exchange processing on the images in the image pair twice to obtain a new image pair.
  • the process may include:
  • S3031 Use the first feature of the first image and the second feature of the second image in each image pair in the second-type sample to generate a new first intermediate image, and use the second feature of the first image And the first feature of the second image generates a new second intermediate image;
  • S3032 Generate a new first image using the first feature of the first intermediate image and the second feature of the second intermediate image, and generate a new first image using the second feature of the first intermediate image and the first feature of the second intermediate image The new second image.
  • the first feature of A obtained by the identity coding network module is A u
  • the first feature of B obtained by the identity coding network module is Bu
  • the second characteristic of A obtained by the attribute coding network module is A v
  • the attribute The second feature of B obtained by the encoding network module is B v .
  • a and B are the first image and the second image of different objects, respectively.
  • the network module can use the identification code and attribute network encoding module respectively acquire again the first intermediate image A 'of the first characteristic A' and a second U wherein A 'V, and the second 'first feature B' and two second intermediate image B u wherein B 'v, and further performing a first network using the generated intermediate image a' of the second characteristic a 'and the second intermediate image V B' second feature B′ v exchange processing, and use the first feature A′ u of the first intermediate image A′ and the second feature B′ v of the second intermediate image B′ to generate a new first image A′′, and use the first intermediate image A′
  • a new picture can be formed by exchanging the second feature of the two images.
  • the difference from the training process of the image pair of the same identity object is that for the second type of sample, due to the first feature
  • the second feature exchange process can be performed, and an image corresponding to the original image can be generated.
  • This process can be a cyclic generation process.
  • the difference between the obtained new image and the corresponding original image, and the first feature and second feature of the new image and the first feature and second feature of the corresponding original image can be used.
  • the first feature of the new image can also be input to the classification module to perform classification processing to obtain the second classification result.
  • the second classification result of the first feature of the final new image can be directly obtained, and for the case of the second type of sample, in addition to the first feature of the final new image, the first feature can be obtained.
  • the second classification result can also be the second classification result of the first feature of the intermediate image.
  • the embodiments of the present disclosure can optimize the neural network according to the above-mentioned second classification result and the difference between the new image and the original image and the difference between the features. That is, the embodiment of the present disclosure can perform feedback adjustment on the neural network according to the loss value of the output result obtained by each network module of the neural network, until the loss value of the neural network meets the preset requirement, if it is lower than the loss threshold, it can be determined to meet the Pre-determined requirements.
  • the loss function of the neural network of the embodiment of the present disclosure may be related to the loss function of the network module, for example, it may be the weighted sum of the loss functions of multiple network modules, based on which the loss value of each network module can be used to obtain the loss value of the neural network In this way, the parameters of each network module of the neural network are adjusted until the preset requirement that the loss value is lower than the loss threshold is met.
  • the loss threshold can be set by those skilled in the art according to requirements, and this disclosure does not specifically limit this.
  • the classification module can obtain the first classification result according to the first feature.
  • the embodiments of the present disclosure can use the first preset method to obtain the The first loss value of the first classification result obtained by the obtained first feature.
  • 5 shows a flowchart of step S400 in the image processing method according to an embodiment of the present disclosure, wherein the process of obtaining the first loss value includes:
  • S402 Using a first preset manner, obtain the first loss value according to the first classification result and the actual classification result corresponding to the image in the image sample group.
  • step S200 when the first feature of the image in the sample is obtained, the classification and recognition of the first feature can be performed by the classification module to obtain the first classification result corresponding to the first feature.
  • the first classification result can be expressed in the form of a matrix, in which each element is expressed as a probability corresponding to each identity tag.
  • the first difference can be obtained.
  • the first difference can be determined as the first loss value.
  • the first classification result and the real classification result may also be input into the first loss function of the classification module to obtain the first loss value, which is not specifically limited in the present disclosure.
  • the loss functions used may be the same or different.
  • the embodiment of the present disclosure can add the loss value of the neural network obtained through the training of the first type of sample and the loss value of the neural network obtained through the training of the second type of sample to obtain the final loss value of the neural network, and The loss value is used to perform feedback adjustment processing on the network.
  • the parameters of each network module of the neural network can be adjusted, or only a part of the network modules can be adjusted. This is not disclosed without specific restrictions.
  • the embodiment of the present disclosure may use a first preset method to obtain the first loss value of the first classification result obtained by the first feature obtained by the identity coding network module.
  • the expression of the first preset mode can be as shown in formula (1):
  • C(I) represents the N-dimensional prediction feature vector corresponding to the first classification result
  • L is the N-dimensional feature vector corresponding to the true label of the corresponding original image (real classification result)
  • L c is the first loss function
  • the corresponding first loss value, i is a variable greater than or equal to 1 and less than or equal to N.
  • the embodiment of the present disclosure can perform feedback adjustment on the parameters of the identity coding network module, the attribute coding network module, and the classification module according to the first loss value, and can also determine the neural network according to the first loss value and the loss values of other network modules.
  • the overall loss value of performs unified feedback adjustment on at least one network module of the neural network, which is not limited in the present disclosure.
  • the embodiment of the present disclosure can also process the new image pair generated by the generating network module to obtain the second loss value of the new image pair and the third loss value of the corresponding feature.
  • the manner of determining the second loss value may include: using a second preset manner to obtain the second loss value of the new image pair obtained by the generating network module relative to the original image pair.
  • a new image pair can be obtained by generating a network, and the embodiment of the present disclosure can determine the second loss value according to the difference between the new image pair and the original image pair.
  • the expression of the second preset mode may be as shown in formula (2):
  • the second loss value corresponding to the new image pair generated by the generating network module can be obtained for the first type of sample.
  • the expression of the second preset mode may be as shown in formula (3):
  • the embodiment of the present disclosure can also obtain the third loss value corresponding to the feature of the new second image pair, where the third loss value can be obtained by using a third preset method.
  • the expression of the third preset mode can be as shown in formula (4):
  • I Xu represents the first feature of the first image X u in the original image pair
  • I Xv represents the second feature of the second image X v in the original image pair
  • I Xv represents the second feature of the second image X v in the original image pair
  • T is the transposition operation
  • L s is the loss value corresponding to the third loss function
  • 2 represents the 2 norm.
  • the third loss value corresponding to the feature of the new image pair generated by the generating network module can be obtained by the classification module.
  • the embodiment of the present disclosure may perform feedback adjustment on the parameters of the generating network module according to the second loss value and the third loss value, or may simultaneously perform feedback adjustment on multiple network modules of the neural network in combination with the first loss value.
  • the weighted sum of the first loss value, the second loss value, and the third loss value can be used to obtain the loss value of the neural network, that is, the loss function of the neural network is
  • the weighted sum of the above-mentioned first loss function, second loss function, and third loss function, the weight of each loss function is not specifically limited in the present disclosure, and those skilled in the art can set it according to needs.
  • the first loss function, the second loss function, and the third loss function when training the image pair based on the first type of sample are the same as the first loss function when the image pair based on the second type is performed.
  • the first loss function, the second loss function, and the third loss function may be different, but they are not a specific limitation of the present disclosure.
  • the neural network of the embodiment of the present disclosure may also include a discrimination network module, which may be used to discriminate the true or false degree (degree of truth) of the generated new image pair. According to the degree of authenticity), the fourth loss value corresponding to the new image pair determined by the identification network module can be obtained according to the degree of authenticity.
  • the discrimination network and the generation network of the embodiments of the present disclosure may constitute a generation confrontation network.
  • the embodiment of the present disclosure may input the generated new image pair to the identification network module of the neural network, and obtain the fourth loss value of the new image pair by using the fourth preset method.
  • D represents the model function of the identification network module
  • E[] represents the expectation
  • X represents the original image corresponding to the new image, that is, the real image.
  • D(X) represents the label feature of the identification network module to the real image, Represents the identification of the tag feature of the new image input by the network module.
  • the element in is a value between zero and one. The closer to 1, the higher the degree of authenticity of the element.
  • the training process of the discrimination network module can be performed separately, that is, the new image generated and the corresponding real image can be input to the discrimination network module, and the discrimination network module can be trained based on the fourth loss function mentioned above, Until the loss value corresponding to the fourth loss function is lower than the loss threshold required by training.
  • step S400 in the embodiment of the present disclosure can also use the first The loss value, the second loss value, the third loss value, and the fourth loss value obtain the loss value of the neural network.
  • the loss function of the neural network is the weighted sum of the first loss function, the second loss function, the third loss function, and the fourth loss function.
  • the weight of each loss function is not specifically limited in this disclosure. Technicians can set it according to their needs.
  • the first loss function, the second loss function, and the third loss function when training the image pair based on the first type of sample are compared with the image pair based on the second pseudo group.
  • the first loss function, the second loss function, and the third loss function may be different, but they are not a specific limitation of the present disclosure.
  • the fifth preset method is used based on the The first loss value, the second loss value, the third loss value, and the fourth loss value obtain the first network loss value of the neural network, where the expression of the fifth preset mode is as shown in formula (6):
  • L intra L c + ⁇ ir L ir + ⁇ s L s + ⁇ adv L adv formula (6);
  • ⁇ ir , ⁇ s and ⁇ adv are the weights of the second loss function, the third loss function, and the fourth loss function, respectively, and L intra is the first network loss value.
  • a sixth preset method is used based on the first loss value, the second loss value, the third loss value, and the fourth loss value.
  • the loss value obtains the second network loss value of the neural network.
  • the expression of the sixth preset mode is shown in formula (7):
  • L inter L c + ⁇ cr L cr + ⁇ s L s + ⁇ adv L adv formula (7);
  • ⁇ cr , ⁇ s and ⁇ adv are the weights of the second loss function, the third loss function and the fourth loss function, respectively, and L inter is the second network loss value.
  • the parameters of the neural network can be adjusted by feedback. For example, multiple network modules (identity coding network module, attribute coding network module, generation network module, identification network module, etc.) can be adjusted by feedback Parameters, until the loss value of the neural network is less than the loss threshold, the training can be terminated, and the neural network optimization is now complete.
  • the parameters of the identity coding network module, the attribute coding network module, and the classification module may be adjusted according to the first loss value
  • the parameters of the generating network module may be adjusted by feedback according to the second loss value and the third loss value
  • the parameters of the identification network module are adjusted according to the fourth loss value until the loss value is less than the loss threshold of the corresponding loss function, that is, the training is terminated. That is to say, the embodiments of the present disclosure can individually perform feedback adjustment and training on any network module, or can uniformly adjust part or all of the neural network modules through the loss value of the neural network. Those skilled in the art can select appropriate ones according to their needs. Way to perform the adjustment process.
  • noise can be added to the image before the image sample group is input to the identity coding network module, for example, to the two images of the image pair. Noise is added to the image area of the object in each image.
  • noise is added by adding an overlay layer to a part of the image area of the human object. The size of the overlay layer can be set by those skilled in the art according to requirements, and the present disclosure does not limit this. It should be noted here that the embodiment of the present disclosure only adds noise to the image input to the identity coding network module, and does not introduce noise to other network modules. In this way, the accuracy of neural network identification can be effectively improved.
  • FIG. 6 shows a schematic diagram of the process of performing network optimization processing using the first type of samples according to an embodiment of the present disclosure, in which two images Xu and Xv for the same object can be input to the identity coding network Eid to obtain the first feature, and the image Xu And Xv are input to the attribute coding network Ea to obtain the second feature, and the first feature is input to the classifier C to obtain the first classification result, and the first loss value Lc is obtained.
  • the image input to the identity coding network Eid can be added with noise, for example, an overlay is added to the area of the person object to block part of the area.
  • the second loss value Lir corresponding to the two new images and two new images can be obtained.
  • the third loss value Ls corresponding to the first feature and the second feature is input to the discrimination network module D to obtain the fourth loss value Ladv.
  • the first loss value Lc, the second loss value Lir, the third loss value Ls, and the fourth loss value Ladv can be used to obtain the loss value of the neural network.
  • the loss value is less than the loss threshold, the training is terminated, otherwise the neural network is fed back and adjusted The parameters of at least one network module.
  • FIG. 7 shows a schematic diagram of a process of performing network optimization processing using a second type of samples according to an embodiment of the present disclosure, in which two images Xu and Yw for different objects can be input to the identity coding network Eid to obtain the first feature, and the image Xu And Yw are input to the attribute coding network Ea to obtain the second feature, and the first feature is input to the classifier C to obtain the first classification result, and the first loss value Lc is obtained.
  • the image input to the identity coding network Eid can be added with noise, for example, an overlay is added to the area of the person object to block part of the area.
  • the embodiments of the present disclosure can effectively extract the first feature (identity feature) in the input image and the second feature other than the first feature, and form a new picture by exchanging the second features of the two images, which can successfully Separate the identity-related features from the non-identity features, and the identity-related features can be effectively used for pedestrian re-identification.
  • the embodiments of the present disclosure provide that any auxiliary information other than the image data set is not required in the training and application stages, and can provide sufficient generation supervision and effectively improve the recognition accuracy.
  • the writing order of the steps does not mean a strict execution order but constitutes any limitation on the implementation process.
  • the specific execution order of each step should be based on its function and possibility.
  • the inner logic is determined.
  • the embodiments of the present disclosure also provide an image processing method, which can apply the neural network obtained by the image optimization method provided by the first aspect to perform image recognition operations, and obtain the recognition result of the identity corresponding to the input image.
  • Fig. 8 shows a flowchart of an image processing method according to an embodiment of the present disclosure, wherein the method includes:
  • the neural network model that meets the requirements can be trained from the first aspect, and the neural network model can be used to perform the operation of identifying the object in the image, that is, the neural network model can be used to form an image recognition and other operations.
  • a database may be included, and the database may include information of multiple personnel objects, such as images of personnel objects, and corresponding identity information, such as name, age, position, and other information.
  • the database may include information of multiple personnel objects, such as images of personnel objects, and corresponding identity information, such as name, age, position, and other information.
  • identity information such as name, age, position, and other information. The present disclosure does not limit this .
  • the embodiment of the present disclosure can compare the first feature of the received input image with the image of the person object in the database to determine the person object that matches it in the database. Since the neural network model of the embodiment of the present disclosure is trained by the above-mentioned embodiment and meets the accuracy requirements, the embodiment of the present disclosure can accurately match the object matching the input image, and then obtain its corresponding identity information.
  • the image processing method of the embodiment of the present disclosure can quickly identify the identity of an image object, and can improve the recognition accuracy.
  • the present disclosure also provides image processing devices, electronic equipment, computer-readable storage media, and programs, all of which can be used to implement any image processing method provided in the present disclosure.
  • image processing devices electronic equipment, computer-readable storage media, and programs, all of which can be used to implement any image processing method provided in the present disclosure.
  • FIG. 9 shows a block diagram of a network optimization device according to an embodiment of the present disclosure.
  • the network optimization device includes:
  • the acquisition module 10 is used to acquire an image sample group, which includes image pairs formed by images of the same object and image pairs formed by images of different objects; a feature coding network module 20, which is used to obtain The first feature and the second feature of the image in the image sample group; the classification module 30 is used to obtain the first classification result according to the first feature of the image; the generation network module 40 is used to compare the images in the image sample group A new image pair is obtained by performing feature exchange processing for generating a new first image using the first feature of the first image and the second feature of the second image in the image pair, and using the first image The second feature and the first feature of the second image generate a new second image; the loss value acquisition module 50 is used to obtain the first loss value of the first classification result and the new image by using a preset method The second loss value of the pair, and the third loss value of the first feature and the second feature of the new image pair; the adjustment module 60 is configured to at least according to the first loss value, the second loss value and the third loss value The three loss values adjust the parameters of the
  • the feature encoding network module includes an identity encoding network module and an attribute encoding network module, wherein the acquisition module is also used to input the two images of the image pair to the identity encoding network Module and an attribute coding network module; and the identity coding network module is used to obtain the first feature of two images in the image pair, and the attribute coding network module is used to obtain two images in the image pair The second feature.
  • the loss value acquisition module is also used to obtain the first classification result corresponding to the first feature acquired through the identity coding network module, and uses a first preset method to obtain a first classification result according to the first classification.
  • the result is the actual classification result corresponding to the image in the image sample group, and the first loss value is obtained.
  • the device further includes: a preprocessing module, which is configured to send the two images of the image pair to the two images of the image pair before inputting the two images of the image pair to the identity coding network module. Noise is added to the image area of the middle object.
  • the generating network module is further configured to perform a feature exchange process on the images in the image pair to obtain the new image pair when the input image pair is an image of the same object. , which includes: generating a new first image using the first feature of the first image and the second feature of the second image in the image pair, and using the second feature of the first image and the first feature of the second image Generate a new second image.
  • the generating network module is further configured to perform feature exchange processing twice on the images in the image pair to obtain a new image pair when the input image pair is an image of a different object. It includes: using the first feature of the first image and the second feature of the second image in the image pair to generate a new first intermediate image, and using the second feature of the first image and the first feature of the second image Generate a new second intermediate image; use the first feature of the first intermediate image and the second feature of the second intermediate image to generate a new first image, use the second feature of the first intermediate image and the second intermediate image The first feature of the image generates a new second image.
  • the loss value acquisition module is further configured to obtain a second loss value of the new image pair obtained by the generating network module relative to the original image pair by using a second preset manner, and the The original image pair corresponds to the new image pair.
  • the loss value acquisition module is further configured to, in a third preset manner, based on the first feature and the second feature of the new image pair and the corresponding first feature and the second feature of the original image pair.
  • the second feature obtains the first feature and the third loss value of the second feature of the new image pair, and the original image pair corresponds to the new image pair.
  • the device further includes: a distinguishing network module, which is configured to receive the new image pair and obtain a label feature representing the true degree of the new image pair; and obtain the loss value The module is also used to obtain a fourth loss value of the new image pair based on the label feature by using a fourth preset manner.
  • the adjustment module is further configured to use the first loss value, the second loss value, the third loss value, and the fourth loss value to obtain the loss value of the neural network; and use the The loss value of the neural network adjusts the parameters of the neural network until the preset requirements are met.
  • the adjustment module is further configured to use a fifth preset method based on the first loss value and the second loss value when the image sample group input to the neural network is an image pair of the same object.
  • the loss value, the third loss value, and the fourth loss value obtain the first network loss value of the neural network; when the image sample group input to the neural network is an image pair of different objects, the sixth preset method is used based on The first loss value, the second loss value, the third loss value, and the fourth loss value obtain the second network loss value of the neural network; based on the sum of the first network loss value and the second network loss value As a result, the loss value of the neural network is obtained.
  • Fig. 10 shows a block diagram of an image processing apparatus according to an embodiment of the present disclosure.
  • the image processing device may include:
  • the receiving module 100 is used to receive input images
  • the recognition module 200 is used to recognize the first feature of the input image through a neural network model
  • An identity determination module 300 configured to determine the identity of the object in the input image based on the first feature
  • the neural network model is a network model obtained after optimization processing by the network optimization method described in any one of the first aspect.
  • the functions or modules contained in the apparatus provided in the embodiments of the present disclosure can be used to execute the methods described in the above method embodiments.
  • the functions or modules contained in the apparatus provided in the embodiments of the present disclosure can be used to execute the methods described in the above method embodiments.
  • the embodiments of the present disclosure also provide a computer-readable storage medium on which computer program instructions are stored, and the computer program instructions implement the foregoing method when executed by a processor.
  • the computer-readable storage medium may be a non-volatile computer-readable storage medium.
  • An embodiment of the present disclosure also provides an electronic device, including: a processor; a memory for storing executable instructions of the processor; wherein the processor is configured as the above method.
  • the embodiments of the present disclosure also provide a computer program product, including computer readable code.
  • the processor in the device executes instructions for implementing the above method provided by any of the above embodiments. .
  • the electronic device can be provided as a terminal, server or other form of device.
  • FIG. 11 shows a block diagram of an electronic device 800 according to an embodiment of the present disclosure.
  • the electronic device 800 may be a mobile phone, a computer, a digital broadcasting terminal, a messaging device, a game console, a tablet device, a medical device, a fitness device, a personal digital assistant, and other terminals.
  • the electronic device 800 may include one or more of the following components: a processing component 802, a memory 804, a power supply component 806, a multimedia component 808, an audio component 810, an input/output (I/O) interface 812, and a sensor component 814 , And communication component 816.
  • the processing component 802 generally controls the overall operations of the electronic device 800, such as operations associated with display, telephone calls, data communications, camera operations, and recording operations.
  • the processing component 802 may include one or more processors 820 to execute instructions to complete all or part of the steps in the above method.
  • the processing component 802 may include one or more modules to facilitate interaction between the processing component 802 and other components.
  • the processing component 802 may include a multimedia module to facilitate the interaction between the multimedia component 808 and the processing component 802.
  • the memory 804 is configured to store various types of data to support operations in the electronic device 800. Examples of these data include instructions for any application or method operating on the electronic device 800, contact data, phone book data, messages, pictures, videos, etc.
  • the memory 804 may be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read only memory (EEPROM), erasable and removable Programmable read only memory (EPROM), programmable read only memory (PROM), read only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.
  • SRAM static random access memory
  • EEPROM electrically erasable programmable read only memory
  • EPROM erasable and removable Programmable read only memory
  • PROM programmable read only memory
  • ROM read only memory
  • magnetic memory flash memory
  • flash memory magnetic disk or optical disk.
  • the power supply component 806 provides power for various components of the electronic device 800.
  • the power supply component 806 may include a power management system, one or more power supplies, and other components associated with the generation, management, and distribution of power for the electronic device 800.
  • the multimedia component 808 includes a screen that provides an output interface between the electronic device 800 and the user.
  • the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive input signals from the user.
  • the touch panel includes one or more touch sensors to sense touch, swipe, and gestures on the touch panel. The touch sensor may not only sense the boundary of the touch or sliding action, but also detect the duration and pressure related to the touch or sliding operation.
  • the multimedia component 808 includes a front camera and/or a rear camera. When the electronic device 800 is in an operation mode, such as a shooting mode or a video mode, the front camera and/or the rear camera can receive external multimedia data. Each front camera and rear camera can be a fixed optical lens system or have focal length and optical zoom capabilities.
  • the audio component 810 is configured to output and/or input audio signals.
  • the audio component 810 includes a microphone (MIC).
  • the microphone is configured to receive an external audio signal.
  • the received audio signal may be further stored in the memory 804 or transmitted via the communication component 816.
  • the audio component 810 further includes a speaker for outputting audio signals.
  • the I/O interface 812 provides an interface between the processing component 802 and a peripheral interface module.
  • the peripheral interface module may be a keyboard, a click wheel, or a button. These buttons may include, but are not limited to: home button, volume button, start button, and lock button.
  • the sensor component 814 includes one or more sensors for providing the electronic device 800 with various aspects of state evaluation.
  • the sensor component 814 can detect the on/off status of the electronic device 800 and the relative positioning of the components.
  • the component is the display and the keypad of the electronic device 800.
  • the sensor component 814 can also detect the electronic device 800 or the electronic device 800.
  • the position of the component changes, the presence or absence of contact between the user and the electronic device 800, the orientation or acceleration/deceleration of the electronic device 800, and the temperature change of the electronic device 800.
  • the sensor assembly 814 may include a proximity sensor configured to detect the presence of nearby objects without any physical contact.
  • the sensor component 814 may also include a light sensor, such as a CMOS or CCD image sensor, for use in imaging applications.
  • the sensor component 814 may further include an acceleration sensor, a gyro sensor, a magnetic sensor, a pressure sensor, or a temperature sensor.
  • the communication component 816 is configured to facilitate wired or wireless communication between the electronic device 800 and other devices.
  • the electronic device 800 can access a wireless network based on a communication standard, such as WiFi, 2G or 3G, or a combination thereof.
  • the communication component 816 receives a broadcast signal or broadcast related information from an external broadcast management system via a broadcast channel.
  • the communication component 816 also includes a near field communication (NFC) module to facilitate short-range communication.
  • the NFC module can be implemented based on radio frequency identification (RFID) technology, infrared data association (IrDA) technology, ultra wideband (UWB) technology, Bluetooth (BT) technology, and other technologies.
  • RFID radio frequency identification
  • IrDA infrared data association
  • UWB ultra wideband
  • Bluetooth Bluetooth
  • the electronic device 800 can be implemented by one or more application specific integrated circuits (ASIC), digital signal processors (DSP), digital signal processing devices (DSPD), programmable logic devices (PLD), field A programmable gate array (FPGA), controller, microcontroller, microprocessor, or other electronic components are implemented to implement the above methods.
  • ASIC application specific integrated circuits
  • DSP digital signal processors
  • DSPD digital signal processing devices
  • PLD programmable logic devices
  • FPGA field A programmable gate array
  • controller microcontroller, microprocessor, or other electronic components are implemented to implement the above methods.
  • a non-volatile computer-readable storage medium such as the memory 804 including computer program instructions, which can be executed by the processor 820 of the electronic device 800 to complete the foregoing method.
  • Fig. 12 is a block diagram showing an electronic device 1900 according to an exemplary embodiment.
  • the electronic device 1900 may be provided as a server. 12
  • the electronic device 1900 includes a processing component 1922, which further includes one or more processors, and a memory resource represented by a memory 1932, for storing instructions executable by the processing component 1922, such as application programs.
  • the application programs stored in the memory 1932 may include one or more modules each corresponding to a set of instructions.
  • the processing component 1922 is configured to execute instructions to perform the above method.
  • the electronic device 1900 may also include a power component 1926 configured to perform power management of the electronic device 1900, a wired or wireless network interface 1950 configured to connect the electronic device 1900 to a network, and an input output (I/O) interface 1958 .
  • the electronic device 1900 can operate based on an operating system stored in the memory 1932, such as Windows ServerTM, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM or the like.
  • a non-volatile computer-readable storage medium such as the memory 1932 including computer program instructions, which can be executed by the processing component 1922 of the electronic device 1900 to complete the foregoing method.
  • the present disclosure may be a system, method, and/or computer program product.
  • the computer program product may include a computer-readable storage medium loaded with computer-readable program instructions for causing the processor to implement various aspects of the present disclosure.
  • the computer-readable storage medium may be a tangible device that can hold and store instructions used by the instruction execution device.
  • the computer-readable storage medium may be, but is not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing.
  • Non-exhaustive list of computer readable storage media include: portable computer disks, hard disks, random access memory (RAM), read only memory (ROM), erasable programmable read only memory (EPROM (Or flash memory), static random access memory (SRAM), portable compact disk read-only memory (CD-ROM), digital versatile disk (DVD), memory stick, floppy disk, mechanical encoding device, such as a computer on which instructions are stored
  • RAM random access memory
  • ROM read only memory
  • EPROM erasable programmable read only memory
  • SRAM static random access memory
  • CD-ROM compact disk read-only memory
  • DVD digital versatile disk
  • memory stick floppy disk
  • mechanical encoding device such as a computer on which instructions are stored
  • the convex structure in the hole card or the groove and any suitable combination of the above.
  • the computer-readable storage medium used herein is not to be interpreted as a transient signal itself, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (eg, optical pulses through fiber optic cables), or through wires The transmitted electrical signal.
  • the computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to various computing/processing devices, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and/or a wireless network.
  • the network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and/or edge servers.
  • the network adapter card or network interface in each computing/processing device receives computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in the computer-readable storage medium in each computing/processing device .
  • Computer program instructions for performing the operations of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state setting data, or in one or more programming languages Source code or object code written in any combination.
  • the programming languages include object-oriented programming languages such as Smalltalk, C++, etc., and conventional procedural programming languages such as "C" language or similar programming languages.
  • the computer-readable program instructions can be executed entirely on the user's computer, partly on the user's computer, as an independent software package, partly on the user's computer and partly on a remote computer, or completely on the remote computer or server carried out.
  • the remote computer may be connected to the user's computer through any kind of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (eg, using an Internet service provider to pass the Internet connection).
  • electronic circuits such as programmable logic circuits, field programmable gate arrays (FPGAs), or programmable logic arrays (PLA), can be personalized by using status information of computer-readable program instructions, which can be Computer-readable program instructions are executed to implement various aspects of the present disclosure.
  • These computer-readable program instructions can be provided to the processor of a general-purpose computer, special-purpose computer, or other programmable data processing device, thereby producing a machine that causes these instructions to be executed by the processor of a computer or other programmable data processing device A device that implements the functions/actions specified in one or more blocks in the flowchart and/or block diagram is generated.
  • the computer-readable program instructions may also be stored in a computer-readable storage medium. These instructions enable the computer, programmable data processing apparatus, and/or other devices to work in a specific manner. Therefore, the computer-readable medium storing the instructions includes An article of manufacture that includes instructions to implement various aspects of the functions/acts specified in one or more blocks in the flowcharts and/or block diagrams.
  • each block in the flowchart or block diagram may represent a module, program segment, or part of an instruction, and the module, program segment, or part of an instruction contains one or more Executable instructions.
  • the functions marked in the blocks may also occur in an order different from that marked in the drawings. For example, two consecutive blocks can actually be executed substantially in parallel, and sometimes they can also be executed in reverse order, depending on the functions involved.
  • each block in the block diagrams and/or flowcharts, and combinations of blocks in the block diagrams and/or flowcharts can be implemented with dedicated hardware-based systems that perform specified functions or actions Or, it can be realized by a combination of dedicated hardware and computer instructions.

Landscapes

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

Abstract

本公开涉及一种网络优化方法及装置、图像处理方法及装置、存储介质，其中，所述网络优化方法包括：获取图像样本组；获取所述图像样本组中图像的第一特征和第二特征，并利用图像的第一特征得到第一分类结果；对图像样本组内图像对执行特征交换处理得到新的图像对；利用预设方式，获得所述第一分类结果的第一损失值、所述新的图像对的第二损失值，以及新的图像对的第一特征和第二特征的第三损失值；至少根据第一损失值、第二损失值和第三损失值调节神经网络的参数，直至满足预设要求。本公开实施例能够有效地提高身份识别的精度。

Description

网络优化方法及装置、图像处理方法及装置、存储介质
本公开要求在2019年01月15日提交中国专利局、申请号为201910036096.X、申请名称为“网络优化方法及装置、图像处理方法及装置、存储介质”的中国专利申请的优先权，其全部内容通过引用结合在本公开中。
技术领域
本公开涉及网络优化领域，特别涉及一种网络优化方法及装置、图像处理方法及装置、存储介质。
背景技术
行人重识别旨在学习有辨别性的特征用于行人检索与匹配。通常情况下，图像数据集中的行人姿态多样性、背景多样性等因素，会影响身份特征的提取。目前，相关技术中通过深度神经网络提取分解特征进行身份识别。
发明内容
本公开实施例提供了一种网络优化技术方案。
根据本公开的一方面，提供了一种网络优化方法，所述方法用于优化神经网络，其包括：
获取图像样本组，所述图像样本组包括由相同对象的图像形成的图像对，以及由不同对象的图像形成的图像对；获取所述图像样本组中图像的第一特征和第二特征，并利用图像的第一特征得到第一分类结果，所述第一特征包括身份特征，所述第二特征包括属性特征；对所述图像样本组内图像对执行特征交换处理得到新的图像对，所述特征交换处理为利用图像对内的第一图像的第一特征以及第二图像的第二特征生成新的第一图像，以及利用第一图像的第二特征以及第二图像的第一特征生成新的第二图像；利用预设方式，获得所述第一分类结果的第一损失值、所述新的图像对的第二损失值，以及所述新的图像对的第一特征和第二特征的第三损失值；至少根据所述第一损失值、第二损失值和第三损失值调节神经网络的参数，直至满足预设要求。
在一些可能的实施方式中，所述获取所述图像样本组中图像的第一特征和第二特征，包括：
将所述图像对的两个图像输入至所述神经网络的身份编码网络模块以及属性编码网络模块；利用所述身份编码网络模块获取所述图像对内的两个图像的第一特征，以及利用所述属性编码网络模块获取所述图像对内的两个图像的第二特征。
在一些可能的实施方式中，所述利用预设方式，获得所述第一分类结果的第一损失值、所述新的图像对的第二损失值，以及所述新的图像对的第一特征和第二特征的第三损失值，包括：
获得通过所述身份编码网络模块获取的第一特征的第一分类结果；利用第一预设方式，根据所述第一分类结果和所述图像样本组中图像对应的真实分类结果，获得所述第一损失值。
在一些可能的实施方式中，在将所述图像对的两个图像输入至所述身份编码网络模块之前，所述方法还包括：向所述图像对的两个图像中对象的图像区域加入噪声。
在一些可能的实施方式中，所述对所述图像样本组内图像对执行特征交换处理得到新的图像对，包括：
将所述图像样本组的图像对中图像的第一特征和第二特征输入至所述神经网络的生成网络模块；通过所述生成网络模块对所述图像样本组内图像对执行所述特征交换处理，得到所述新的图像对。
在一些可能的实施方式中，在输入的图像对为相同对象的图像的情况下，对所述图像样本组内图像对执行特征交换处理得到新的图像对，包括：对所述图像对内的图像执 行一次特征交换处理得到所述新的图像对；
所述对所述图像对内的图像执行一次特征交换处理得到所述新的图像对，包括：
利用图像对中的第一图像的第一特征以及第二图像的第二特征生成新的第一图像，以及利用所述第一图像的第二特征以及第二图像的第一特征生成新的第二图像。
在一些可能的实施方式中，在输入的图像对为不同对象的图像的情况下，对所述图像样本组内图像对执行特征交换处理得到新的图像对，包括：对所述图像对内的图像执行两次特征交换处理得到新的图像对；
所述对所述图像对内的图像执行两次特征交换处理得到新的图像对，包括：
利用图像对中的第一图像的第一特征以及第二图像的第二特征生成新的第一中间图像，以及利用所述第一图像的第二特征以及第二图像的第一特征生成新的第二中间图像；利用所述第一中间图像的第一特征以及第二中间图像的第二特征生成新的第一图像，利用所述第一中间图像的第二特征以及第二中间图像的第一特征生成新的第二图像。
在一些可能的实施方式中，所述利用预设方式，获得所述第一分类结果的第一损失值、所述新的图像对的第二损失值，以及所述新的图像对的第一特征和第二特征的第三损失值，包括：利用第二预设方式，获得通过所述生成网络模块获得的新的图像对相对于原始图像对的第二损失值，所述原始图像对与所述新的图像对相对应。
在一些可能的实施方式中，所述利用预设方式，获得所述第一分类结果的第一损失值、所述新的图像对的第二损失值，以及所述新的图像对的第一特征和第二特征的第三损失值，包括：
按照第三预设方式，基于所述新的图像对的第一特征和第二特征以及对应的原始图像对的第一特征和第二特征，得到所述新的图像对的第一特征和第二特征的第三损失值，所述原始图像对与所述新的图像对相对应。
在一些可能的实施方式中，在对所述图像样本组内图像对执行特征交换处理得到新的图像对之后，所述方法还包括：将生成的新的图像对输入至所述神经网络的辨别网络模块，得到表示所述新的图像对的真实程度的标签特征；利用第四预设方式，基于所述标签特征获得所述新的图像对的第四损失值。
在一些可能的实施方式中，所述至少根据所述第一损失值、第二损失值和第三损失值调节神经网络的参数，直至满足预设要求，包括：
利用所述第一损失值、第二损失值、第三损失值以及第四损失值得到所述神经网络的损失值；利用所述神经网络的损失值调节所述神经网络的参数，直至满足预设要求。
在一些可能的实施方式中，所述利用所述第一损失值、第二损失值、第三损失值以及第四损失值得到所述神经网络的损失值，包括：
在输入至所述神经网络的图像样本组为相同对象的图像对时，利用第五预设方式基于所述第一损失值、第二损失值、第三损失值以及第四损失值得到所述神经网络的第一网络损失值；在输入至所述神经网络的图像样本组为不同对象的图像对时，利用第六预设方式基于所述第一损失值、第二损失值、第三损失值以及第四损失值得到所述神经网络的第二网络损失值；基于所述第一网络损失值和第二网络损失值的加和结果得到所述神经网络的损失值。
根据本公开的第二方面，提供了一种图像处理方法，其包括：
接收输入图像；通过神经网络模型识别所述输入图像的第一特征；基于所述第一特征确定所述输入图像中的对象的身份；其中，所述神经网络模型为通过第一方面中任意一项所述的网络优化方法优化处理后得到的网络模型。
根据本公开的第三方面，提供了一种网络优化装置，其包括：
获取模块，其用于获取图像样本组，所述图像样本组包括由相同对象的图像形成的 图像对，以及由不同对象的图像形成的图像对；特征编码网络模块，其用于获取所述图像样本组中图像的第一特征和第二特征；分类模块，其用于根据图像的第一特征得到第一分类结果；生成网络模块，其用于对所述图像样本组内图像对执行特征交换处理得到新的图像对，所述特征交换处理为利用图像对内的第一图像的第一特征以及第二图像的第二特征生成新的第一图像，以及利用第一图像的第二特征以及第二图像的第一特征生成新的第二图像；损失值获取模块，其用于利用预设方式，获得所述第一分类结果的第一损失值、所述新的图像对的第二损失值，以及所述新的图像对的第一特征和第二特征的第三损失值；调节模块，其用于至少根据所述第一损失值、第二损失值和第三损失值调节神经网络的参数，直至满足预设要求。
在一些可能的实施方式中，所述特征编码网络模块包括身份编码网络模块和属性编码网络模块，其中，
所述获取模块还用于将所述图像对的两个图像输入至所述身份编码网络模块以及属性编码网络模块；并且所述身份编码网络模块用于获取所述图像对内的两个图像的第一特征，以及所述属性编码网络模块用于获取所述图像对内的两个图像的第二特征。
在一些可能的实施方式中，所述损失值获取模块还用获得通过所述身份编码网络模块获取的第一特征对应的第一分类结果，并利用第一预设方式，根据所述第一分类结果和所述图像样本组中图像对应的真实分类结果，获得所述第一损失值。
在一些可能的实施方式中，所述装置还包括：预处理模块，其用于在将所述图像对的两个图像输入至所述身份编码网络模块之前，向所述图像对的两个图像中对象的图像区域加入噪声。
在一些可能的实施方式中，所述生成网络模块还用于在输入的图像对为相同对象的图像的情况下，对所述图像对内的图像执行一次特征交换处理得到所述新的图像对；具体用于利用图像对中的第一图像的第一特征以及第二图像的第二特征生成新的第一图像，以及利用所述第一图像的第二特征以及第二图像的第一特征生成新的第二图像。
在一些可能的实施方式中，所述生成网络模块还用于在输入的图像对为不同对象的图像的情况下，对所述图像对内的图像执行两次特征交换处理得到新的图像对；具体用于利用图像对中的第一图像的第一特征以及第二图像的第二特征生成新的第一中间图像，以及利用所述第一图像的第二特征以及第二图像的第一特征生成新的第二中间图像；利用所述第一中间图像的第一特征以及第二中间图像的第二特征生成新的第一图像，利用所述第一中间图像的第二特征以及第二中间图像的第一特征生成新的第二图像。
在一些可能的实施方式中，所述损失值获取模块还用于利用第二预设方式，获得通过所述生成网络模块获得的新的图像对相对于原始图像对的第二损失值，所述原始图像对与所述新的图像对相对应。
在一些可能的实施方式中，所述损失值获取模块还用于按照第三预设方式，基于所述新的图像对的第一特征和第二特征以及对应的原始图像对的第一特征和第二特征，得到所述新的图像对的第一特征和第二特征的第三损失值，所述原始图像对与所述新的图像对相对应。
在一些可能的实施方式中，所述装置还包括：
辨别网络模块，其用于接收所述新的图像对，并得到表示所述新的图像对的真实程度的标签特征；所述损失值获取模块还用于利用第四预设方式，基于所述标签特征获得所述新的图像对的第四损失值。
在一些可能的实施方式中，所述调节模块还用于利用所述第一损失值、第二损失值、第三损失值以及第四损失值得到所述神经网络的损失值；以及利用所述神经网络的损失值调节所述神经网络的参数，直至满足预设要求。
在一些可能的实施方式中，所述调节模块还用于在输入至所述神经网络的图像样本组为相同对象的图像对时，利用第五预设方式基于所述第一损失值、第二损失值、第三损失值以及第四损失值得到所述神经网络的第一网络损失值；在输入至所述神经网络的图像样本组为不同对象的图像对时，利用第六预设方式基于所述第一损失值、第二损失值、第三损失值以及第四损失值得到所述神经网络的第二网络损失值；基于所述第一网络损失值和第二网络损失值的加和结果得到所述神经网络的损失值。
根据本公开的第四方面，提供了一种图像处理装置，其包括：
接收模块，其用于接收输入图像；识别模块，其用于通过神经网络模型识别所述输入图像的第一特征；身份确定模块，其用于基于所述第一特征确定所述输入图像中的对象的身份；其中，所述神经网络模型为通过第一方面中任意一项所述的网络优化方法优化处理后得到的网络模型。
根据本公开的第五方面，提供了一种电子设备，其包括：
处理器；用于存储处理器可执行指令的存储器；其中，所述处理器被配置为：执行第一方面和第二方面中任意一项所述的方法。
根据本公开的第六方面，提供了一种计算机可读存储介质，其上存储有计算机程序指令，所述计算机程序指令被处理器执行时实现第一方面和第二方面中任意一项所述的方法。
根据本公开的一方面，提供了一种计算机程序，其中，所述计算机程序包括计算机可读代码，当所述计算机可读代码在电子设备中运行时，所述电子设备中的处理器执行用于实现上述网络优化方法。
本公开实施例，可以有效地将提取输入图像中的第一特征(身份特征)以及第一特征以外的第二特征，通过交换两张图像的第二特征而形成新的图片，从而可以成功地将与身份相关的特征和与身份无关的特征分离开，其中与身份相关的特征可以有效地用于行人重识别。本公开实施例提供可以在训练和应用阶段都不需要任何除了图像数据集之外的辅助信息，并且能够提供充分的生成监督，且有效的提高识别精度。
应当理解的是，以上的一般描述和后文的细节描述仅是示例性和解释性的，而非限制本公开。
根据下面参考附图对示例性实施例的详细说明，本公开的其它特征及方面将变得清楚。
附图说明
此处的附图被并入说明书中并构成本说明书的一部分，这些附图示出了符合本公开的实施例，并与说明书一起用于说明本公开的技术方案。
图1示出根据本公开实施例的一种网络优化方法的流程图；
图2示出根据本公开实施例的网络优化方法中步骤S200的流程图；
图3示出根据本公开实施例网络优化方法步骤S300的流程图；
图4示出根据本公开实施例网络优化方法步骤S303的流程图；
图5示出根据本公开实施例的图像处理方法中步骤S400的流程图；
图6示出根据本公开实施例利用第一类样本执行网络优化处理的过程示意图；
图7示出根据本公开实施例利用第二类样本执行网络优化处理的过程示意图；
图8示出根据本公开实施例的图像处理方法的流程图；
图9示出根据本公开实施例的一种网络优化装置的框图；
图10示出根据本公开实施例的一种图像处理装置的框图；
图11示出根据本公开实施例的一种电子设备800的框图；
图12示出根据本公开实施例的一种电子设备1900的框图。
具体实施方式
以下将参考附图详细说明本公开的各种示例性实施例、特征和方面。附图中相同的附图标记表示功能相同或相似的元件。尽管在附图中示出了实施例的各种方面，但是除非特别指出，不必按比例绘制附图。
在这里专用的词“示例性”意为“用作例子、实施例或说明性”。这里作为“示例性”所说明的任何实施例不必解释为优于或好于其它实施例。
本文中术语“和/或”，仅仅是一种描述关联对象的关联关系，表示可以存在三种关系，例如，A和/或B，可以表示：单独存在A，同时存在A和B，单独存在B这三种情况。另外，本文中术语“至少一种”表示多种中的任意一种或多种中的至少两种的任意组合，例如，包括A、B、C中的至少一种，可以表示包括从A、B和C构成的集合中选择的任意一个或多个元素。
另外，为了更好地说明本公开，在下文的具体实施方式中给出了众多的具体细节。本领域技术人员应当理解，没有某些具体细节，本公开同样可以实施。在一些实例中，对于本领域技术人员熟知的方法、手段、元件和电路未作详细描述，以便于凸显本公开的主旨。
本公开实施例提供了一种网络优化方法，其可以用于训练神经网络或者其他机器学习网络，例如，可以应用于对目标用户进行人脸识别、身份验证等场景的机器学习网络的训练过程，还可以应用于对鉴别图像真伪等对精度要求比较高的网络的训练过程，本公开不对具体的应用场景进行限制，只要使用本公开提供的网络优化方法实施的过程均在本公开的保护范围内。本公开实施例以神经网络为例进行说明，但不对此进行具体限定。通过本公开实施例的网络优化方法训练后，可以提高网络的对人物对象的识别精度，同时不需要输入图像以外的其他辅助信息，具有简单方便的特点。
本公开实施例提供的网络优化方案可以由终端设备、服务器或其它类型的电子设备执行，其中，终端设备可以为用户设备(User Equipment，UE)、移动设备、用户终端、终端、蜂窝电话、无绳电话、个人数字处理(Personal Digital Assistant，PDA)、手持设备、计算设备、车载设备、可穿戴设备等。在一些可能的实现方式中，该网络优化方法可以通过处理器调用存储器中存储的计算机可读指令的方式来实现。
图1示出根据本公开实施例的一种网络优化方法的流程图，如图1所示，本公开实施例的网络优化方法可以包括：
S100：获取图像样本组，所述图像样本组包括由相同对象的图像形成的图像对，以及由不同对象的图像形成的图像对；
S200：获取所述图像样本组中图像的第一特征和第二特征，并利用图像的第一特征得到第一分类结果，所述第一特征包括身份特征，所述第二特征包括属性特征；
S300：对所述图像样本组内图像执行特征交换处理得到新的图像对，所述特征交换处理为利用图像对内的第一图像的第一特征以及第二图像的第二特征生成新的第一图像，以及利用第一图像的第二特征以及第二图像的第一特征生成新的第二图像；
S400：利用预设方式，获得所述第一分类结果的第一损失值、所述新的图像对的第二损失值，以及所述新的图像对的第一特征和第二特征的第三损失值；
S500：至少根据所述第一损失值、第二损失值和第三损失值，调节神经网络的参数，直至满足预设要求。
本公开实施例中，在通过本公开实施例训练神经网络时，可以首先向神经网络输入图像样本组，该图像样本组作为训练神经网络的样本图像。其中，在本公开实施例中，图像样本组可以包括两类图像样本，第一类样本为由相同对象的不同图像组成的图像对，第二类样本为由不同对象的不同图像组成的图像对，即，在第一类样本中，每个图像对 内的图像是相同对象的不同图像，在第二类样本中，每个图像对内的图像是不同对象的不同图像。其中每个图像对可以包括两个图像，如下述第一图像和第二图像。另外，本公开实施例可以分别利用该两种类型的图像样本对神经网络进行训练。
进一步地，本公开实施例的图像样本组中的至少一个图像可以具有对应的身份标识，该身份标识可以与图像中的对象相对应，用于分辨图像中的人物对象的身份。本公开实施例中，图像样本组中的至少一个图像可以具有与其对应的对象对应的真实分类标签，其可以由矩阵的形式表示，根据该真实分类标签可以比较神经网络模型的分类结果的准确程度，如可以确定相应的损失值。
在一些可能的实施方式中，获取图像样本组的方式可以包括：利用通信组件接收其他电子设备传输的图像样本组，如从服务器、手机、任意的计算机设备等接收图像样本组。其中图像样本组中的至少一个图像可以是通过摄像头采集的视频图像，经编码处理后得到的多个图像对，但不作为本公开的具体限定。
在获取图像样本组后，即可以执行神经网络的具体优化过程。在步骤S200中，首先可以识别每个图像对的第一图像和第二图像的第一特征和第二特征。其中，第一特征可以包括图像中的对象的身份特征，如服饰的颜色、形状、配饰特征等，第二特征可以为第一特征以外的特征，如属性特征，其可以包括人物对象的姿态特征、背景特征、环境特征等。下面对获取第一特征和第二特征方式进行举例说明。
图2示出根据本公开实施例的网络优化方法中步骤S200的流程图，其中，所述获取所述图像样本组中图像的第一特征和第二特征，并利用图像的第一特征得到第一分类结果，包括：
S201：将所述图像对的两个图像输入至所述神经网络的身份编码网络模块以及属性编码网络模块；
S202：利用所述身份编码网络模块获取所述图像对内的两个图像的第一特征，以及利用所述属性编码网络模块获取所述图像对内的两个图像的第二特征；
S203：利用所述神经网络的分类模块得到所述第一特征对应的第一分类结果。
其中，本公开实施例的神经网络可以包括身份编码网络模块和属性编码网络模块，身份编码网络模块可以用于识别图像中对象的身份特征，属性编码网络模块可以用于识别图像中对象的属性特征。因此，可以将获取的图像样本组中的至少一个图像对分别输入至上述身份编码网络模块和属性编码网络模块。通过身份编码网络模块可以获取接收的图像对内的两个图像的第一特征，以及利用属性编码网络模块获取接收的图像对内的两个图像的第二特征。例如输入的图像对中的两个图像分别用A和B表示，则通过身份编码网络模块获得A的第一特征为A u，通过身份编码网络模块获得B的第一特征为B u，通过属性编码网络模块获得A的第二特征为A v，以及通过属性编码网络模块获得B的第二特征为B v。
其中，身份编码网络模块可以采用预设的人物特征提取算法提取图像中的第一特征，或者也可以包括卷积模块、池化模块等模块单元，以执行第一特征的获取，对于身份编码网络模块的结构，本公开实施例对此不进行具体限定，只要能够提取图像中的第一特征，即可以作为本公开实施例的身份编码网络模块。
同样的，属性编码网络模块也可以采用预设的姿态以及背景特征算法提取图像中的第二特征，或者也可以包括卷积模块等模块单元，对于属性编码网络模块的结构，本公开实施例对此不进行具体限定，只要能够提取图像中的第二特征，即可以作为本公开实施例的属性编码网络模块。
在提取了图像对中两个图像的第一特征和第二特征之后，本公开实施例可以执行利用第一特征进行分类识别的操作，同时还可以执行后续的特征交换处理。
本公开实施例的神经网络中还可以包括分类模块，身份编码网络模块的输出侧可以与分类模块的输入侧连接，从而接收所述身份编码网络模块输出的第一特征，分类模块可以根据接收到的第一特征得到第一分类结果。其中，所述第一分类结果用于表示与第一特征对应的身份标识的预测结果，其预测结果可以按照矩阵的形式体现，该矩阵的元素为预测对象标识的概率。本公开实施例的分类模块的构成可以自行设定，其可以利用设定分类原则得到与第一特征对应的第一分类结果，其中只要能够执行第一特征的分类即可以作为本公开实施例。在得到第一分类结果之后，可以获得该第一分类结果对应的第一损失值，并可以进一步根据该第一损失值确定神经网络的损失值，反馈调节网络中的参数。
另外，在得到每个图像的第一特征和第二特征之后，可以执行图像对的每两个图像之间的特征交换处理。其中，如上述实施例所述，特征交换处理可以为将图像对中的第一图像的第二特征和第二图像的第二特征进行交换，并基于第一特征和交换后的第二特征得到新的图像。
通过特征交换处理，可以将一个图像的第一特征与另一个图像的第二特征进行组合，形成新的图像，利用该新的图像执行分类，可以有效的实现基于身份特征对人物身份的识别，而降低背景、姿态等属性的影响。
图3示出根据本公开实施例的图像处理方法中步骤S300的流程图，其中，所述对所述图像样本组内图像对执行特征交换处理得到新的图像对，可以包括：
S301：将所述图像样本组的图像对输入至所述神经网络的生成网络模块；
S302：通过所述生成网络模块对所述图像样本组内图像对执行所述特征交换处理，得到所述新的图像对。
本公开实施例中的神经网络还可以包括生成网络模块，该生成网络模块可以对身份编码网络模块和属性编码网络模块得到的第一特征和第二特征执行特征交换处理，并根据交换后的特征得到新的图像。具体的，如上述实施例所述，本公开实施例输入的图像样本组可以包括两种类型的图像样本组。其中，第一类样本中的图像对为相同对象的图像。对于第一类样本的图像对，本公开实施例可以对每个图像对内的图像执行一次特征交换处理。
其中，对于第一类样本，所述对所述图像样本组内图像执行特征交换处理得到新的图像对，可以包括：对所述图像对内的图像执行一次特征交换处理得到所述新的图像对。该过程可以包括：
利用每个图像对中的第一图像的第一特征以及第二图像的第二特征生成新的第一图像，以及利用所述第一图像的第二特征以及第二图像的第一特征生成新的第二图像。
其中，由于第一类样本内的图像对中的两个图像为同一对象的不同图像，因此在执行特征交换处理后得到的新图像仍可以为同一对象的图像。在执行完成特征交换处理之后，可以利用得到的新的图像与对应的原始图像之间的差异，以及新的图像的第一特征和第二特征与对应的原始图像的第一特征和第二特征的差异，确定神经网络的损失值，并可以直接根据生成的新图像执行识别分类，此时可以将生成的新的图像对输入至分类模块，执行分类得到第二分类结果。
例如，第一类样本中的图像对包括图像A和图像B，通过身份编码网络模块可以获得A的第一特征为A u，通过身份编码网络模块获得B的第一特征为B u，通过属性编码网络模块获得A的第二特征为A v，以及通过属性编码网络模块获得B的第二特征为B v。A和B分别为相同对象的第一图像和第二图像，第一图像和第二图像不同。在执行特征交换处理时，可以利用A的第一特征A u和B的第二特征B v得到新的第一图像A′，以及利用B的第一特征B u和A的第二特征A v得到新的第二图像B′。
如上所述，本公开实施例的神经网络可以包括生成网络模块，生成网络模块可以用于根据接收的第一特征和第二特征生成新的图像。例如生成网络模块可以包括至少一个卷积单元，或者也可以包括其他的处理单元，通过生成网络模块可以得到第一特征和第二特征对应的图像。即可以通过生成网络完成上述第二特征的交换以及基于交换后的特征生成图像的过程。
通过上述特征交换处理，可以通过交换两张图像的第二特征而形成新的图片，从而可以成功地将与身份相关的特征和与身份无关的特征分离开，通过该种方式对神经网络进行训练，可以提高神经网络对于身份特征的识别精度。
另外，本公开实施例的图像样本组还可以包括第二类样本组，其内的图像对为不同对象的图像。对于第二类样本内的图像对，本公开实施例可以对每个图像对内的图像执行两次特征交换处理。
针对第二类样本组，图4示出根据本公开实施例网络优化方法步骤S303的流程图，其中所述在输入的图像对为不同对象的图像的情况下，所述对所述图像样本组内图像执行特征交换处理得到新的图像对，可以包括：对所述图像对内的图像执行两次特征交换处理，得到新的图像对，该过程可以包括：
S3031：利用所述第二类样本内每个图像对中的第一图像的第一特征以及第二图像的第二特征生成新的第一中间图像，以及利用所述第一图像的第二特征以及第二图像的第一特征生成新的第二中间图像；
S3032：利用所述第一中间图像的第一特征以及第二中间图像的第二特征生成新的第一图像，利用所述第一中间图像的第二特征以及第二中间图像的第一特征生成新的第二图像。
例如，通过身份编码网络模块获得A的第一特征为A u，通过身份编码网络模块获得B的第一特征为B u，通过属性编码网络模块获得A的第二特征为A v，以及通过属性编码网络模块获得B的第二特征为B v。A和B分别为不同对象的第一图像和第二图像。在执行第一次特征交换处理时，可以利用A的第一特征A u和B的第二特征B v得到新的第一中间图像A′，以及利用B的第一特征B u和A的第二特征A v得到新的第二中间图像B′。对应的，在执行第二次特征交换处理时，可以再次利用身份编码网络模块和属性编码网络模块分别获取第一中间图像A′的第一特征A′ u和第二特征A′ v，以及第二中间图像B′的第一特征B′ u和第二特征B′ v，并进一步利用生成网络执行第一中间图像A′的第二特征A′ v以及第二中间图像B′的第二特征B′ v的交换处理，并利用第一中间图像A′的第一特征A′ u和第二中间图像B′的第二特征B′ v生成新的第一图像A″，以及利用第一中间图像A′的第二特征A′ v和第二中间图像B′的第一特征B′ u生成新的第二图像B″。
通过上述两次特征交换处理，可以通过交换两张图像的第二特征而形成新的图片，与相同身份对象的图像对的训练过程不同之处在于，对于第二类样本，由于第一次特征交换处理后没有直接的像素级监督，因此可以进行第二次特征交换处理，可以生成与原始图像对应的图像，该过程可以成为循环生成过程。
在执行完成特征交换处理之后，可以利用得到的新的图像与对应的原始图像之间的差异，以及新的图像的第一特征和第二特征与对应的原始图像的第一特征和第二特征的差异，同时，该新的图像的第一特征也可以被输入至分类模块执行分类处理得到第二分类结果。其中，对于第一类样本的情况，可以直接得到最终的新的图像的第一特征的第二分类结果，对于第二类样本的情况，除了可以得到最终的新的图像的第一特征的第二分类结果，也可以得到中间图像的第一特征的第二分类结果。本公开实施例可以根据上述第二分类结果以及新的图像与原始图像之间的差异和特征之间的差异对神经网络进行优化。即本公开实施例可以根据神经网络的每个网络模块得到的输出结果的损失值对神 经网络进行反馈调节，直至神经网络的损失值满足预设要求，如低于损失阈值时可以确定为满足该预设要求。本公开实施例的神经网络的损失函数可以与网络模块的损失函数有关，例如可以为多个网络模块的损失函数的加权和，基于此可以利用每个网络模块的损失值得到神经网络的损失值，以此对神经网络每个网络模块的参数进行调整，直至满足损失值低于损失阈值的预设要求，该损失阈值本领域技术人员可以根据需求设定，本公开对此不进行具体限定。
下面，对本公开实施例的反馈调节过程进行详细说明。
其中，在通过身份编码网络模块得到图像的第一特征之后，分类模块可以根据该第一特征得到第一分类结果，本公开实施例可以利用第一预设方式，获得通过所述身份编码网络模块获得的第一特征所得到的第一分类结果的第一损失值。其中，图5示出根据本公开实施例的图像处理方法中步骤S400的流程图，其中获取第一损失值的过程包括：
S401：获得通过所述身份编码网络模块获取的第一特征的第一分类结果；
S402：利用第一预设方式，根据所述第一分类结果和所述图像样本组中图像对应的真实分类结果，获得所述第一损失值。
如上述实施例所述，在步骤S200中，在获得样本中的图像的第一特征时，可以通过分类模块执行该第一特征的分类识别，得到与第一特征对应的第一分类结果，该第一分类结果可以用矩阵的形式表示，其中每个元素表示为与每个身份标签对应的概率，通过将该第一分类结果与真实分类结果进行对比，可以得到第一差值，本公开实施例可以将该第一差值确定为第一损失值。或者在其他实施例中，也可以将第一分类结果和真实分类结果输入至分类模块的第一损失函数中，得到第一损失值，本公开对此不进行具体限定。
本公开实施例中，在通过第一类样本和通过第二类样本对神经网络进行训练时，所采用的损失函数可以相同，或者也可以不同。并且，本公开实施例可以将在通过第一类样本训练得到的神经网络的损失值和通过第二类样本训练得到神经网络的损失值进行加和处理，得到最终的神经网络的损失值，并利用该损失值对网络进行反馈调节处理，其中在反馈调节的过程中可以调节神经网络每个网络模块的参数，也可以只调节其中一部分网络模块的参数，对此不公开不作具体限制。
首先，本公开实施例可以利用第一预设方式获得通过所述身份编码网络模块获得的第一特征所得到的第一分类结果的第一损失值。其中，第一预设方式的表达式可以如公式(1)所示：
Figure PCTCN2019118922-appb-000001
其中，
Figure PCTCN2019118922-appb-000002
表示元素相乘，C(I)表示第一分类结果对应的N维预测特征向量，L为对应的原始图像的真实标签对应的N维特征向量(真实分类结果)，L c为第一损失函数对应的第一损失值，i为大于或者等于1且小于或者等于N的变量。
通过上述方式可以得到分类模块得到第一分类结果的第一损失值。其中，本公开实施例可以根据该第一损失值对身份编码网络模块、属性编码网络模块以及分类模块的参数进行反馈调节，也可以根据该第一损失值和其他网络模块的损失值确定神经网络的整体损失值，对神经网络的至少一个网络模块进行统一的反馈调节，本公开对此不作限定。
其次，本公开实施例还可以对通过生成网络模块生成的新的图像对进行处理，得到新的图像对的第二损失值以及对应特征的第三损失值。其中，确定第二损失值的方式可以包括：利用第二预设方式，获得通过所述生成网络模块获得的新的图像对相对于原始 图像对的第二损失值。
本公开实施例中，通过生成网络可以得到新的图像对，本公开实施例可以根据新的图像对与原始图像对之间的差异，确定第二损失值。
其中，对于第一类样本，所述第二预设方式的表达式可以如公式(2)所示：
Figure PCTCN2019118922-appb-000003
其中，X u和X v分别为原始图像对中的第一图像和第二图像，
Figure PCTCN2019118922-appb-000004
和
Figure PCTCN2019118922-appb-000005
分别为新的图像对中的新的第一图像和新的第二图像，L ir为第二损失函数对应的第二损失值，其中|||| 1表示1范数。
通过上述方式可以得到对于第一类样本，通过生成网络模块生成的新的图像对对应的第二损失值。
其中，对于第二类样本，所述第二预设方式的表达式可以如公式(3)所示：
Figure PCTCN2019118922-appb-000006
其中，X u为原始图像对中的第一图像，
Figure PCTCN2019118922-appb-000007
为新的图像对中的第一图像，Y w为原始图像对中的第二图像，
Figure PCTCN2019118922-appb-000008
为新的图像对中的第二图像。
另外，本公开实施例还可以获得新的第二图像对的特征所对应的第三损失值，其中，可以利用第三预设方式获得所述第三损失值。
其中，第三预设方式的表达式可以如公式(4)所示：
Figure PCTCN2019118922-appb-000009
其中，I Xu表示原始图像对中的第一图像X u的第一特征，
Figure PCTCN2019118922-appb-000010
为新的第一图像
Figure PCTCN2019118922-appb-000011
的第一特征，I Xv表示原始图像对中的第二图像X v的第二特征，
Figure PCTCN2019118922-appb-000012
为新的第二图像
Figure PCTCN2019118922-appb-000013
的第二特征，T为转置运算，L s为第三损失函数对应的损失值，|||| 2表示2范数。
通过上述方式可以得到分类模块得到生成网络模块生成的新的图像对的特征对应的第三损失值。
同样的，本公开实施例可以分别根据第二损失值以及第三损失值对生成网络模块的参数进行反馈调节，也可以结合第一损失值对神经网络的多个网络模块同时进行反馈调节。例如在本公开的一些可能的实施方式中，可以分别利用上述第一损失值、第二损失值以及第三损失值的加权和得到神经网络的损失值，也就是说，神经网络的损失函数为上述第一损失函数、第二损失函数和第三损失函数的加权和，每个损失函数的权值本公开不进行具体限定，本领域技术人员可以根据需求自行设定。在得到的损失值大于损失阈值时，反馈调节多个网络模块的参数，直至损失值小于损失阈值，即可以终止训练，此时神经网络优化完成。而且，在本公开实施例中，在基于第一类样本的图像对进行训练时的第一损失函数、第二损失函数以及第三损失函数，与基于第二类本的图像对进行 时的第一损失函数、第二损失函数以及第三损失函数可以不同，但不作为本公开的具体限定。
另外，为了加强本公开实施例的神经网络的训练精度，本公开实施例的神经网络还可以包括辨别网络模块，该辨别网络模块可以用于辨别生成的新的图像对的真假程度(真实程度的标签特征)，可以根据该真假程度获得辨别网络模块确定的生成与新的图像对对应的第四损失值。本公开实施例的辨别网络和生成网络可以构成生成对抗网络。对于该生成对抗网络的具体结构本领域技术人员可以根据现有技术手段选择合适的结构进行配置，本公开对此不进行具体限定。其中，本公开实施例可以将生成的新的图像对输入至所述神经网络的辨别网络模块，并利用第四预设方式获得所述新的图像对的第四损失值。
其中，第四预设方式的表达式如公式(5)所示：
Figure PCTCN2019118922-appb-000014
其中，D表示辨别网络模块的模型函数，E[]表示期望，X表示新的图像对应的原始图像，即真实图像，
Figure PCTCN2019118922-appb-000015
表示输入至辨别网络模块的新的图像，以及D(X)表示辨别网络模块对真实图像的标签特征，
Figure PCTCN2019118922-appb-000016
表示辨别网络模块对输入的新的图像的标签特征。其中，
Figure PCTCN2019118922-appb-000017
中的元素为零至一之间的数值，越接近1,表示该元素真实程度越高。
在本公开实施例中，可以单独执行辨别网络模块的训练过程，即可以向辨别网络模块输入生成的新的图像以及对应的真实图像，并基于上述第四损失函数得到对辨别网络模块进行训练，直到第四损失函数对应的损失值低于训练要求的损失阈值。
在另一些可能的实施例中，也可以将辨别网络模块与前述的身份编码网络模块、属性编码网络模块以及生成网络模块同时进行训练，对应的，本公开实施例步骤S400也可以利用上述第一损失值、第二损失值、第三损失值和第四损失值获得神经网络的损失值。也就是说，神经网络的损失函数为上述第一损失函数、第二损失函数、第三损失函数以及第四损失函数的加权和，每个损失函数的权值本公开不进行具体限定，本领域技术人员可以根据需求自行设定。在得到的损失值大于损失阈值时，反馈调节神经网络多个网络模块的参数，直至损失值小于损失阈值，即可以终止训练，此时神经网络优化完成。
而且，在本公开实施例中，在基于第一类样本的图像对进行训练时的第一损失函数、第二损失函数以及第三损失函数，与基于第二赝本组的图像对进行时的第一损失函数、第二损失函数以及第三损失函数可以不同，但不作为本公开的具体限定。
在本公开的一些可能的实施例中，在输入至身份编码网络模块和属性编码网络模块的图像样本组为相同对象的图像对时(第一类样本)，利用第五预设方式基于所述第一损失值、第二损失值、第三损失值以及第四损失值得到所述神经网络的第一网络损失值，其中，第五预设方式的表达式如公式(6)所示：
L intra＝L c+λ irL ir+λ sL s+λ advL adv  公式(6)；
其中，λ ir、λ s和λ adv分别为第二损失函数、第三损失函数以及第四损失函数的权重，L intra为第一网络损失值。
在输入至身份编码网络模块和属性编码网络模块的图像样本组为不同对象的图像对时，利用第六预设方式基于所述第一损失值、第二损失值、第三损失值以及第四损失值得到所述神经网络的第二网络损失值。其中，第六预设方式的表达式如公式(7)所示：
L inter＝L c+λ crL cr+λ sL s+λ advL adv  公式(7)；
其中，λ cr、λ s和λ adv分别为第二损失函数、第三损失函数以及第四损失函数的权重，L inter为第二网络损失值。
本公开实施例可以根据第一网络损失值以及第二网络损失值的加和结果，得到神经网络的损失值，即神经网络的损失值L＝L intra+L inter。训练过程中，在得到的损失值大于损失阈值时，反馈调节神经网络的参数，例如可以反馈调节多个网络模块(身份编码网络模块、属性编码网络模块、生成网络模块以及辨别网络模块等)的参数，直至神经网络的损失值小于损失阈值，即可以终止训练，此时神经网络优化完成。或者，在其他实施例中，也可以根据第一损失值调节身份编码网络模块、属性编码网络模块以及分类模块的参数，以及根据第二损失值和第三损失值反馈调节生成网络模块的参数，根据第四损失值调节辨别网络模块的参数，直至损失值小于相应损失函数的损失阈值，即终止训练。即本公开实施例可以分别对任意一个网络模块单独进行反馈调节和训练，也可以通过神经网络的损失值，对神经网络部分或全部网络模块进行统一调节，本领域技术人员可以根据需求选择适当的方式执行该调节过程。
另外，在本公开实施例中，为了提高神经网络的身份特征的识别精度，在将所述图像样本组输入至身份编码网络模块之前还可以在图像中添加噪声，例如向所述图像对的两个图像中对象的图像区域加入噪声。本公开实施例中，通过在人物对象的图像区域的部分区域内加入覆盖层的方式加入噪声，所述覆盖层的大小本领域技术人员可以根据需求自行设定，本公开对此不进行限定。在此需要说明的是，本公开实施例仅在输入至身份编码网络模块的图像加入噪声，对其他网络模块不引入噪声。通过该方式可以有效的提高神经网络的身份识别的精度。
为了更加清楚的说明本公开实施例，下面举例对第一类样本和第二类样本的训练过程进行说明。
图6示出根据本公开实施例利用第一类样本执行网络优化处理的过程示意图，其中可以将针对同一对象的两个图像Xu和Xv输入至身份编码网络Eid获得第一特征，以及将图像Xu和Xv输入至属性编码网络Ea获得第二特征，将第一特征输入至分类器C得到第一分类结果，并得到第一损失值Lc。其中，输入至身份编码网络Eid的图像可以被加入噪声，例如在人物对象的区域部分加入覆盖图，以遮挡部分区域。
对图像对中的第二特征交换，并利用生成网络模块G得到交换处理后的两个新的图像，此时可以获得两个新的图像对应的第二损失值Lir，以及两个新的图像对应的第一特征和第二特征对应的第三损失值Ls，将新的图像输入至辨别网络模块D，得到第四损失值Ladv。此时可以利用第一损失值Lc、第二损失值Lir、第三损失值Ls和第四损失值Ladv得到神经网络的损失值，在该损失值小于损失阈值时终止训练，否则反馈调节神经网络的至少一个网络模块的参数。
图7示出根据本公开实施例利用第二类样本执行网络优化处理的过程示意图，其中可以将针对不同对象的两个图像Xu和Yw输入至身份编码网络Eid获得第一特征，以及将图像Xu和Yw输入至属性编码网络Ea获得第二特征，将第一特征输入至分类器C得到第一分类结果，并得到第一损失值Lc。其中，输入至身份编码网络Eid的图像可以被加入噪声，例如在人物对象的区域部分加入覆盖图，以遮挡部分区域。
对图像对中的第二特征交换，并利用生成器得到交换处理后的两个中间图像，进一步利用身份编码网络模块Eid和属性编码网络模块Ea得到这两个中间图像的第一特征和第二特征，继而交换中间图像的第二特征得到新的图像。此时可以获得两个新的图像对 应的第二损失值Ls，以及两个新的图像对应的第一特征和第二特征对应的第三损失值Lcr，将中间图像或者新的图像输入至辨别网络模块D，得到第四损失值Ladv。此时可以利用第一损失值、第二损失值、第三损失值和第四损失值得到神经网络的损失值，在该损失值小于损失阈值时终止训练，否则反馈调节神经网络的至少一个网络模块的参数。
本公开实施例，可以有效地将提取输入图像中的第一特征(身份特征)以及第一特征以外的第二特征，通过交换两张图像的第二特征而形成新的图片，从而可以成功地将与身份相关的特征和与身份无关的特征分离开，其中与身份相关的特征可以有效地用于行人重识别。本公开实施例提供可以在训练和应用阶段都不需要任何除了图像数据集之外的辅助信息，并且能够提供充分的生成监督，且有效的提高识别精度。
本领域技术人员可以理解，在具体实施方式的上述方法中，各步骤的撰写顺序并不意味着严格的执行顺序而对实施过程构成任何限定，各步骤的具体执行顺序应当以其功能和可能的内在逻辑确定。
另外，本公开实施例还提供了一种图像处理方法，该方法可以应用由第一方面提供的图像优化方法得到的神经网络执行图像识别的操作，获得与输入图像对应身份的识别结果。
图8示出根据本公开实施例的图像处理方法的流程图，其中所述方法包括：
S10：接收输入图像；
S20：通过神经网络模型识别所述输入图像的第一特征；
S30：基于所述第一特征确定所述输入图像中的对象的身份；其中，所述神经网络模型为通过第一方面中任意一项所述的网络优化方法优化处理后得到的网络模型。
本公开实施例中，可以由第一方面训练得到满足要求的神经网络模型，并利用该神经网络模型执行图像中对象的身份识别的操作，即可以利用神经网络模型形成能够执行图像识别等操作的图像处理装置，利用该装置可以执行上述身份识别过程。
本公开实施例中，可以包括数据库，该数据库内可以包括多个人员对象的信息，例如人员对象的图像、以及对应的身份信息，如姓名、年龄、职位等信息，本公开对此不进行限定。
在接入到输入图像后，本公开实施例可以将接收的输入图像的第一特征与数据库中的人员对象的图像进行对比，判定数据库中与之匹配的人员对象。由于本公开实施例的神经网络模型经上述实施例进行训练，并满足精度要求，从而本公开实施例可以精确的匹配出与输入图像匹配的对象，并进而获得其相应的身份信息。
本公开实施例的图像处理方法能够快速的识别图像对象的身份，并能够提高识别精度。
可以理解，本公开提及的上述各个方法实施例，在不违背原理逻辑的情况下，均可以彼此相互结合形成结合后的实施例，限于篇幅，本公开不再赘述。
此外，本公开还提供了图像处理装置、电子设备、计算机可读存储介质、程序，上述均可用来实现本公开提供的任一种图像处理方法，相应技术方案和描述和参见方法部分的相应记载，不再赘述。
图9示出根据本公开实施例的网络优化装置的框图，如图9所示，所述网络优化装置包括：
获取模块10，其用于获取图像样本组，所述图像样本组包括由相同对象的图像形成的图像对，以及由不同对象的图像形成的图像对；特征编码网络模块20，其用于获取所述图像样本组中图像的第一特征和第二特征；分类模块30，其用于根据图像的第一特征得到第一分类结果；生成网络模块40，其用于对所述图像样本组内图像对执行特征交换处理得到新的图像对，所述特征交换处理为利用图像对内的第一图像的第一特征以及第 二图像的第二特征生成新的第一图像，以及利用第一图像的第二特征以及第二图像的第一特征生成新的第二图像；损失值获取模块50，其用于利用预设方式，获得所述第一分类结果的第一损失值、所述新的图像对的第二损失值，以及所述新的图像对的第一特征和第二特征的第三损失值；调节模块60，其用于至少根据所述第一损失值、第二损失值和第三损失值调节神经网络的参数，直至满足预设要求。
在一些可能的实施方式中，所述特征编码网络模块包括身份编码网络模块和属性编码网络模块，其中，所述获取模块还用于将所述图像对的两个图像输入至所述身份编码网络模块以及属性编码网络模块；并且所述身份编码网络模块用于获取所述图像对内的两个图像的第一特征，以及所述属性编码网络模块用于获取所述图像对内的两个图像的第二特征。
在一些可能的实施方式中，所述损失值获取模块还用获得通过所述身份编码网络模块获取的第一特征对应的第一分类结果，并利用第一预设方式，根据所述第一分类结果和所述图像样本组中图像对应的真实分类结果，获得所述第一损失值。
在一些可能的实施方式中，所述装置还包括：预处理模块，其用于在将所述图像对的两个图像输入至所述身份编码网络模块之前，向所述图像对的两个图像中对象的图像区域加入噪声。
在一些可能的实施方式中，所述生成网络模块还用于在输入的图像对为相同对象的图像的情况下，对所述图像对内的图像执行一次特征交换处理得到所述新的图像对，其包括：利用图像对中的第一图像的第一特征以及第二图像的第二特征生成新的第一图像，以及利用所述第一图像的第二特征以及第二图像的第一特征生成新的第二图像。
在一些可能的实施方式中，所述生成网络模块还用于在输入的图像对为不同对象的图像的情况下，对所述图像对内的图像执行两次特征交换处理得到新的图像对，其包括：利用图像对中的第一图像的第一特征以及第二图像的第二特征生成新的第一中间图像，以及利用所述第一图像的第二特征以及第二图像的第一特征生成新的第二中间图像；利用所述第一中间图像的第一特征以及第二中间图像的第二特征生成新的第一图像，利用所述第一中间图像的第二特征以及第二中间图像的第一特征生成新的第二图像。
在一些可能的实施方式中，所述损失值获取模块还用于利用第二预设方式，获得通过所述生成网络模块获得的新的图像对相对于原始图像对的第二损失值，所述原始图像对与所述新的图像对相对应。
在一些可能的实施方式中，所述损失值获取模块还用于按照第三预设方式，基于所述新的图像对的第一特征和第二特征以及对应的原始图像对的第一特征和第二特征，得到所述新的图像对的第一特征和第二特征的第三损失值，所述原始图像对与所述新的图像对相对应。
在一些可能的实施方式中，所述装置还包括：辨别网络模块，其用于接收所述新的图像对，并得到表示所述新的图像对的真实程度的标签特征；所述损失值获取模块还用于利用第四预设方式，基于所述标签特征获得所述新的图像对的第四损失值。
在一些可能的实施方式中，所述调节模块还用于利用所述第一损失值、第二损失值、第三损失值以及第四损失值得到所述神经网络的损失值；以及利用所述神经网络的损失值调节所述神经网络的参数，直至满足预设要求。
在一些可能的实施方式中，所述调节模块还用于在输入至所述神经网络的图像样本组为相同对象的图像对时，利用第五预设方式基于所述第一损失值、第二损失值、第三损失值以及第四损失值得到所述神经网络的第一网络损失值；在输入至所述神经网络的图像样本组为不同对象的图像对时，利用第六预设方式基于所述第一损失值、第二损失值、第三损失值以及第四损失值得到所述神经网络的第二网络损失值；基于所述第一网 络损失值和第二网络损失值的加和结果得到所述神经网络的损失值。
图10示出根据本公开实施例的一种图像处理装置的框图。如图所示，图像处理装置可以包括：
接收模块100，其用于接收输入图像；
识别模块200，其用于通过神经网络模型识别所述输入图像的第一特征；
身份确定模块300，其用于基于所述第一特征确定所述输入图像中的对象的身份；
其中，所述神经网络模型为通过第一方面中任意一项所述的网络优化方法优化处理后得到的网络模型。
在一些实施例中，本公开实施例提供的装置具有的功能或包含的模块可以用于执行上文方法实施例描述的方法，其具体实现可以参照上文方法实施例的描述，为了简洁，这里不再赘述。
本公开实施例还提出一种计算机可读存储介质，其上存储有计算机程序指令，所述计算机程序指令被处理器执行时实现上述方法。计算机可读存储介质可以是非易失性计算机可读存储介质。
本公开实施例还提出一种电子设备，包括：处理器；用于存储处理器可执行指令的存储器；其中，所述处理器被配置为上述方法。
本公开实施例还提供了一种计算机程序产品，包括计算机可读代码，当计算机可读代码在设备上运行时，设备中的处理器执行用于实现如上任一实施例提供的上述方法的指令。
电子设备可以被提供为终端、服务器或其它形态的设备。
图11示出根据本公开实施例的一种电子设备800的框图。例如，电子设备800可以是移动电话，计算机，数字广播终端，消息收发设备，游戏控制台，平板设备，医疗设备，健身设备，个人数字助理等终端。
参照图11，电子设备800可以包括以下一个或多个组件：处理组件802，存储器804，电源组件806，多媒体组件808，音频组件810，输入/输出(I/O)的接口812，传感器组件814，以及通信组件816。
处理组件802通常控制电子设备800的整体操作，诸如与显示，电话呼叫，数据通信，相机操作和记录操作相关联的操作。处理组件802可以包括一个或多个处理器820来执行指令，以完成上述的方法的全部或部分步骤。此外，处理组件802可以包括一个或多个模块，便于处理组件802和其他组件之间的交互。例如，处理组件802可以包括多媒体模块，以方便多媒体组件808和处理组件802之间的交互。
存储器804被配置为存储各种类型的数据以支持在电子设备800的操作。这些数据的示例包括用于在电子设备800上操作的任何应用程序或方法的指令，联系人数据，电话簿数据，消息，图片，视频等。存储器804可以由任何类型的易失性或非易失性存储设备或者它们的组合实现，如静态随机存取存储器(SRAM)，电可擦除可编程只读存储器(EEPROM)，可擦除可编程只读存储器(EPROM)，可编程只读存储器(PROM)，只读存储器(ROM)，磁存储器，快闪存储器，磁盘或光盘。
电源组件806为电子设备800的各种组件提供电力。电源组件806可以包括电源管理系统，一个或多个电源，及其他与为电子设备800生成、管理和分配电力相关联的组件。
多媒体组件808包括在所述电子设备800和用户之间的提供一个输出接口的屏幕。在一些实施例中，屏幕可以包括液晶显示器(LCD)和触摸面板(TP)。如果屏幕包括触摸面板，屏幕可以被实现为触摸屏，以接收来自用户的输入信号。触摸面板包括一个或多个触摸传感器以感测触摸、滑动和触摸面板上的手势。所述触摸传感器可以不仅感测触摸或滑动动作的边界，而且还检测与所述触摸或滑动操作相关的持续时间和压力。在一 些实施例中，多媒体组件808包括一个前置摄像头和/或后置摄像头。当电子设备800处于操作模式，如拍摄模式或视频模式时，前置摄像头和/或后置摄像头可以接收外部的多媒体数据。每个前置摄像头和后置摄像头可以是一个固定的光学透镜系统或具有焦距和光学变焦能力。
音频组件810被配置为输出和/或输入音频信号。例如，音频组件810包括一个麦克风(MIC)，当电子设备800处于操作模式，如呼叫模式、记录模式和语音识别模式时，麦克风被配置为接收外部音频信号。所接收的音频信号可以被进一步存储在存储器804或经由通信组件816发送。在一些实施例中，音频组件810还包括一个扬声器，用于输出音频信号。
I/O接口812为处理组件802和外围接口模块之间提供接口，上述外围接口模块可以是键盘，点击轮，按钮等。这些按钮可包括但不限于：主页按钮、音量按钮、启动按钮和锁定按钮。
传感器组件814包括一个或多个传感器，用于为电子设备800提供各个方面的状态评估。例如，传感器组件814可以检测到电子设备800的打开/关闭状态，组件的相对定位，例如所述组件为电子设备800的显示器和小键盘，传感器组件814还可以检测电子设备800或电子设备800一个组件的位置改变，用户与电子设备800接触的存在或不存在，电子设备800方位或加速/减速和电子设备800的温度变化。传感器组件814可以包括接近传感器，被配置用来在没有任何的物理接触时检测附近物体的存在。传感器组件814还可以包括光传感器，如CMOS或CCD图像传感器，用于在成像应用中使用。在一些实施例中，该传感器组件814还可以包括加速度传感器，陀螺仪传感器，磁传感器，压力传感器或温度传感器。
通信组件816被配置为便于电子设备800和其他设备之间有线或无线方式的通信。电子设备800可以接入基于通信标准的无线网络，如WiFi，2G或3G，或它们的组合。在一个示例性实施例中，通信组件816经由广播信道接收来自外部广播管理系统的广播信号或广播相关信息。在一个示例性实施例中，所述通信组件816还包括近场通信(NFC)模块，以促进短程通信。例如，在NFC模块可基于射频识别(RFID)技术，红外数据协会(IrDA)技术，超宽带(UWB)技术，蓝牙(BT)技术和其他技术来实现。
在示例性实施例中，电子设备800可以被一个或多个应用专用集成电路(ASIC)、数字信号处理器(DSP)、数字信号处理设备(DSPD)、可编程逻辑器件(PLD)、现场可编程门阵列(FPGA)、控制器、微控制器、微处理器或其他电子元件实现，用于执行上述方法。
在示例性实施例中，还提供了一种非易失性计算机可读存储介质，例如包括计算机程序指令的存储器804，上述计算机程序指令可由电子设备800的处理器820执行以完成上述方法。
图12是根据一示例性实施例示出的一种电子设备1900的框图。例如，电子设备1900可以被提供为一服务器。参照图12，电子设备1900包括处理组件1922，其进一步包括一个或多个处理器，以及由存储器1932所代表的存储器资源，用于存储可由处理组件1922的执行的指令，例如应用程序。存储器1932中存储的应用程序可以包括一个或一个以上的每一个对应于一组指令的模块。此外，处理组件1922被配置为执行指令，以执行上述方法。
电子设备1900还可以包括一个电源组件1926被配置为执行电子设备1900的电源管理，一个有线或无线网络接口1950被配置为将电子设备1900连接到网络，和一个输入输出(I/O)接口1958。电子设备1900可以操作基于存储在存储器1932的操作系统，例如Windows ServerTM，Mac OS XTM，UnixTM,LinuxTM，FreeBSDTM或类似。
在示例性实施例中，还提供了一种非易失性计算机可读存储介质，例如包括计算机程序指令的存储器1932，上述计算机程序指令可由电子设备1900的处理组件1922执行以完成上述方法。
本公开可以是系统、方法和/或计算机程序产品。计算机程序产品可以包括计算机可读存储介质，其上载有用于使处理器实现本公开的各个方面的计算机可读程序指令。
计算机可读存储介质可以是可以保持和存储由指令执行设备使用的指令的有形设备。计算机可读存储介质例如可以是――但不限于――电存储设备、磁存储设备、光存储设备、电磁存储设备、半导体存储设备或者上述的任意合适的组合。计算机可读存储介质的更具体的例子(非穷举的列表)包括：便携式计算机盘、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)、静态随机存取存储器(SRAM)、便携式压缩盘只读存储器(CD-ROM)、数字多功能盘(DVD)、记忆棒、软盘、机械编码设备、例如其上存储有指令的打孔卡或凹槽内凸起结构、以及上述的任意合适的组合。这里所使用的计算机可读存储介质不被解释为瞬时信号本身，诸如无线电波或者其他自由传播的电磁波、通过波导或其他传输媒介传播的电磁波(例如，通过光纤电缆的光脉冲)、或者通过电线传输的电信号。
这里所描述的计算机可读程序指令可以从计算机可读存储介质下载到各个计算/处理设备，或者通过网络、例如因特网、局域网、广域网和/或无线网下载到外部计算机或外部存储设备。网络可以包括铜传输电缆、光纤传输、无线传输、路由器、防火墙、交换机、网关计算机和/或边缘服务器。每个计算/处理设备中的网络适配卡或者网络接口从网络接收计算机可读程序指令，并转发该计算机可读程序指令，以供存储在各个计算/处理设备中的计算机可读存储介质中。
用于执行本公开操作的计算机程序指令可以是汇编指令、指令集架构(ISA)指令、机器指令、机器相关指令、微代码、固件指令、状态设置数据、或者以一种或多种编程语言的任意组合编写的源代码或目标代码，所述编程语言包括面向对象的编程语言—诸如Smalltalk、C++等，以及常规的过程式编程语言—诸如“C”语言或类似的编程语言。计算机可读程序指令可以完全地在用户计算机上执行、部分地在用户计算机上执行、作为一个独立的软件包执行、部分在用户计算机上部分在远程计算机上执行、或者完全在远程计算机或服务器上执行。在涉及远程计算机的情形中，远程计算机可以通过任意种类的网络—包括局域网(LAN)或广域网(WAN)—连接到用户计算机，或者，可以连接到外部计算机(例如利用因特网服务提供商来通过因特网连接)。在一些实施例中，通过利用计算机可读程序指令的状态信息来个性化定制电子电路，例如可编程逻辑电路、现场可编程门阵列(FPGA)或可编程逻辑阵列(PLA)，该电子电路可以执行计算机可读程序指令，从而实现本公开的各个方面。
这里参照根据本公开实施例的方法、装置(系统)和计算机程序产品的流程图和/或框图描述了本公开的各个方面。应当理解，流程图和/或框图的每个方框以及流程图和/或框图中各方框的组合，都可以由计算机可读程序指令实现。
这些计算机可读程序指令可以提供给通用计算机、专用计算机或其它可编程数据处理装置的处理器，从而生产出一种机器，使得这些指令在通过计算机或其它可编程数据处理装置的处理器执行时，产生了实现流程图和/或框图中的一个或多个方框中规定的功能/动作的装置。也可以把这些计算机可读程序指令存储在计算机可读存储介质中，这些指令使得计算机、可编程数据处理装置和/或其他设备以特定方式工作，从而，存储有指令的计算机可读介质则包括一个制造品，其包括实现流程图和/或框图中的一个或多个方框中规定的功能/动作的各个方面的指令。
也可以把计算机可读程序指令加载到计算机、其它可编程数据处理装置、或其它设 备上，使得在计算机、其它可编程数据处理装置或其它设备上执行一系列操作步骤，以产生计算机实现的过程，从而使得在计算机、其它可编程数据处理装置、或其它设备上执行的指令实现流程图和/或框图中的一个或多个方框中规定的功能/动作。
附图中的流程图和框图显示了根据本公开的多个实施例的系统、方法和计算机程序产品的可能实现的体系架构、功能和操作。在这点上，流程图或框图中的每个方框可以代表一个模块、程序段或指令的一部分，所述模块、程序段或指令的一部分包含一个或多个用于实现规定的逻辑功能的可执行指令。在有些作为替换的实现中，方框中所标注的功能也可以以不同于附图中所标注的顺序发生。例如，两个连续的方框实际上可以基本并行地执行，它们有时也可以按相反的顺序执行，这依所涉及的功能而定。也要注意的是，框图和/或流程图中的每个方框、以及框图和/或流程图中的方框的组合，可以用执行规定的功能或动作的专用的基于硬件的系统来实现，或者可以用专用硬件与计算机指令的组合来实现。
以上已经描述了本公开的各实施例，上述说明是示例性的，并非穷尽性的，并且也不限于所披露的各实施例。在不偏离所说明的各实施例的范围和精神的情况下，对于本技术领域的普通技术人员来说许多修改和变更都是显而易见的。本文中所用术语的选择，旨在最好地解释各实施例的原理、实际应用或对市场中技术的技术改进，或者使本技术领域的其它普通技术人员能理解本文披露的各实施例。

Claims (28)

  1. 一种网络优化方法，所述方法用于优化神经网络，其特征在于，包括：
    获取图像样本组，所述图像样本组包括由相同对象的图像形成的图像对，以及由不同对象的图像形成的图像对；
    获取所述图像样本组中图像的第一特征和第二特征，并利用图像的第一特征得到第一分类结果，所述第一特征包括身份特征，所述第二特征包括属性特征；
    对所述图像样本组内图像对执行特征交换处理得到新的图像对，所述特征交换处理为利用所述图像对内的第一图像的第一特征以及第二图像的第二特征生成新的第一图像，以及利用所述第一图像的第二特征以及所述第二图像的第一特征生成新的第二图像；
    利用预设方式，获得所述第一分类结果的第一损失值、所述新的图像对的第二损失值，以及所述新的图像对的第一特征和第二特征的第三损失值；
    至少根据所述第一损失值、所述第二损失值和所述第三损失值调节神经网络的参数，直至满足预设要求。
  2. 根据权利要求1所述的方法，其特征在于，所述获取所述图像样本组中图像的第一特征和第二特征，包括：
    将所述图像对的两个图像输入至所述神经网络的身份编码网络模块以及属性编码网络模块；
    利用所述身份编码网络模块获取所述图像对内的两个图像的第一特征，以及利用所述属性编码网络模块获取所述图像对内的两个图像的第二特征。
  3. 根据权利要求2所述的方法，其特征在于，所述利用预设方式，获得所述第一分类结果的第一损失值、所述新的图像对的第二损失值，以及所述新的图像对的第一特征和第二特征的第三损失值，包括：
    获得通过所述身份编码网络模块获取的所述第一特征的所述第一分类结果；
    利用第一预设方式，根据所述第一分类结果和所述图像样本组中图像对应的真实分类结果，获得所述第一损失值。
  4. 根据权利要求2或3所述的方法，其特征在于，在将所述图像对的两个图像输入至所述身份编码网络模块之前，所述方法还包括：
    向所述图像对的两个图像中对象的图像区域加入噪声。
  5. 根据权利要求1-4中任意一项所述的方法，其特征在于，所述对所述图像样本组内图像对执行特征交换处理得到新的图像对，包括：
    将所述图像样本组的图像对中图像的第一特征和第二特征输入至所述神经网络的生成网络模块；
    通过所述生成网络模块对所述图像样本组内图像对执行所述特征交换处理，得到所述新的图像对。
  6. 根据权利要求1-5中任意一项所述的方法，其特征在于，在输入的图像对为相同对象的图像的情况下，对所述图像样本组内图像对执行特征交换处理得到新的图像对，包括：对所述图像对内的图像执行一次特征交换处理得到所述新的图像对；
    所述对所述图像对内的图像执行一次特征交换处理得到所述新的图像对，包括：
    利用所述图像对中的第一图像的第一特征以及第二图像的第二特征生成新的第一图像，以及利用所述第一图像的第二特征以及所述第二图像的第一特征生成新的第二图像。
  7. 根据权利要求1-6中任意一项所述的方法，其特征在于，在输入的图像对为不同对象的图像的情况下，对所述图像样本组内图像对执行特征交换处理得到新的图像对，包括：对所述图像对内的图像执行两次特征交换处理得到新的图像对；
    所述对所述图像对内的图像执行两次特征交换处理得到新的图像对，包括：
    利用所述图像对中的第一图像的第一特征以及第二图像的第二特征生成新的第一中 间图像，以及利用所述第一图像的第二特征以及第二图像的第一特征生成新的第二中间图像；
    利用所述第一中间图像的第一特征以及所述第二中间图像的第二特征生成新的第一图像，利用所述第一中间图像的第二特征以及第二中间图像的第一特征生成新的第二图像。
  8. 根据权利要求5-7中任意一项所述的方法，其特征在于，所述利用预设方式，获得所述第一分类结果的第一损失值、所述新的图像对的第二损失值，以及所述新的图像对的第一特征和第二特征的第三损失值，包括：
    利用第二预设方式，获得通过所述生成网络模块获得的新的图像对相对于原始图像对的第二损失值，所述原始图像对与所述新的图像对相对应。
  9. 根据权利要求1-8中任意一项所述的方法，其特征在于，所述利用预设方式，获得所述第一分类结果的第一损失值、所述新的图像对的第二损失值，以及所述新的图像对的第一特征和第二特征的第三损失值，包括：
    按照第三预设方式，基于所述新的图像对的第一特征和第二特征以及原始图像对的第一特征和第二特征，得到所述新的图像对的第一特征和第二特征的第三损失值，所述原始图像对与所述新的图像对相对应。
  10. 根据权利要求1-9中任意一项所述方法，其特征在于，在对所述图像样本组内图像对执行特征交换处理得到新的图像对之后，所述方法还包括：
    将生成的所述新的图像对输入至所述神经网络的辨别网络模块，得到表示所述新的图像对的真实程度的标签特征；
    利用第四预设方式，基于所述标签特征获得所述新的图像对的第四损失值。
  11. 根据权利要求10所述的方法，其特征在于，所述至少根据所述第一损失值、所述第二损失值和所述第三损失值调节神经网络的参数，直至满足预设要求，包括：
    利用所述第一损失值、所述第二损失值、所述第三损失值以及所述第四损失值得到所述神经网络的损失值；
    利用所述神经网络的损失值调节所述神经网络的参数，直至满足预设要求。
  12. 根据权利要求11所述的方法，其特征在于，所述利用所述第一损失值、所述第二损失值、所述第三损失值以及所述第四损失值得到所述神经网络的损失值，包括：
    在输入至所述神经网络的图像样本组为相同对象的图像对时，利用第五预设方式基于所述第一损失值、所述第二损失值、所述第三损失值以及所述第四损失值得到所述神经网络的第一网络损失值；
    在输入至所述神经网络的图像样本组为不同对象的图像对时，利用第六预设方式基于所述第一损失值、所述第二损失值、所述第三损失值以及所述第四损失值得到所述神经网络的第二网络损失值；
    基于所述第一网络损失值和所述第二网络损失值的加和结果得到所述神经网络的损失值。
  13. 一种图像处理方法，其特征在于，包括：
    接收输入图像；
    通过神经网络模型识别所述输入图像的第一特征；
    基于所述第一特征确定所述输入图像中的对象的身份；
    其中，所述神经网络模型为通过权利要求1-12中任意一项所述的网络优化方法优化处理后得到的网络模型。
  14. 一种网络优化装置，其包括：
    获取模块，其用于获取图像样本组，所述图像样本组包括由相同对象的图像形成的 图像对，以及由不同对象的图像形成的图像对；
    特征编码网络模块，其用于获取所述图像样本组中图像的第一特征和第二特征；
    分类模块，用于根据图像的第一特征得到第一分类结果；
    生成网络模块，用于对所述图像样本组内图像对执行特征交换处理得到新的图像对，所述特征交换处理为利用所述图像对内的第一图像的第一特征以及第二图像的第二特征生成新的第一图像，以及利用所述第一图像的第二特征以及所述第二图像的第一特征生成新的第二图像；
    损失值获取模块，用于利用预设方式，获得所述第一分类结果的第一损失值、所述新的图像对的第二损失值，以及所述新的图像对的第一特征和第二特征的第三损失值；
    调节模块，用于至少根据所述第一损失值、所述第二损失值和所述第三损失值调节神经网络的参数，直至满足预设要求。
  15. 根据权利要求14所述的装置，其特征在于，所述特征编码网络模块包括身份编码网络模块和属性编码网络模块，其中，
    所述获取模块还用于将所述图像对的两个图像输入至所述身份编码网络模块以及属性编码网络模块；并且
    所述身份编码网络模块用于获取所述图像对内的两个图像的第一特征，以及所述属性编码网络模块用于获取所述图像对内的两个图像的第二特征。
  16. 根据权利要求15所述的装置，其特征在于，所述损失值获取模块还用获得通过所述身份编码网络模块获取的所述第一特征对应的所述第一分类结果，并利用第一预设方式，根据所述第一分类结果和所述图像样本组中图像对应的真实分类结果，获得所述第一损失值。
  17. 根据权利要求15或16所述的装置，其特征在于，所述装置还包括：
    预处理模块，用于在将所述图像对的两个图像输入至所述身份编码网络模块之前，向所述图像对的两个图像中对象的图像区域加入噪声。
  18. 根据权利要求14-17中任意一项所述的装置，其特征在于，所述生成网络模块还用于在输入的图像对为相同对象的图像的情况下，对所述图像对内的图像执行一次特征交换处理得到所述新的图像对；
    所述生成网络模块具体用于利用图像对中的第一图像的第一特征以及第二图像的第二特征生成新的第一图像，以及利用所述第一图像的第二特征以及第二图像的第一特征生成新的第二图像。
  19. 根据权利要求14-18中任意一项所述的装置，其特征在于，所述生成网络模块还用于在输入的图像对为不同对象的图像的情况下，对所述图像对内的图像执行两次特征交换处理得到新的图像对；
    所述生成网络模块具体用于利用所述图像对中的第一图像的第一特征以及第二图像的第二特征生成新的第一中间图像，以及利用所述第一图像的第二特征以及所述第二图像的第一特征生成新的第二中间图像；利用所述第一中间图像的第一特征以及所述第二中间图像的第二特征生成新的第一图像，利用所述第一中间图像的第二特征以及第二中间图像的第一特征生成新的第二图像。
  20. 根据权利要求18-19中任意一项所述的装置，其特征在于，所述损失值获取模块还用于利用第二预设方式，获得通过所述生成网络模块获得的新的图像对相对于原始图像对的第二损失值，所述原始图像对与所述新的图像对相对应。
  21. 根据权利要求14-20中任意一项所述的装置，其特征在于，所述损失值获取模块还用于按照第三预设方式，基于所述新的图像对的第一特征和第二特征以及原始图像对的第一特征和第二特征，得到所述新的图像对的第一特征和第二特征的第三损失值，所 述原始图像对与所述新的图像对相对应。
  22. 根据权利要求14-21中任意一项所述的装置，其特征在于，所述装置还包括：
    辨别网络模块，用于接收所述新的图像对，并得到表示所述新的图像对的真实程度的标签特征；
    所述损失值获取模块，还用于利用第四预设方式，基于所述标签特征获得所述新的图像对的第四损失值。
  23. 根据权利要求22所述的装置，其特征在于，所述调节模块还用于利用所述第一损失值、所述第二损失值、所述第三损失值以及所述第四损失值得到所述神经网络的损失值；以及
    利用所述神经网络的损失值调节所述神经网络的参数，直至满足预设要求。
  24. 根据权利要求23所述的装置，其特征在于，所述调节模块还用于在输入至所述神经网络的图像样本组为相同对象的图像对时，利用第五预设方式基于所述第一损失值、第二损失值、第三损失值以及第四损失值得到所述神经网络的第一网络损失值；
    在输入至所述神经网络的图像样本组为不同对象的图像对时，利用第六预设方式基于所述第一损失值、所述第二损失值、所述第三损失值以及所述第四损失值得到所述神经网络的第二网络损失值；
    基于所述第一网络损失值和所述第二网络损失值的加和结果得到所述神经网络的损失值。
  25. 一种图像处理装置，其特征在于，包括：
    接收模块，用于接收输入图像；
    识别模块，用于通过神经网络模型识别所述输入图像的第一特征；
    身份确定模块，用于基于所述第一特征确定所述输入图像中的对象的身份；
    其中，所述神经网络模型为通过权利要求1-12中任意一项所述的网络优化方法优化处理后得到的网络模型。
  26. 一种图像处理装置，其特征在于，包括：
    处理器；
    用于存储处理器可执行指令的存储器；
    其中，所述处理器被配置为：执行权利要求1至12中任意一项所述的方法，或者执行如权利要求13所述的方法。
  27. 一种计算机可读存储介质，其上存储有计算机程序指令，其特征在于，所述计算机程序指令被处理器执行时实现权利要求1至12中任意一项所述的方法，或者实现权利要求13所述的方法。
  28. 一种计算机程序，其特征在于，所述计算机程序包括计算机可读代码，当所述计算机可读代码在电子设备中运行时，所述电子设备中的处理器执行用于实现1至12中任意一项所述的方法，或者实现权利要求13所述的方法。
PCT/CN2019/118922 2019-01-15 2019-11-15 网络优化方法及装置、图像处理方法及装置、存储介质 Ceased WO2020147414A1 (zh)

Priority Applications (4)

Application Number Priority Date Filing Date Title
KR1020207033113A KR102454515B1 (ko) 2019-01-15 2019-11-15 네트워크 최적화 방법 및 장치, 이미지 처리 방법 및 장치, 및 기억 매체
SG11202009395SA SG11202009395SA (en) 2019-01-15 2019-11-15 Network optimization method and apparatus, image processing method and apparatus, and storage medium
JP2020552759A JP7074877B2 (ja) 2019-01-15 2019-11-15 ネットワーク最適化方法及び装置、画像処理方法及び装置、記憶媒体及びコンピュータプログラム
US17/037,654 US11416703B2 (en) 2019-01-15 2020-09-29 Network optimization method and apparatus, image processing method and apparatus, and storage medium

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910036096.X 2019-01-15
CN201910036096.XA CN111435432B (zh) 2019-01-15 2019-01-15 网络优化方法及装置、图像处理方法及装置、存储介质

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US17/037,654 Continuation US11416703B2 (en) 2019-01-15 2020-09-29 Network optimization method and apparatus, image processing method and apparatus, and storage medium

Publications (1)

Publication Number Publication Date
WO2020147414A1 true WO2020147414A1 (zh) 2020-07-23

Family

ID=71580810

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/118922 Ceased WO2020147414A1 (zh) 2019-01-15 2019-11-15 网络优化方法及装置、图像处理方法及装置、存储介质

Country Status (7)

Country Link
US (1) US11416703B2 (zh)
JP (1) JP7074877B2 (zh)
KR (1) KR102454515B1 (zh)
CN (1) CN111435432B (zh)
SG (1) SG11202009395SA (zh)
TW (1) TWI757668B (zh)
WO (1) WO2020147414A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2022070106A1 (en) * 2020-09-29 2022-04-07 Servicenow Canada Inc. Systems and methods for enforcing constraints in character recognition

Families Citing this family (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112132832B (zh) * 2020-08-21 2021-09-28 苏州浪潮智能科技有限公司 一种增强图像实例分割的方法、系统、设备及介质
CN112989085B (zh) * 2021-01-29 2023-07-25 腾讯科技（深圳）有限公司 图像处理方法、装置、计算机设备及存储介质
CN113469249B (zh) * 2021-06-30 2024-04-09 阿波罗智联(北京)科技有限公司 图像分类模型训练方法、分类方法、路侧设备和云控平台
KR102815333B1 (ko) * 2021-11-25 2025-06-04 중앙대학교 산학협력단 적대적 공격을 이용한 물체추적 성능 향상 시스템
CN114119976B (zh) * 2021-11-30 2024-05-14 广州文远知行科技有限公司 语义分割模型训练、语义分割的方法、装置及相关设备
CN115641609A (zh) * 2022-10-11 2023-01-24 中国科学院信息工程研究所 一种用于行人重识别的基于样本孤立机制的身份隐私保护方法和系统
TWI824796B (zh) * 2022-10-26 2023-12-01 鴻海精密工業股份有限公司 圖像分類方法、電腦設備及儲存介質
CN119206402B (zh) * 2024-09-29 2025-07-11 西南交通大学 用于火灾烟雾检测的网络模型优化方法、系统及存储介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106778604A (zh) * 2015-12-15 2017-05-31 西安电子科技大学 基于匹配卷积神经网络的行人再识别方法
US20170308740A1 (en) * 2012-09-07 2017-10-26 Stone Lock Global, Inc. Biometric identification systems and methods
CN107704838A (zh) * 2017-10-19 2018-02-16 北京旷视科技有限公司 目标对象的属性识别方法及装置
CN109101866A (zh) * 2018-06-05 2018-12-28 中国科学院自动化研究所 基于分割剪影的行人再识别方法及系统
CN109190470A (zh) * 2018-07-27 2019-01-11 北京市商汤科技开发有限公司 行人重识别方法及装置

Family Cites Families (40)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9025864B2 (en) * 2010-06-01 2015-05-05 Hewlett-Packard Development Company, L.P. Image clustering using a personal clothing model
CN103366177B (zh) * 2012-03-28 2016-12-07 佳能株式会社 对象检测分类器生成方法和设备、图像对象检测方法和设备
WO2014160426A1 (en) * 2013-03-13 2014-10-02 Kofax, Inc. Classifying objects in digital images captured using mobile devices
TW201520907A (zh) * 2013-11-29 2015-06-01 Nat Inst Chung Shan Science & Technology 影像式機踏車與行人偵測系統及其方法
TWI589468B (zh) * 2014-12-25 2017-07-01 財團法人車輛研究測試中心 行人偵測系統
GB201501510D0 (en) * 2015-01-29 2015-03-18 Apical Ltd System
US11205270B1 (en) * 2015-03-25 2021-12-21 Amazon Technologies, Inc. Collecting user pattern descriptors for use in tracking a movement of a user within a materials handling facility
US9953217B2 (en) * 2015-11-30 2018-04-24 International Business Machines Corporation System and method for pose-aware feature learning
CN106096727B (zh) * 2016-06-02 2018-12-07 腾讯科技（深圳）有限公司 一种基于机器学习的网络模型构造方法及装置
US10163042B2 (en) 2016-08-02 2018-12-25 International Business Machines Corporation Finding missing persons by learning features for person attribute classification based on deep learning
KR102634166B1 (ko) * 2016-10-04 2024-02-08 한화비전 주식회사 다중 크기 컨볼루션 블록 층을 이용한 얼굴 인식 장치
KR20180057096A (ko) * 2016-11-21 2018-05-30 삼성전자주식회사 표정 인식과 트레이닝을 수행하는 방법 및 장치
CN108229267B (zh) * 2016-12-29 2020-10-16 北京市商汤科技开发有限公司 对象属性检测、神经网络训练、区域检测方法和装置
KR102036963B1 (ko) * 2017-01-03 2019-11-29 한국과학기술원 Cnn 기반의 와일드 환경에 강인한 얼굴 검출 방법 및 시스템
US10255681B2 (en) * 2017-03-02 2019-04-09 Adobe Inc. Image matting using deep learning
CN108229276B (zh) * 2017-03-31 2020-08-11 北京市商汤科技开发有限公司 神经网络训练及图像处理方法、装置和电子设备
US10824910B2 (en) * 2017-05-04 2020-11-03 Htc Corporation Image processing method, non-transitory computer readable storage medium and image processing system
US20180322623A1 (en) * 2017-05-08 2018-11-08 Aquifi, Inc. Systems and methods for inspection and defect detection using 3-d scanning
KR102139729B1 (ko) * 2017-06-09 2020-07-31 한국과학기술원 전자 장치 및 학습 모델의 재학습 방법
FR3073311A1 (fr) * 2017-11-09 2019-05-10 Centralesupelec Procede d'estimation de pose d'une camera dans le referentiel d'une scene tridimensionnelle, dispositif, systeme de realite augmentee et programme d'ordinateur associe
CN108230359B (zh) * 2017-11-12 2021-01-26 北京市商汤科技开发有限公司 目标检测方法和装置、训练方法、电子设备、程序和介质
CN108229379A (zh) * 2017-12-29 2018-06-29 广东欧珀移动通信有限公司 图像识别方法、装置、计算机设备和存储介质
CN108090470B (zh) * 2018-01-10 2020-06-23 浙江大华技术股份有限公司 一种人脸对齐方法及装置
JP7062960B2 (ja) * 2018-01-12 2022-05-09 株式会社リコー 情報処理システム、プログラム、情報処理装置及び情報出力方法
CN110097606B (zh) * 2018-01-29 2023-07-07 微软技术许可有限责任公司 面部合成
CN108875779A (zh) * 2018-05-07 2018-11-23 深圳市恒扬数据股份有限公司 神经网络的训练方法、装置及终端设备
US11048919B1 (en) * 2018-05-30 2021-06-29 Amazon Technologies, Inc. Person tracking across video instances
CN109145150B (zh) * 2018-06-15 2021-02-12 深圳市商汤科技有限公司 目标匹配方法及装置、电子设备和存储介质
US12249074B2 (en) * 2018-06-29 2025-03-11 Hinge Health, Inc. Human pose analysis system and method
CN109117744A (zh) 2018-07-20 2019-01-01 杭州电子科技大学 一种用于人脸验证的孪生神经网络训练方法
CN108986891A (zh) * 2018-07-24 2018-12-11 北京市商汤科技开发有限公司 医疗影像处理方法及装置、电子设备及存储介质
CN109089133B (zh) * 2018-08-07 2020-08-11 北京市商汤科技开发有限公司 视频处理方法及装置、电子设备和存储介质
CN109102025A (zh) * 2018-08-15 2018-12-28 电子科技大学 基于深度学习联合优化的行人重识别方法
AU2019333924B2 (en) * 2018-09-04 2024-08-22 EMvision Medical Devices Ltd Apparatus and process for medical imaging
CN109389069B (zh) * 2018-09-28 2021-01-05 北京市商汤科技开发有限公司 注视点判断方法和装置、电子设备和计算机存储介质
US10922534B2 (en) * 2018-10-26 2021-02-16 At&T Intellectual Property I, L.P. Identifying and addressing offensive actions in visual communication sessions
GB2578789A (en) * 2018-11-09 2020-05-27 Sony Corp A method, apparatus and computer program for image processing
JP6869490B2 (ja) * 2018-12-28 2021-05-12 オムロン株式会社 欠陥検査装置、欠陥検査方法、及びそのプログラム
CN110020633B (zh) * 2019-04-12 2022-11-04 腾讯科技（深圳）有限公司 姿态识别模型的训练方法、图像识别方法及装置
US20210201661A1 (en) * 2019-12-31 2021-07-01 Midea Group Co., Ltd. System and Method of Hand Gesture Detection

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170308740A1 (en) * 2012-09-07 2017-10-26 Stone Lock Global, Inc. Biometric identification systems and methods
CN106778604A (zh) * 2015-12-15 2017-05-31 西安电子科技大学 基于匹配卷积神经网络的行人再识别方法
CN107704838A (zh) * 2017-10-19 2018-02-16 北京旷视科技有限公司 目标对象的属性识别方法及装置
CN109101866A (zh) * 2018-06-05 2018-12-28 中国科学院自动化研究所 基于分割剪影的行人再识别方法及系统
CN109190470A (zh) * 2018-07-27 2019-01-11 北京市商汤科技开发有限公司 行人重识别方法及装置

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2022070106A1 (en) * 2020-09-29 2022-04-07 Servicenow Canada Inc. Systems and methods for enforcing constraints in character recognition

Also Published As

Publication number Publication date
KR20200143478A (ko) 2020-12-23
TWI757668B (zh) 2022-03-11
JP7074877B2 (ja) 2022-05-24
CN111435432B (zh) 2023-05-26
US11416703B2 (en) 2022-08-16
TW202029062A (zh) 2020-08-01
KR102454515B1 (ko) 2022-10-13
CN111435432A (zh) 2020-07-21
JP2021517321A (ja) 2021-07-15
US20210012154A1 (en) 2021-01-14
SG11202009395SA (en) 2020-10-29

Similar Documents

Publication Publication Date Title
CN111435432B (zh) 网络优化方法及装置、图像处理方法及装置、存储介质
TWI724736B (zh) 圖像處理方法及裝置、電子設備、儲存媒體和電腦程式
JP6916970B2 (ja) ビデオ処理方法及び装置、電子機器並びに記憶媒体
JP7110412B2 (ja) 生体検出方法及び装置、電子機器並びに記憶媒体
TWI747325B (zh) 目標對象匹配方法及目標對象匹配裝置、電子設備和電腦可讀儲存媒介
CN111259967B (zh) 图像分类及神经网络训练方法、装置、设备及存储介质
CN111310616A (zh) 图像处理方法及装置、电子设备和存储介质
WO2021196401A1 (zh) 图像重建方法及装置、电子设备和存储介质
WO2020155606A1 (zh) 面部识别方法及装置、电子设备和存储介质
CN111524521A (zh) 声纹提取模型训练方法和声纹识别方法、及其装置和介质
CN111242303B (zh) 网络训练方法及装置、图像处理方法及装置
CN110532956B (zh) 图像处理方法及装置、电子设备和存储介质
CN111582383B (zh) 属性识别方法及装置、电子设备和存储介质
WO2020192113A1 (zh) 图像处理方法及装置、电子设备和存储介质
TWI738349B (zh) 圖像處理方法及圖像處理裝置、電子設備和電腦可讀儲存媒體
WO2019237870A1 (zh) 目标匹配方法及装置、电子设备和存储介质
CN107220614B (zh) 图像识别方法、装置及计算机可读存储介质
CN108197585A (zh) 脸部识别方法和装置
CN109522937B (zh) 图像处理方法及装置、电子设备和存储介质
TW202119281A (zh) 人臉識別方法、電子設備和儲存介質
CN111178115B (zh) 对象识别网络的训练方法及系统
CN110135349A (zh) 识别方法、装置、设备及存储介质
WO2023024439A1 (zh) 一种行为识别方法及装置、电子设备和存储介质
CN110070046B (zh) 人脸图像识别方法及装置、电子设备和存储介质
WO2021147199A1 (zh) 网络训练方法及装置、图像处理方法及装置

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: 19910103

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2020552759

Country of ref document: JP

Kind code of ref document: A

ENP Entry into the national phase

Ref document number: 20207033113

Country of ref document: KR

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205N DATED 07/09/2021)

122 Ep: pct application non-entry in european phase

Ref document number: 19910103

Country of ref document: EP

Kind code of ref document: A1