CN114239753A - Migratable image identification method and device - Google Patents

Migratable image identification method and device Download PDF

Info

Publication number
CN114239753A
CN114239753A CN202210164555.4A CN202210164555A CN114239753A CN 114239753 A CN114239753 A CN 114239753A CN 202210164555 A CN202210164555 A CN 202210164555A CN 114239753 A CN114239753 A CN 114239753A
Authority
CN
China
Prior art keywords
image
loss
representing
target domain
entropy
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202210164555.4A
Other languages
Chinese (zh)
Other versions
CN114239753B (en
Inventor
张凯
王帆
韩忠义
房体品
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.)
Shandong Liju Robot Technology Co ltd
Original Assignee
Shandong Liju Robot Technology 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 Shandong Liju Robot Technology Co ltd filed Critical Shandong Liju Robot Technology Co ltd
Priority to CN202210164555.4A priority Critical patent/CN114239753B/en
Publication of CN114239753A publication Critical patent/CN114239753A/en
Application granted granted Critical
Publication of CN114239753B publication Critical patent/CN114239753B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting

Landscapes

  • Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Image Analysis (AREA)

Abstract

The invention relates to a transferable image identification method and a device, which relate to the technical field of image identification, and the method comprises the following steps: determining an image type of an input image recognition model; when the input image is a source domain image with a label, enabling the source domain image with the label to pass through a feature extractor and a category predictor, and determining cross entropy loss; when the input image is an unlabeled target domain image, enabling the target domain image to pass through a feature extractor and a domain discriminator and simultaneously pass through the feature extractor and a category predictor; determining the countermeasure loss according to the output result of the domain discriminator and the similarity of the center points of the target domain image and each source domain image; determining information maximization loss according to an output result of the class predictor; the image recognition model is optimized according to cross entropy loss, counter loss and information maximization loss. Through the technical scheme, the performance of target image recognition can be effectively improved, the marks for target image recognition are effectively reduced, and manpower and material resources are greatly reduced.

Description

Migratable image identification method and device
Technical Field
The invention relates to the technical field of image recognition, in particular to a transferable image recognition method and device.
Background
Migratory image recognition refers to a technology for guiding the current label-free image to accurately recognize by using labeled images with similar distribution but different distribution when image recognition is performed. In the big data era, it has become a benign development trend to analyze the value information implied in the data to guide people's life and production. However, in a real scene, it is very easy to collect a large amount of label-free data, and it is very time-consuming and labor-consuming to perform accurate manual labeling on some tasks, such as accurate labeling of a large-scale sensor image. Under the limitation, the task of current image recognition can be guided by the existing labeled image and by utilizing the similarity between the labeled image and the distribution of the image to be recognized. For example, sensors a and B capture images of the same task, and the image of sensor a (generally referred to as a source domain image) has been labeled, and since the image types captured in sensors a and B are the same, sensor B can perform efficient image recognition by virtue of the data already labeled by sensor a, without performing extensive labeling of the data obtained by sensor B. However, due to the difference of the internal structures and the like of the sensor a and the sensor B, the image data collected by the two devices have distribution difference, and therefore, how to accurately identify the image collected by the sensor B (generally referred to as a target domain image) in the presence of the image distribution difference is a difficult point in the current transferable image identification problem.
The traditional method comprises the following steps: accurate labeling is performed on data collected by a sensor, a model is retrained and used for an image recognition task, but the process generates expensive manpower waste, and accurate manual labeling on all collected data is extremely unrealistic in a big data background.
Disclosure of Invention
In order to overcome the problems in the related art, the invention provides a transferable image identification method and a device.
According to a first aspect of the embodiments of the present invention, there is provided a migratable image recognition method, including:
determining an image type of an input image recognition model, wherein the image type comprises a labeled source domain image and an unlabeled target domain image, and the image recognition model comprises a feature extractor, a category predictor and a domain discriminator;
when the input image is a labeled source domain image, enabling the labeled source domain image to pass through the feature extractor and the category predictor, and determining cross entropy loss;
when the input image is an unlabeled target domain image, enabling the target domain image to pass through the feature extractor and the domain discriminator and simultaneously pass through the feature extractor and the class predictor;
determining the countermeasure loss according to the output result of the domain discriminator and the similarity of the central points of the target domain image and each source domain image;
determining information maximization loss according to an output result of the category predictor;
optimizing the image recognition model according to the cross entropy loss, the immunity loss and the information maximization loss.
In one embodiment, preferably, the method further comprises:
acquiring a target image to be identified;
and identifying the target image according to the image identification model so as to determine the category of the target image.
In one embodiment, the cross entropy loss is preferably calculated using the following first formula:
Figure 396069DEST_PATH_IMAGE001
wherein the content of the first and second substances,D s representing the image of all the source domain(s),L CE (D s ) Representing the cross-entropy loss of all source-domain images, E represents the periodThe physician can watch the disease,x s features that represent the image of the source domain,y s a label class representing the source domain image,
Figure 126258DEST_PATH_IMAGE002
the indication function is represented by a representation of,σdenotes the softmax function, log denotes the log function,
Figure 905996DEST_PATH_IMAGE003
and K represents the total class number of the image.
In one embodiment, preferably, determining the countermeasure loss according to the output result of the domain discriminator and the similarity of the target domain image and the center point of each source domain image includes:
determining an initial countermeasure loss from the output result of the domain discriminator, wherein the initial countermeasure loss is calculated using a second calculation formula:
Figure 737423DEST_PATH_IMAGE004
wherein the content of the first and second substances,L d_initial (D i ) Representing the initial countermeasure loss for the ith target field image,D i representing the image of the i-th target domain,x t representing the features of the ith target domain image,D(G(x t ) Represents the output of the ith target domain image through the feature extractor and then through the domain discriminator, which is equivalent to a binary classification problem,d i a binary label representing the ith target domain image, for indicating whether the target domain image belongs to the source domain or the target domain, and maximizingL d_initial (D i ) Enabling the domain discriminator to carry out feature level alignment;
determining the clustering center of each category of image through the features of all source domain images output by the feature extractor, wherein the clustering center is calculated by adopting the following third calculation formula;
Figure 551795DEST_PATH_IMAGE005
wherein the content of the first and second substances,c k the cluster center of the image representing the kth class,x s representing the features of the source domain image S,y s a label class representing the source domain image S,D s representing the image of all the source domain(s),
Figure 187307DEST_PATH_IMAGE006
the indication function is represented by a representation of,G(x s) The feature which represents the feature of the source domain image S and is output after passing through the feature extractor;
calculating the similarity between each target domain image and the cluster center closest to the target domain image, and taking the similarity as the initial weight of the target domain image to resist loss, wherein the weight is calculated by adopting the following fourth calculation formula;
Figure 454341DEST_PATH_IMAGE007
wherein the content of the first and second substances,w t the weight corresponding to the initial countermeasure loss representing the ith target domain image,D f the cosine similarity is shown in the figure,c k the cluster center of the image representing the kth class,x t features representing an ith target domain image;
calculating a countermeasure loss corresponding to the target domain image according to the initial countermeasure loss and a weight corresponding thereto, wherein the countermeasure loss is calculated using a fifth calculation formula:
Figure 715558DEST_PATH_IMAGE008
wherein the content of the first and second substances,L d (D i ) Representing the countermeasure loss of the ith target domain image,w t the weight corresponding to the initial countermeasure loss representing the ith target domain image,x t representing the features of the ith target domain image,D(G(x t ) Represents the output result of the ith target domain image after passing through the feature extractor and the domain discriminator,d i a binary label representing the ith target domain image.
In one embodiment, preferably, determining the information maximization loss according to the output result of the class predictor comprises:
calculating the entropy minimization loss and the class average entropy maximization loss of the target domain image according to the output result of the class predictor;
calculating the information maximization loss according to the entropy minimization loss and the class average entropy maximization loss;
wherein the entropy minimization loss is calculated using the following sixth calculation formula:
Figure 384436DEST_PATH_IMAGE009
wherein the content of the first and second substances,L ent (D t ) Representing the loss of minimum entropy in the said entropy,D t representing all of the target domain images,σdenotes the softmax function, H (G: (G) (G))x t ) Represents the output result of the target domain image after passing through the feature extractor and the label predictor, K represents the total class number of the image, E represents the expectation,x t representing a target domain image;
calculating the class average entropy maximization loss by adopting the following seventh calculation formula:
Figure 689385DEST_PATH_IMAGE010
wherein the content of the first and second substances,L div (D t ) Represents the average entropy maximization penalty of the class,
Figure 443714DEST_PATH_IMAGE011
represents the average probability vector after softmax of all samples of class k;
wherein the information maximization loss is calculated by adopting the following eighth calculation formula:
L IM = L ent + L div
L IM indicating that the information is maximally lost,L ent (D t ) Representing the loss of minimum entropy in the said entropy,L div (D t ) Representing the class average entropy maximization loss.
In one embodiment, preferably, optimizing the image recognition model according to the cross-entropy loss, the immunity loss, and the information maximization loss comprises:
determining a model final loss according to the cross entropy loss, the immunity loss and the information maximization loss, wherein the model final loss is calculated by adopting a ninth calculation formula as follows:
L = L CE (D s ) - L d (D t ) + βL IM
l represents the final loss of the model,L CE (D s ) Representing the cross-entropy loss as a function of time,L d (D t ) The loss of confrontation is expressed as,L IM indicating that the information is maximally lost,βindicating the equalization parameters.
According to a second aspect of the embodiments of the present invention, there is provided a migratable image recognition apparatus including:
the image recognition system comprises a first determination module, a second determination module and a third determination module, wherein the first determination module is used for determining the image type of an input image recognition model, the image type comprises a labeled source domain image and an unlabeled target domain image, and the image recognition model comprises a feature extractor, a category predictor and a domain discriminator;
the first processing module is used for enabling the source domain image with the label to pass through the feature extractor and the class predictor and determining cross entropy loss when the input image is the source domain image with the label;
the second processing module is used for enabling the target domain image to pass through the feature extractor and the domain discriminator and simultaneously pass through the feature extractor and the class predictor when the input image is the target domain image without a label;
the second determining module is used for determining the countermeasure loss according to the output result of the domain discriminator and the similarity of the central points of the target domain image and each source domain image;
the third determining module is used for determining information maximization loss according to the output result of the category predictor;
an optimization module to optimize the image recognition model according to the cross entropy loss, the immunity loss, and the information maximization loss.
In one embodiment, preferably, the apparatus further comprises:
the acquisition module is used for acquiring a target image to be identified;
and the identification module is used for identifying the target image according to the image identification model so as to determine the category of the target image.
In one embodiment, the cross entropy loss is preferably calculated using the following first formula:
Figure 243043DEST_PATH_IMAGE012
wherein the content of the first and second substances,D s representing the image of all the source domain(s),L CE (D s ) Represents the cross-entropy loss of all source domain images, E represents the expectation,x s features that represent the image of the source domain,y s to representThe label category of the source domain image,
Figure 766428DEST_PATH_IMAGE013
the indication function is represented by a representation of,σdenotes the softmax function, log denotes the log function,
Figure 196272DEST_PATH_IMAGE003
and K represents the total class number of the image.
In one embodiment, preferably, the second determining module is configured to:
determining an initial countermeasure loss from the output result of the domain discriminator, wherein the initial countermeasure loss is calculated using a second calculation formula:
Figure 250947DEST_PATH_IMAGE014
wherein the content of the first and second substances,L d_initial (D i ) Representing the initial countermeasure loss for the ith target field image,D i representing the image of the i-th target domain,x t representing the features of the ith target domain image,D(G(x t ) Represents the output of the ith target domain image through the feature extractor and then through the domain discriminator, which is equivalent to a binary classification problem,d i a binary label representing the ith target domain image, for indicating whether the target domain image belongs to the source domain or the target domain, and maximizingL d_initial (D i ) Enabling the domain discriminator to carry out feature level alignment;
determining the clustering center of each category of image through the features of all source domain images output by the feature extractor, wherein the clustering center is calculated by adopting the following third calculation formula;
Figure 526071DEST_PATH_IMAGE015
wherein the content of the first and second substances,c k the cluster center of the image representing the kth class,x s representing the features of the source domain image S,y s a label class representing the source domain image S,D s representing the image of all the source domain(s),
Figure 231858DEST_PATH_IMAGE013
the indication function is represented by a representation of,G(x s) The feature which represents the feature of the source domain image S and is output after passing through the feature extractor;
calculating the similarity between each target domain image and the cluster center closest to the target domain image, and taking the similarity as the initial weight of the target domain image to resist loss, wherein the weight is calculated by adopting the following fourth calculation formula;
Figure 567025DEST_PATH_IMAGE016
wherein the content of the first and second substances,w t the weight corresponding to the initial countermeasure loss representing the ith target domain image,D f the cosine similarity is shown in the figure,c k the cluster center of the image representing the kth class,x t features representing an ith target domain image;
calculating a countermeasure loss corresponding to the target domain image according to the initial countermeasure loss and a weight corresponding thereto, wherein the countermeasure loss is calculated using a fifth calculation formula:
Figure 607531DEST_PATH_IMAGE017
wherein the content of the first and second substances,L d (D i ) Representing the countermeasure loss of the ith target domain image,w t the weight corresponding to the initial countermeasure loss representing the ith target domain image,x t indicates the ith itemThe characteristics of the domain-marked image,D(G(x t ) Represents the output result of the ith target domain image after passing through the feature extractor and the domain discriminator,d i a binary label representing an ith target domain image;
the third determining module is to:
calculating the entropy minimization loss and the class average entropy maximization loss of the target domain image according to the output result of the class predictor;
calculating the information maximization loss according to the entropy minimization loss and the class average entropy maximization loss;
wherein the entropy minimization loss is calculated using the following sixth calculation formula:
Figure 686345DEST_PATH_IMAGE018
wherein the content of the first and second substances,L ent (D t ) Representing the loss of minimum entropy in the said entropy,D t representing all of the target domain images,σdenotes the softmax function, H (G: (G) (G))x t ) Represents the output result of the target domain image after passing through the feature extractor and the label predictor, K represents the total class number of the image, E represents the expectation,x t representing a target domain image;
calculating the class average entropy maximization loss by adopting the following seventh calculation formula:
Figure 246640DEST_PATH_IMAGE019
wherein the content of the first and second substances,L div (D t ) Represents the average entropy maximization penalty of the class,
Figure 18287DEST_PATH_IMAGE020
represents the average probability vector after softmax of all samples of class k;
wherein the information maximization loss is calculated by adopting the following eighth calculation formula:
L IM = L ent + L div
L IM indicating that the information is maximally lost,L ent (D t ) Representing the loss of minimum entropy in the said entropy,L div (D t ) Representing the class average entropy maximization loss;
the optimization module is configured to:
determining a model final loss according to the cross entropy loss, the immunity loss and the information maximization loss, wherein the model final loss is calculated by adopting a ninth calculation formula as follows:
L = L CE (D s ) - L d (D t ) + βL IM
l represents the final loss of the model,L CE (D s ) Representing the cross-entropy loss as a function of time,L d (D t ) The loss of confrontation is expressed as,L IM indicating that the information is maximally lost,βindicating the equalization parameters.
According to a third aspect of embodiments of the present invention, there is provided a computer readable storage medium having stored thereon computer instructions which, when executed by a processor, implement the steps of the method of any one of the first aspect.
The technical scheme provided by the embodiment of the invention can have the following beneficial effects:
compared with the prior art, the technical scheme of the invention not only utilizes the weighted countermeasure loss to optimize the feature extractor module, but also utilizes the cross entropy loss and the information maximization loss to optimize the label predictor module, thereby effectively improving the performance of target image identification, effectively reducing the labels for the target image identification and greatly reducing manpower and material resources.
It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention, as claimed.
Drawings
The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and together with the description, serve to explain the principles of the invention.
FIG. 1 is a flow diagram illustrating a migratable image recognition method in accordance with an exemplary embodiment.
Fig. 2 is a schematic diagram illustrating a migratable image recognition method in accordance with an exemplary embodiment.
FIG. 3 is a flow chart illustrating another migratable image recognition method in accordance with an exemplary embodiment.
Fig. 4 is a block diagram illustrating a migratable image recognition device in accordance with an exemplary embodiment.
Fig. 5 is a block diagram illustrating a migratable image recognition device in accordance with an exemplary embodiment.
Detailed Description
Reference will now be made in detail to the exemplary embodiments, examples of which are illustrated in the accompanying drawings. When the following description refers to the accompanying drawings, like numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the invention, as detailed in the appended claims.
Fig. 1 is a flow chart illustrating a method of migratable image recognition, as shown in fig. 1, in accordance with an exemplary embodiment, the method comprising:
step S101, determining an image type of an input image recognition model, wherein the image type comprises a labeled source domain image and a non-labeled target domain image, and the image recognition model comprises a feature extractor, a category predictor and a domain discriminator as shown in FIG. 2;
step S102, when the input image is a source domain image with a label, enabling the source domain image with the label to pass through the feature extractor and the category predictor, and determining cross entropy loss;
step S103, when the input image is a label-free target domain image, enabling the target domain image to pass through the feature extractor and the domain discriminator and simultaneously pass through the feature extractor and the class predictor;
step S104, determining the countermeasure loss according to the output result of the domain discriminator and the similarity of the center points of the target domain image and each source domain image;
step S105, determining information maximization loss according to an output result of the category predictor;
and S106, optimizing the image recognition model according to the cross entropy loss, the immunity loss and the information maximization loss.
Compared with the prior art, the technical scheme of the invention not only utilizes the weighted countermeasure loss to optimize the feature extractor module, but also utilizes the cross entropy loss and the information maximization loss to optimize the label predictor module, thereby effectively improving the performance of target image identification, effectively reducing the labels for the target image identification and greatly reducing manpower and material resources.
FIG. 3 is a flow chart illustrating another migratable image recognition method in accordance with an exemplary embodiment.
As shown in fig. 3, in one embodiment, preferably, the method further comprises:
step S301, acquiring a target image to be identified;
step S302, the target image is identified according to the image identification model so as to determine the category of the target image.
In one embodiment, the cross entropy loss is preferably calculated using the following first formula:
Figure 968925DEST_PATH_IMAGE021
wherein the content of the first and second substances,D s representing the image of all the source domain(s),L CE (D s ) Represents the cross-entropy loss of all source domain images, E represents the expectation,x s features that represent the image of the source domain,y s a label class representing the source domain image,
Figure 664480DEST_PATH_IMAGE002
the indication function is represented by a representation of,σdenotes the softmax function, log denotes the log function,
Figure 16964DEST_PATH_IMAGE003
and K represents the total class number of the image.
In one embodiment, preferably, determining the countermeasure loss according to the output result of the domain discriminator and the similarity of the target domain image and the center point of each source domain image includes:
determining an initial countermeasure loss from the output result of the domain discriminator, wherein the initial countermeasure loss is calculated using a second calculation formula:
Figure 959512DEST_PATH_IMAGE022
wherein the content of the first and second substances,L d_initial (D i ) Representing the initial countermeasure loss for the ith target field image,D i representing the image of the i-th target domain,x t representing the features of the ith target domain image,D(G(x t ) Represents the output of the ith target domain image through the feature extractor and then through the domain discriminator, which is equivalent to a binary classification problem,d i represents the ithBinary label of target domain image for indicating whether the target domain image belongs to source domain or target domain, and maximizingL d_initial (D i ) Enabling the domain discriminator to carry out feature level alignment;
determining the clustering center of each category of image through the features of all source domain images output by the feature extractor, wherein the clustering center is calculated by adopting the following third calculation formula;
Figure 646714DEST_PATH_IMAGE005
wherein the content of the first and second substances,c k the cluster center of the image representing the kth class,x s representing the features of the source domain image S,y s a label class representing the source domain image S,D s representing the image of all the source domain(s),
Figure 67331DEST_PATH_IMAGE023
the indication function is represented by a representation of,G(x s) The feature which represents the feature of the source domain image S and is output after passing through the feature extractor;
calculating the similarity between each target domain image and the cluster center closest to the target domain image, and taking the similarity as the initial weight of the target domain image to resist loss, wherein the weight is calculated by adopting the following fourth calculation formula;
Figure 539901DEST_PATH_IMAGE007
wherein the content of the first and second substances,w t the weight corresponding to the initial countermeasure loss representing the ith target domain image,D f the cosine similarity is shown in the figure,c k the cluster center of the image representing the kth class,x t features representing an ith target domain image;
calculating a countermeasure loss corresponding to the target domain image according to the initial countermeasure loss and a weight corresponding thereto, wherein the countermeasure loss is calculated using a fifth calculation formula:
Figure 715667DEST_PATH_IMAGE008
wherein the content of the first and second substances,L d (D i ) Representing the countermeasure loss of the ith target domain image,w t the weight corresponding to the initial countermeasure loss representing the ith target domain image,x t representing the features of the ith target domain image,D(G(x t ) Represents the output result of the ith target domain image after passing through the feature extractor and the domain discriminator,d i a binary label representing the ith target domain image.
In one embodiment, preferably, determining the information maximization loss according to the output result of the class predictor comprises:
calculating the entropy minimization loss and the class average entropy maximization loss of the target domain image according to the output result of the class predictor;
calculating the information maximization loss according to the entropy minimization loss and the class average entropy maximization loss;
wherein the entropy minimization loss is calculated using the following sixth calculation formula:
Figure 975920DEST_PATH_IMAGE024
wherein the content of the first and second substances,L ent (D t ) Representing the loss of minimum entropy in the said entropy,D t representing all of the target domain images,σdenotes the softmax function, H (G: (G) (G))x t ) Represents the output result of the target domain image after passing through the feature extractor and the label predictor, K represents the total class number of the image, E represents the expectation,x t representing a target domain image;
calculating the class average entropy maximization loss by adopting the following seventh calculation formula:
Figure 262545DEST_PATH_IMAGE010
wherein the content of the first and second substances,L div (D t ) Represents the average entropy maximization penalty of the class,
Figure 324042DEST_PATH_IMAGE011
represents the average probability vector after softmax of all samples of class k;
wherein the information maximization loss is calculated by adopting the following eighth calculation formula:
L IM = L ent + L div
L IM indicating that the information is maximally lost,L ent (D t ) Representing the loss of minimum entropy in the said entropy,L div (D t ) Representing the class average entropy maximization loss.
In one embodiment, preferably, optimizing the image recognition model according to the cross-entropy loss, the immunity loss, and the information maximization loss comprises:
determining a model final loss according to the cross entropy loss, the immunity loss and the information maximization loss, wherein the model final loss is calculated by adopting a ninth calculation formula as follows:
L = L CE (D s ) - L d (D t ) + βL IM
l represents the final loss of the model,L CE (D s ) Representing the cross-entropy loss as a function of time,L d (D t ) The loss of confrontation is expressed as,L IM indicating that the information is maximally lost,βindicating the equalization parameters.
Fig. 4 is a block diagram illustrating a migratable image recognition device in accordance with an exemplary embodiment.
As shown in fig. 4, according to a second aspect of the embodiments of the present invention, there is provided a migratable image recognition apparatus including:
a first determining module 41, configured to determine an image type of an input image recognition model, where the image type includes a labeled source domain image and an unlabeled target domain image, and the image recognition model includes a feature extractor, a class predictor, and a domain discriminator;
a first processing module 42, configured to, when the input image is a labeled source domain image, pass the labeled source domain image through the feature extractor and the class predictor, and determine a cross entropy loss;
a second processing module 43, configured to, when the input image is an unlabeled target domain image, make the target domain image pass through the feature extractor and the domain discriminator, and pass through the feature extractor and the class predictor at the same time;
a second determining module 44, configured to determine a countermeasure loss according to an output result of the domain discriminator and a similarity between the target domain image and a center point of each source domain image;
a third determining module 45, configured to determine information maximization loss according to an output result of the class predictor;
an optimization module 46 for optimizing the image recognition model based on the cross-entropy loss, the immunity loss, and the information maximization loss.
Fig. 5 is a block diagram illustrating a migratable image recognition device in accordance with an exemplary embodiment.
As shown in fig. 5, in one embodiment, preferably, the apparatus further comprises:
an obtaining module 51, configured to obtain a target image to be identified;
and the identifying module 52 is configured to identify the target image according to the image identification model to determine the category of the target image.
In one embodiment, the cross entropy loss is preferably calculated using the following first formula:
Figure 421442DEST_PATH_IMAGE025
wherein the content of the first and second substances,D s representing the image of all the source domain(s),L CE (D s ) Represents the cross-entropy loss of all source domain images, E represents the expectation,x s features that represent the image of the source domain,y s a label class representing the source domain image,
Figure 896285DEST_PATH_IMAGE013
the indication function is represented by a representation of,σdenotes the softmax function, log denotes the log function,
Figure 924284DEST_PATH_IMAGE003
and K represents the total class number of the image.
In one embodiment, preferably, the second determining module is configured to:
determining an initial countermeasure loss from the output result of the domain discriminator, wherein the initial countermeasure loss is calculated using a second calculation formula:
Figure 105867DEST_PATH_IMAGE026
wherein the content of the first and second substances,L d_initial (D i ) Representing the initial countermeasure loss for the ith target field image,D i representing the image of the i-th target domain,x t representing the features of the ith target domain image,D(G(x t ) Representing the output result of the i-th target domain image passing through the feature extractor and then passing through the domain discriminatorThe output of the classifier is equivalent to a binary classification problem,d i a binary label representing the ith target domain image, for indicating whether the target domain image belongs to the source domain or the target domain, and maximizingL d_initial (D i ) Enabling the domain discriminator to carry out feature level alignment;
determining the clustering center of each category of image through the features of all source domain images output by the feature extractor, wherein the clustering center is calculated by adopting the following third calculation formula;
Figure 607124DEST_PATH_IMAGE015
wherein the content of the first and second substances,c k the cluster center of the image representing the kth class,x s representing the features of the source domain image S,y s a label class representing the source domain image S,D s representing the image of all the source domain(s),
Figure 506947DEST_PATH_IMAGE023
the indication function is represented by a representation of,G(x s) The feature which represents the feature of the source domain image S and is output after passing through the feature extractor;
calculating the similarity between each target domain image and the cluster center closest to the target domain image, and taking the similarity as the initial weight of the target domain image to resist loss, wherein the weight is calculated by adopting the following fourth calculation formula;
Figure 135375DEST_PATH_IMAGE016
wherein the content of the first and second substances,w t the weight corresponding to the initial countermeasure loss representing the ith target domain image,D f the cosine similarity is shown in the figure,c k the cluster center of the image representing the kth class,x t features representing an ith target domain image;
calculating a countermeasure loss corresponding to the target domain image according to the initial countermeasure loss and a weight corresponding thereto, wherein the countermeasure loss is calculated using a fifth calculation formula:
Figure 437043DEST_PATH_IMAGE027
wherein the content of the first and second substances,L d (D i ) Representing the countermeasure loss of the ith target domain image,w t the weight corresponding to the initial countermeasure loss representing the ith target domain image,x t representing the features of the ith target domain image,D(G(x t ) Represents the output result of the ith target domain image after passing through the feature extractor and the domain discriminator,d i a binary label representing an ith target domain image;
the third determining module is to:
calculating the entropy minimization loss and the class average entropy maximization loss of the target domain image according to the output result of the class predictor;
calculating the information maximization loss according to the entropy minimization loss and the class average entropy maximization loss;
wherein the entropy minimization loss is calculated using the following sixth calculation formula:
Figure 610666DEST_PATH_IMAGE028
wherein the content of the first and second substances,L ent (D t ) Representing the loss of minimum entropy in the said entropy,D t representing all of the target domain images,σdenotes the softmax function, H (G: (G) (G))x t ) Represents the output result of the target domain image after passing through the feature extractor and the label predictor, K represents the total class number of the image, E represents the expectation,x t representing a target domain image;
calculating the class average entropy maximization loss by adopting the following seventh calculation formula:
Figure 60102DEST_PATH_IMAGE019
wherein the content of the first and second substances,L div (D t ) Represents the average entropy maximization penalty of the class,
Figure 429904DEST_PATH_IMAGE020
represents the average probability vector after softmax of all samples of class k;
wherein the information maximization loss is calculated by adopting the following eighth calculation formula:
L IM = L ent + L div
L IM indicating that the information is maximally lost,L ent (D t ) Representing the loss of minimum entropy in the said entropy,L div (D t ) Representing the class average entropy maximization loss;
the optimization module is configured to:
determining a model final loss according to the cross entropy loss, the immunity loss and the information maximization loss, wherein the model final loss is calculated by adopting a ninth calculation formula as follows:
L = L CE (D s ) - L d (D t ) + βL IM
l represents the final loss of the model,L CE (D s ) Representing the cross-entropy loss as a function of time,L d (D t ) The loss of confrontation is expressed as,L IM indicating that the information is maximally lost,βindicating the equalization parameters.
According to a third aspect of embodiments of the present invention, there is provided a computer readable storage medium having stored thereon computer instructions which, when executed by a processor, implement the steps of the method of any one of the first aspect.
According to a fourth aspect of the embodiments of the present invention, there is provided a migratable image recognition system, the system including:
a processor;
a memory for storing processor-executable instructions;
wherein the processor is configured to:
determining an image type of an input image recognition model, wherein the image type comprises a labeled source domain image and an unlabeled target domain image, and the image recognition model comprises a feature extractor, a category predictor and a domain discriminator;
when the input image is a labeled source domain image, enabling the labeled source domain image to pass through the feature extractor and the category predictor, and determining cross entropy loss;
when the input image is an unlabeled target domain image, enabling the target domain image to pass through the feature extractor and the domain discriminator and simultaneously pass through the feature extractor and the class predictor;
determining the countermeasure loss according to the output result of the domain discriminator and the similarity of the central points of the target domain image and each source domain image;
determining information maximization loss according to an output result of the category predictor;
optimizing the image recognition model according to the cross entropy loss, the immunity loss and the information maximization loss.
It is further understood that the term "plurality" means two or more, and other terms are analogous. "and/or" describes the association relationship of the associated objects, meaning that there may be three relationships, e.g., a and/or B, which may mean: a exists alone, A and B exist simultaneously, and B exists alone. The character "/" generally indicates that the former and latter associated objects are in an "or" relationship. The singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise.
It will be further understood that the terms "first," "second," and the like are used to describe various information and that such information should not be limited by these terms. These terms are only used to distinguish one type of information from another and do not denote a particular order or importance. Indeed, the terms "first," "second," and the like are fully interchangeable. For example, first information may also be referred to as second information, and similarly, second information may also be referred to as first information, without departing from the scope of the present invention.
It is further to be understood that while operations are depicted in the drawings in a particular order, this is not to be understood as requiring that such operations be performed in the particular order shown or in serial order, or that all illustrated operations be performed, to achieve desirable results. In certain environments, multitasking and parallel processing may be advantageous.
Other embodiments of the invention will be apparent to those skilled in the art from consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the invention following, in general, the principles of the invention and including such departures from the present disclosure as come within known or customary practice within the art to which the invention pertains. It is intended that the specification and examples be considered as exemplary only, with a true scope and spirit of the invention being indicated by the following claims.
It will be understood that the invention is not limited to the precise arrangements described above and shown in the drawings and that various modifications and changes may be made without departing from the scope thereof. The scope of the invention is limited only by the appended claims.

Claims (10)

1. A migratory image recognition method, the method comprising:
determining an image type of an input image recognition model, wherein the image type comprises a labeled source domain image and an unlabeled target domain image, and the image recognition model comprises a feature extractor, a category predictor and a domain discriminator;
when the input image is a labeled source domain image, enabling the labeled source domain image to pass through the feature extractor and the category predictor, and determining cross entropy loss;
when the input image is an unlabeled target domain image, enabling the target domain image to pass through the feature extractor and the domain discriminator and simultaneously pass through the feature extractor and the class predictor;
determining the countermeasure loss according to the output result of the domain discriminator and the similarity of the central points of the target domain image and each source domain image;
determining information maximization loss according to an output result of the category predictor;
optimizing the image recognition model according to the cross entropy loss, the immunity loss and the information maximization loss.
2. The method of claim 1, further comprising:
acquiring a target image to be identified;
and identifying the target image according to the image identification model so as to determine the category of the target image.
3. The method of claim 1, wherein the cross-entropy loss is calculated using a first formula:
Figure 407841DEST_PATH_IMAGE001
wherein the content of the first and second substances,D s representing the image of all the source domain(s),L CE (D s ) Represents the cross-entropy loss of all source domain images, E represents the expectation,x s features that represent the image of the source domain,y s a label class representing the source domain image,
Figure 204895DEST_PATH_IMAGE002
the indication function is represented by a representation of,σdenotes the softmax function, log denotes the log function,
Figure 813731DEST_PATH_IMAGE003
and K represents the total class number of the image.
4. The method of claim 1, wherein determining the countermeasure loss according to the output of the domain discriminator and the similarity of the target domain image to the center point of each source domain image comprises:
determining an initial countermeasure loss from the output result of the domain discriminator, wherein the initial countermeasure loss is calculated using a second calculation formula:
Figure 49541DEST_PATH_IMAGE004
wherein the content of the first and second substances,L d_initial (D i ) Representing the initial countermeasure loss for the ith target field image,D i representing the image of the i-th target domain,x t representing the features of the ith target domain image,D(G(x t ) Represents the output of the ith target domain image through the feature extractor and then through the domain discriminator, which is equivalent to a binary classification problem,d i a binary label representing the ith target domain image, for indicating whether the target domain image belongs to the source domain or the target domain, and maximizingL d_initial (D i ) Enabling the domain discriminator to carry out feature level alignment;
determining the clustering center of each category of image through the features of all source domain images output by the feature extractor, wherein the clustering center is calculated by adopting the following third calculation formula;
Figure 60222DEST_PATH_IMAGE005
wherein the content of the first and second substances,c k the cluster center of the image representing the kth class,x s representing the features of the source domain image S,y s a label class representing the source domain image S,D s representing the image of all the source domain(s),
Figure 339762DEST_PATH_IMAGE006
the indication function is represented by a representation of,G(x s) The feature which represents the feature of the source domain image S and is output after passing through the feature extractor;
calculating the similarity between each target domain image and the cluster center closest to the target domain image, and taking the similarity as the initial weight of the target domain image to resist loss, wherein the weight is calculated by adopting the following fourth calculation formula;
Figure 435894DEST_PATH_IMAGE007
wherein the content of the first and second substances,w t the weight corresponding to the initial countermeasure loss representing the ith target domain image,D f the cosine similarity is shown in the figure,c k the cluster center of the image representing the kth class,x t features representing an ith target domain image;
calculating a countermeasure loss corresponding to the target domain image according to the initial countermeasure loss and a weight corresponding thereto, wherein the countermeasure loss is calculated using a fifth calculation formula:
Figure 147498DEST_PATH_IMAGE008
wherein the content of the first and second substances,L d (D i ) Representing the countermeasure loss of the ith target domain image,w t the weight corresponding to the initial countermeasure loss representing the ith target domain image,x t representing the features of the ith target domain image,D(G(x t ) Represents the output result of the ith target domain image after passing through the feature extractor and the domain discriminator,d i a binary label representing the ith target domain image.
5. The method of claim 1, wherein determining an information maximization penalty based on the output of the class predictor comprises:
calculating the entropy minimization loss and the class average entropy maximization loss of the target domain image according to the output result of the class predictor;
calculating the information maximization loss according to the entropy minimization loss and the class average entropy maximization loss;
wherein the entropy minimization loss is calculated using the following sixth calculation formula:
Figure 75003DEST_PATH_IMAGE009
wherein the content of the first and second substances,L ent (D t ) Representing the loss of minimum entropy in the said entropy,D t representing all of the target domain images,σdenotes the softmax function, H (G: (G) (G))x t ) Represents the output result of the target domain image after passing through the feature extractor and the label predictor, K represents the total class number of the image, E represents the expectation,x t representing a target domain image;
calculating the class average entropy maximization loss by adopting the following seventh calculation formula:
Figure 213860DEST_PATH_IMAGE010
wherein the content of the first and second substances,L div (D t ) Represents the average entropy maximization penalty of the class,
Figure 610338DEST_PATH_IMAGE011
represents the average probability vector after softmax of all samples of class k;
wherein the information maximization loss is calculated by adopting the following eighth calculation formula:
L IM = L ent + L div
L IM indicating that the information is maximally lost,L ent (D t ) Representing the loss of minimum entropy in the said entropy,L div (D t ) Representing the class average entropy maximization loss.
6. The method of claim 1, wherein optimizing the image recognition model based on the cross-entropy loss, the immunity loss, and the information maximization loss comprises:
determining a model final loss according to the cross entropy loss, the immunity loss and the information maximization loss, wherein the model final loss is calculated by adopting a ninth calculation formula as follows:
L = L CE (D s ) - L d (D t ) + βL IM
l represents the final loss of the model,L CE (D s ) Representing the cross-entropy loss as a function of time,L d (D t ) The loss of confrontation is expressed as,L IM indicating that the information is maximally lost,βindicating the equalization parameters.
7. A migrateable image recognition apparatus, comprising:
the image recognition system comprises a first determination module, a second determination module and a third determination module, wherein the first determination module is used for determining the image type of an input image recognition model, the image type comprises a labeled source domain image and an unlabeled target domain image, and the image recognition model comprises a feature extractor, a category predictor and a domain discriminator;
the first processing module is used for enabling the source domain image with the label to pass through the feature extractor and the class predictor and determining cross entropy loss when the input image is the source domain image with the label;
the second processing module is used for enabling the target domain image to pass through the feature extractor and the domain discriminator and simultaneously pass through the feature extractor and the class predictor when the input image is the target domain image without a label;
the second determining module is used for determining the countermeasure loss according to the output result of the domain discriminator and the similarity of the central points of the target domain image and each source domain image;
the third determining module is used for determining information maximization loss according to the output result of the category predictor;
an optimization module to optimize the image recognition model according to the cross entropy loss, the immunity loss, and the information maximization loss.
8. The apparatus of claim 7, further comprising:
the acquisition module is used for acquiring a target image to be identified;
and the identification module is used for identifying the target image according to the image identification model so as to determine the category of the target image.
9. The apparatus of claim 7, wherein the cross entropy loss is calculated using a first formula:
Figure 860054DEST_PATH_IMAGE012
wherein the content of the first and second substances,D s representing the image of all the source domain(s),L CE (D s ) Represents the cross-entropy loss of all source domain images, E represents the expectation,x s features that represent the image of the source domain,y s a label class representing the source domain image,
Figure 845327DEST_PATH_IMAGE013
the indication function is represented by a representation of,σdenotes the softmax function, log denotes the log function,
Figure 217403DEST_PATH_IMAGE003
and K represents the total class number of the image.
10. The apparatus of claim 7, wherein the second determining module is configured to:
determining an initial countermeasure loss from the output result of the domain discriminator, wherein the initial countermeasure loss is calculated using a second calculation formula:
Figure 288127DEST_PATH_IMAGE014
wherein the content of the first and second substances,L d_initial (D i ) Representing the initial countermeasure loss for the ith target field image,D i representing the image of the i-th target domain,x t representing the features of the ith target domain image,D(G(x t ) Means that the ith target domain image passes through the feature extractor and then undergoes domain judgmentThe output result of the discriminator, the output result of the domain discriminator is equivalent to a binary classification problem,d i a binary label representing the ith target domain image, for indicating whether the target domain image belongs to the source domain or the target domain, and maximizingL d_initial (D i ) Enabling the domain discriminator to carry out feature level alignment;
determining the clustering center of each category of image through the features of all source domain images output by the feature extractor, wherein the clustering center is calculated by adopting the following third calculation formula;
Figure 653118DEST_PATH_IMAGE015
wherein the content of the first and second substances,c k the cluster center of the image representing the kth class,x s representing the features of the source domain image S,y s a label class representing the source domain image S,D s representing the image of all the source domain(s),
Figure 492898DEST_PATH_IMAGE016
the indication function is represented by a representation of,G(x s) The feature which represents the feature of the source domain image S and is output after passing through the feature extractor;
calculating the similarity between each target domain image and the cluster center closest to the target domain image, and taking the similarity as the initial weight of the target domain image to resist loss, wherein the weight is calculated by adopting the following fourth calculation formula;
Figure 973558DEST_PATH_IMAGE017
wherein the content of the first and second substances,w t the weight corresponding to the initial countermeasure loss representing the ith target domain image,D f the cosine similarity is shown in the figure,c k in a cluster of images representing the kth classThe heart is provided with a plurality of heart-shaped grooves,x t features representing an ith target domain image;
calculating a countermeasure loss corresponding to the target domain image according to the initial countermeasure loss and a weight corresponding thereto, wherein the countermeasure loss is calculated using a fifth calculation formula:
Figure 328316DEST_PATH_IMAGE018
wherein the content of the first and second substances,L d (D i ) Representing the countermeasure loss of the ith target domain image,w t the weight corresponding to the initial countermeasure loss representing the ith target domain image,x t representing the features of the ith target domain image,D(G(x t ) Represents the output result of the ith target domain image after passing through the feature extractor and the domain discriminator,d i a binary label representing an ith target domain image;
the third determining module is to:
calculating the entropy minimization loss and the class average entropy maximization loss of the target domain image according to the output result of the class predictor;
calculating the information maximization loss according to the entropy minimization loss and the class average entropy maximization loss;
wherein the entropy minimization loss is calculated using the following sixth calculation formula:
Figure 185413DEST_PATH_IMAGE019
wherein the content of the first and second substances,L ent (D t ) Representing the loss of minimum entropy in the said entropy,D t representing all of the target domain images,σdenotes the softmax function, H (G: (G) (G))x t ) Represents the output result of the target domain image after passing through the feature extractor and the label predictor, K represents the total class number of the image,e indicates that it is desired to,x t representing a target domain image;
calculating the class average entropy maximization loss by adopting the following seventh calculation formula:
Figure 692749DEST_PATH_IMAGE020
，
wherein the content of the first and second substances,L div (D t ) Represents the average entropy maximization penalty of the class,
Figure 344310DEST_PATH_IMAGE021
represents the average probability vector after softmax of all samples of class k;
wherein the information maximization loss is calculated by adopting the following eighth calculation formula:
L IM = L ent + L div
L IM indicating that the information is maximally lost,L ent (D t ) Representing the loss of minimum entropy in the said entropy,L div (D t ) Representing the class average entropy maximization loss;
the optimization module is configured to:
determining a model final loss according to the cross entropy loss, the immunity loss and the information maximization loss, wherein the model final loss is calculated by adopting a ninth calculation formula as follows:
L = L CE (D s ) - L d (D t ) + βL IM
l represents the final loss of the model,L CE (D s ) To representThe cross-entropy loss is a loss of,L d (D t ) The loss of confrontation is expressed as,L IM indicating that the information is maximally lost,βindicating the equalization parameters.
CN202210164555.4A 2022-02-23 2022-02-23 Migratable image identification method and device Active CN114239753B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210164555.4A CN114239753B (en) 2022-02-23 2022-02-23 Migratable image identification method and device

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210164555.4A CN114239753B (en) 2022-02-23 2022-02-23 Migratable image identification method and device

Publications (2)

Publication Number Publication Date
CN114239753A true CN114239753A (en) 2022-03-25
CN114239753B CN114239753B (en) 2022-07-22

Family

ID=80747785

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210164555.4A Active CN114239753B (en) 2022-02-23 2022-02-23 Migratable image identification method and device

Country Status (1)

Country Link
CN (1) CN114239753B (en)

Citations (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109740682A (en) * 2019-01-08 2019-05-10 南京大学 A kind of image-recognizing method converting and generate model based on domain
CN110322446A (en) * 2019-07-01 2019-10-11 华中科技大学 A kind of domain adaptive semantic dividing method based on similarity space alignment
US20200082221A1 (en) * 2018-09-06 2020-03-12 Nec Laboratories America, Inc. Domain adaptation for instance detection and segmentation
CN111814871A (en) * 2020-06-13 2020-10-23 浙江大学 Image classification method based on reliable weight optimal transmission
US10839269B1 (en) * 2020-03-20 2020-11-17 King Abdulaziz University System for fast and accurate visual domain adaptation
CN113128287A (en) * 2019-12-31 2021-07-16 暗物智能科技(广州)有限公司 Method and system for training cross-domain facial expression recognition model and facial expression recognition
CN113159199A (en) * 2021-04-27 2021-07-23 广东工业大学 Cross-domain image classification method based on structural feature enhancement and class center matching
CN113435509A (en) * 2021-06-28 2021-09-24 山东力聚机器人科技股份有限公司 Small sample scene classification and identification method and system based on meta-learning
CN113435546A (en) * 2021-08-26 2021-09-24 广东众聚人工智能科技有限公司 Migratable image recognition method and system based on differentiation confidence level
CN113469186A (en) * 2021-06-30 2021-10-01 华侨大学 Cross-domain migration image segmentation method based on small amount of point labels
CN113688867A (en) * 2021-07-20 2021-11-23 广东工业大学 Cross-domain image classification method
CN113792609A (en) * 2021-08-19 2021-12-14 北京理工大学 Method for detecting image abnormity through aligning multi-spectrum cross-domain features in automatic driving

Patent Citations (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200082221A1 (en) * 2018-09-06 2020-03-12 Nec Laboratories America, Inc. Domain adaptation for instance detection and segmentation
CN109740682A (en) * 2019-01-08 2019-05-10 南京大学 A kind of image-recognizing method converting and generate model based on domain
CN110322446A (en) * 2019-07-01 2019-10-11 华中科技大学 A kind of domain adaptive semantic dividing method based on similarity space alignment
CN113128287A (en) * 2019-12-31 2021-07-16 暗物智能科技(广州)有限公司 Method and system for training cross-domain facial expression recognition model and facial expression recognition
US10839269B1 (en) * 2020-03-20 2020-11-17 King Abdulaziz University System for fast and accurate visual domain adaptation
US20210390355A1 (en) * 2020-06-13 2021-12-16 Zhejiang University Image classification method based on reliable weighted optimal transport (rwot)
CN111814871A (en) * 2020-06-13 2020-10-23 浙江大学 Image classification method based on reliable weight optimal transmission
CN113159199A (en) * 2021-04-27 2021-07-23 广东工业大学 Cross-domain image classification method based on structural feature enhancement and class center matching
CN113435509A (en) * 2021-06-28 2021-09-24 山东力聚机器人科技股份有限公司 Small sample scene classification and identification method and system based on meta-learning
CN113469186A (en) * 2021-06-30 2021-10-01 华侨大学 Cross-domain migration image segmentation method based on small amount of point labels
CN113688867A (en) * 2021-07-20 2021-11-23 广东工业大学 Cross-domain image classification method
CN113792609A (en) * 2021-08-19 2021-12-14 北京理工大学 Method for detecting image abnormity through aligning multi-spectrum cross-domain features in automatic driving
CN113435546A (en) * 2021-08-26 2021-09-24 广东众聚人工智能科技有限公司 Migratable image recognition method and system based on differentiation confidence level

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
JOGENDRA NATH KUNDU ET AL: "Universal Source-Free Domain Adaptation", 《CVPR 2020》 *
YEN-CHANG HSU ET AL: "LEARNING TO CLUSTER IN ORDER TO TRANSFER ACROSS DOMAINS AND TASKS", 《ARXIV》 *
周强: "面向图像分类的迀移学习算法研究", 《中国博士学位论文全文数据库信息科技辑》 *
杨潇 等: "基于域适应的煤矿环境监控图像语义分割", 《煤炭学报》 *

Also Published As

Publication number Publication date
CN114239753B (en) 2022-07-22

Similar Documents

Publication Publication Date Title
CN109949317B (en) Semi-supervised image example segmentation method based on gradual confrontation learning
CN111680614B (en) Abnormal behavior detection method based on video monitoring
CN113435546B (en) Migratable image recognition method and system based on differentiation confidence level
CN112232371B (en) American license plate recognition method based on YOLOv3 and text recognition
CN110728694B (en) Long-time visual target tracking method based on continuous learning
CN110458022B (en) Autonomous learning target detection method based on domain adaptation
CN112561960B (en) Multi-target tracking repositioning method based on track similarity measurement learning
CN113283282A (en) Weak supervision time sequence action detection method based on time domain semantic features
CN115861738A (en) Category semantic information guided remote sensing target detection active sampling method
CN111461121A (en) Electric meter number identification method based on YO L OV3 network
CN111898528B (en) Data processing method, device, computer readable medium and electronic equipment
CN110909645B (en) Crowd counting method based on semi-supervised manifold embedding
CN116681961A (en) Weak supervision target detection method based on semi-supervision method and noise processing
CN113673631B (en) Abnormal image detection method and device
CN114239753B (en) Migratable image identification method and device
CN114492657A (en) Plant disease classification method and device, electronic equipment and storage medium
CN111696070A (en) Multispectral image fusion power internet of things fault point detection method based on deep learning
CN113449676B (en) Pedestrian re-identification method based on two-way interaction-based disentanglement learning
CN115565008A (en) Transferable image recognition detection system, method and computer readable storage medium
CN116403074B (en) Semi-automatic image labeling method and device based on active labeling
Meena Deshpande License Plate Detection and Recognition using YOLO v4
CN115131588B (en) Image robust clustering method based on fuzzy clustering
CN116843368B (en) Marketing data processing method based on ARMA model
CN112990145B (en) Group-sparse-based age estimation method and electronic equipment
CN110781942B (en) Semi-supervised image classification method and system

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant
CP02 Change in the address of a patent holder

Address after: Room 1409, Floor 14, Building 1, High tech Zone Entrepreneurship Center, No. 177, Gaoxin 6th Road, Rizhao, Shandong 276801

Patentee after: Shandong Liju Robot Technology Co.,Ltd.

Address before: 276808 No.99, Yuquan 2nd Road, antonwei street, Lanshan District, Rizhao City, Shandong Province

Patentee before: Shandong Liju Robot Technology Co.,Ltd.

CP02 Change in the address of a patent holder
CB03 Change of inventor or designer information

Inventor after: Zhao Yue

Inventor after: Zhang Kai

Inventor after: Wang Fan

Inventor after: Han Zhongyi

Inventor after: Fang Tipin

Inventor before: Zhang Kai

Inventor before: Wang Fan

Inventor before: Han Zhongyi

Inventor before: Fang Tipin

CB03 Change of inventor or designer information