CN112784494A - Training method of false positive recognition model, target recognition method and device - Google Patents

Training method of false positive recognition model, target recognition method and device Download PDF

Info

Publication number
CN112784494A
CN112784494A CN202110113779.8A CN202110113779A CN112784494A CN 112784494 A CN112784494 A CN 112784494A CN 202110113779 A CN202110113779 A CN 202110113779A CN 112784494 A CN112784494 A CN 112784494A
Authority
CN
China
Prior art keywords
target
recognition model
recognition
model
false positive
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
CN202110113779.8A
Other languages
Chinese (zh)
Other versions
CN112784494B (en
Inventor
夏威
高欣
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.)
Suzhou Institute of Biomedical Engineering and Technology of CAS
Original Assignee
Suzhou Institute of Biomedical Engineering and Technology of CAS
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 Suzhou Institute of Biomedical Engineering and Technology of CAS filed Critical Suzhou Institute of Biomedical Engineering and Technology of CAS
Priority to CN202110113779.8A priority Critical patent/CN112784494B/en
Publication of CN112784494A publication Critical patent/CN112784494A/en
Application granted granted Critical
Publication of CN112784494B publication Critical patent/CN112784494B/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
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
    • G06F2119/02—Reliability analysis or reliability optimisation; Failure analysis, e.g. worst case scenario performance, failure mode and effects analysis [FMEA]

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • Computer Hardware Design (AREA)
  • Geometry (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Image Analysis (AREA)

Abstract

The invention relates to the technical field of image processing, in particular to a training method of a false positive identification model, a target identification method and a device, wherein the training method comprises the steps of obtaining a target identification model and obtaining at least one disturbed target identification model, wherein the disturbed target identification model is obtained by training output characteristic information of a preset convolution layer in the target identification model after random disturbance processing; determining a first uncertainty characteristic vector based on the relation between a first recognition result of each disturbed target recognition model to the target image and a second recognition result of the target recognition model to the target image; and training the false positive recognition model according to the first uncertainty characteristic vector and the label of the target image to determine the target false positive recognition model. The uncertainty characteristic vector is used for training the false positive recognition model, and the false positive recognition model is subsequently used for screening the recognition result, so that the accuracy of the recognition result can be improved.

Description

Training method of false positive recognition model, target recognition method and device
Technical Field
The invention relates to the technical field of image processing, in particular to a training method of a false positive identification model, a target identification method and a target identification device.
Background
The Convolutional Neural Network (CNN) is the most widely applied deep learning method in the field of image processing, and is used for object recognition by inputting a picture into the CNN and then transmitting the picture to a plurality of convolutional layers and pooling layers in the CNN, so that automatic extraction of image features can be realized, feature representation of an object can be automatically learned layer by layer, and abstraction and description of object layering can be realized. Among them, object recognition has been widely used in various fields, for example, in unmanned driving and medical image segmentation.
No matter what method is adopted to carry out the components of the target recognition model, the components need to rely on the convolutional layer for feature extraction, and the result of the feature extraction directly influences the result of the target recognition. For some targets with correct identification, the error of feature extraction has little influence on the identification result; for some targets with identification errors (namely false positives), the target identification result is unstable under the condition of characteristic extraction errors, and the probability value is mainly reflected to have large fluctuation, so that the accuracy of target identification is influenced.
Disclosure of Invention
In view of this, embodiments of the present invention provide a training method for a false positive recognition model, a target recognition method, and an apparatus thereof, so as to solve the problem of low target recognition accuracy.
According to a first aspect, an embodiment of the present invention provides a training method for a false positive recognition model, where the training method includes:
acquiring a target recognition model and at least one disturbed target recognition model, wherein the disturbed target recognition model is obtained by training after random disturbance processing is carried out on output characteristic information of a preset convolution layer in the target recognition model;
determining a first uncertainty characteristic vector based on the relation between a first recognition result of each disturbed target recognition model to a target image and a second recognition result of the target recognition model to the target image;
and training the false positive recognition model according to the first uncertainty characteristic vector and the label of the target image to determine the target false positive recognition model.
According to the training method of the false positive recognition model provided by the embodiment of the invention, on the basis of not changing the structure of the target recognition model, random disturbance processing is carried out on the output characteristic information generated after convolution to obtain the first recognition result under the random disturbance condition, the uncertainty characteristic vector of the recognition result can be quantitatively evaluated based on the first recognition result, then the false positive recognition model is trained by utilizing the uncertainty characteristic vector, the false positive recognition model which is more accurate in recognition can be obtained, and the recognition result is subsequently screened by utilizing the false positive recognition model, so that the accuracy of the recognition result can be improved.
With reference to the first aspect, in a first implementation manner of the first aspect, the obtaining at least one perturbed target recognition model includes:
inputting a sample image into the target identification model, and extracting output characteristic information of a preset convolution layer in the target identification model;
randomly selecting a plurality of disturbance coordinate points;
processing the pixel value corresponding to the disturbance coordinate point in the output characteristic information by using the disturbance coordinate point to obtain a disturbance characteristic diagram;
and training the target recognition model based on the disturbance characteristic diagram to obtain the disturbance target recognition model.
According to the training method of the false positive identification model, the output characteristic information is disturbed in a mode of randomly selecting the disturbance coordinate point, on one hand, the randomness of disturbance can be guaranteed, on the other hand, the processing mode is simple, and the training efficiency of the false positive identification model is improved.
With reference to the first implementation manner of the first aspect, in a second implementation manner of the first aspect, the processing, by using the disturbance coordinate point, a pixel value corresponding to the disturbance coordinate point in the output feature information to obtain a disturbance feature map includes:
determining a disturbance area by taking each disturbance coordinate point as a center;
and setting the pixel value corresponding to the disturbance area in the output characteristic information as a preset value to obtain the disturbance characteristic graph.
According to the training method of the false positive identification model provided by the embodiment of the invention, the disturbance area is determined by taking the disturbance coordinate point as the center, and the pixel value corresponding to the disturbance area in the output characteristic information is processed to adjust the disturbance size, so that the reliability of the false positive identification model is further ensured.
With reference to the first aspect, in a third implementation manner of the first aspect, the determining a first uncertainty feature vector based on a relationship between a first recognition result of the perturbed target recognition model on a target image and a second recognition result of the target recognition model on the target image includes:
corresponding to each disturbed target recognition model, calculating the overlapping rate of each first target recognition area in the disturbed target recognition model and the corresponding second target recognition area in the target recognition model;
adjusting the recognition probability of the first target recognition area by using the overlapping rate to obtain the target recognition probability of each first target recognition area corresponding to each target category;
and carrying out statistical analysis on the target identification probability corresponding to each disturbance target identification model to determine the first uncertainty characteristic vector.
According to the training method for the false positive identification model, provided by the embodiment of the invention, the identification probability of the identification result corresponding to the disturbance identification model is adjusted by utilizing the disturbance target identification model and the overlapping rate corresponding to the identification result of the target identification model, so that the identification probability can be changed along with the change of the overlapping rate, and the reliability of the false positive identification model is further ensured.
With reference to the first aspect, in a fourth implementation manner of the first aspect, the determining a first uncertainty feature vector based on a relationship between a first recognition result of the perturbed target recognition model on a target image and a second recognition result of the target recognition model on the target image includes:
calculating the average value of the probabilities that all pixels in the second target identification region belong to each target category to obtain the target identification probability of the first target identification region corresponding to each target category;
and carrying out statistical analysis on the target identification probability corresponding to each disturbance target identification model to determine the first uncertainty characteristic vector.
With reference to the third implementation manner or the fourth implementation manner of the first aspect, in a fifth implementation manner of the first aspect, the statistically analyzing the target recognition probability corresponding to each of the disturbed target recognition models to determine the first uncertainty feature vector includes:
calculating a statistical value of the target identification probability corresponding to each target category by using the target identification probability corresponding to each disturbed target identification model, wherein the statistical value of the target identification probability comprises at least one of entropy, standard deviation, root mean square error, range and average absolute deviation;
and determining a first uncertainty characteristic vector corresponding to each target class based on the statistic value of the target identification probability corresponding to each target class.
According to the training method of the false positive identification model provided by the embodiment of the invention, the statistical value of the target identification probability can accurately measure the heterogeneity or the non-uniformity of data, so that the statistical value of the target identification probability is used for calculating the first uncertainty characteristic vector corresponding to each target category, and the reliability of the calculation result can be ensured.
With reference to the first aspect, in a sixth implementation manner of the first aspect, the training the false positive recognition model according to the first uncertainty feature vector and the label of the target image to determine a target false positive recognition model includes:
inputting the first uncertainty characteristic vector into the false positive identification model to obtain a prediction classification result;
and adjusting parameters of the false positive identification model by using the label of the target image and the prediction classification result to determine the target false positive identification model.
According to the training method of the false positive identification model provided by the embodiment of the invention, the false positive identification model is trained by using the uncertain feature vectors, so that the false positive rate can be reduced, and the accuracy rate of target identification is finally improved.
According to a second aspect, an embodiment of the present invention further provides a target identification method, where the identification method includes:
acquiring an image to be identified;
respectively inputting the images to be recognized into a target recognition model and at least one disturbed target recognition model to respectively obtain a third recognition result and a fourth recognition result of the images to be recognized, wherein the disturbed target recognition model is obtained by training after random disturbance processing is carried out on output characteristic information of a preset convolution layer in the target recognition model;
determining a second uncertainty feature vector based on a relationship of the third recognition result and the fourth recognition result;
and inputting the second uncertainty feature vector into the false positive recognition model, and screening the third recognition result to determine a target recognition result, wherein the false positive recognition model is obtained by training according to the training method of the false positive recognition model in the first aspect of the invention or any embodiment of the first aspect.
According to the target identification method provided by the embodiment of the invention, the false positive identification model is used for screening the target identification result, so that the false positive rate can be reduced, the identification accuracy is improved, and the further optimization of the identification result of the target identification model is realized; the method has strong universality and can be applied to all target identification based on the convolutional neural network.
According to a third aspect, an embodiment of the present invention further provides a training apparatus for a false positive recognition model, where the training apparatus includes:
the system comprises a first acquisition module, a second acquisition module and a third acquisition module, wherein the first acquisition module is used for acquiring a target recognition model and acquiring at least one disturbed target recognition model, and the disturbed target recognition model is obtained by training after random disturbance processing is carried out on output characteristic information of a preset convolution layer in the target recognition model;
the first determining module is used for determining a first uncertainty characteristic vector based on the relation between a first recognition result of each disturbed target recognition model to a target image and a second recognition result of the target recognition model to the target image;
and the training module is used for training the false positive recognition model according to the first uncertainty characteristic vector and the label of the target image to determine the target false positive recognition model.
According to the training device for the false positive identification model, provided by the embodiment of the invention, on the basis of not changing the structure of the target identification model, random disturbance processing is carried out on output characteristic information generated after convolution, a first identification result under a random disturbance condition is obtained, uncertainty characteristic vectors of the identification result can be quantitatively evaluated based on the first identification result, the false positive identification model is trained by using the uncertainty characteristic vectors, the false positive identification model which is identified more accurately can be obtained, and the identification result is subsequently screened by using the false positive identification model, so that the accuracy of the identification result can be improved.
According to a fourth aspect, an embodiment of the present invention further provides an object recognition apparatus, where the object recognition apparatus includes:
the second acquisition module is used for acquiring an image to be identified;
the recognition module is used for respectively inputting the images to be recognized into a target recognition model and at least one disturbed target recognition model to respectively obtain a third recognition result and a fourth recognition result of the images to be recognized, and the disturbed target recognition model is obtained by training after random disturbance processing is carried out on output characteristic information of a preset convolution layer in the target recognition model;
a second determining module, configured to determine a second uncertainty feature vector based on a relationship between the third recognition result and the fourth recognition result;
and a screening module, configured to input the second uncertainty feature vector into the false positive recognition model, screen the third recognition result, and determine a target recognition result, where the false positive recognition model is obtained by training according to the training method of the false positive recognition model in the first aspect of the present invention or any embodiment of the first aspect.
According to the target recognition device provided by the embodiment of the invention, the false positive recognition model is used for screening the target recognition result, so that the false positive rate can be reduced, the recognition accuracy is improved, and further optimization of the recognition result of the target recognition model is realized; the method has strong universality and can be applied to all target identification based on the convolutional neural network.
According to a fourth aspect, an embodiment of the present invention provides an electronic device, including: a memory and a processor, the memory and the processor being communicatively connected to each other, the memory having stored therein computer instructions, and the processor executing the computer instructions to perform the method for training a false positive recognition model according to the first aspect or any one of the embodiments of the first aspect, or to perform the method for target recognition according to the second aspect.
According to a fifth aspect, an embodiment of the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to execute the training method of the false positive recognition model according to the first aspect or any one of the embodiments of the first aspect, or execute the target recognition method according to the second aspect.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, and it is obvious that the drawings in the following description are some embodiments of the present invention, and other drawings can be obtained by those skilled in the art without creative efforts.
FIG. 1 is a flow diagram of a method of training a false positive identification model according to an embodiment of the invention;
FIG. 2 is a flow diagram of a method of training a false positive identification model according to an embodiment of the invention;
FIG. 3 is a flow diagram of a method of training a false positive identification model according to an embodiment of the invention;
FIG. 4 is a flow chart of a method of object recognition according to an embodiment of the present invention;
FIG. 5 is a block diagram of a training apparatus for a false positive recognition model according to an embodiment of the present invention;
FIG. 6 is a block diagram of a target recognition apparatus according to an embodiment of the present invention;
fig. 7 is a schematic diagram of a hardware structure of an electronic device according to an embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some, but not all, embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
The false positive identification model provided by the embodiment of the invention is a classification model and is used for judging whether the identification result is false positive or not. Specifically, in the process of target identification, an uncertainty characteristic vector can be formed by using an identification result, and the uncertainty characteristic vector is input into a false positive identification model to screen false positives in the identification result to obtain a more accurate identification result.
The target recognition may be target detection, target segmentation, or the like, and is not limited herein. According to the method provided by the embodiment of the invention, on the basis of not changing the network structure of the original target identification model, the uncertainty characteristic vector corresponding to the identification result of the target identification model is calculated; and then, the uncertainty characteristic vector is input into a false positive identification model to identify a false positive target detection or segmentation result, so that the false positive rate is reduced, further optimization of the target detection or segmentation result of the target identification model is realized, and the accuracy is finally improved.
In accordance with an embodiment of the present invention, there is provided an embodiment of a method for training a false positive recognition model, wherein the steps illustrated in the flowchart of the figure may be performed in a computer system, such as a set of computer-executable instructions, and wherein, although a logical order is illustrated in the flowchart, in some cases, the steps illustrated or described may be performed in an order different than that illustrated.
In this embodiment, a training method of a false positive recognition model is provided, which can be used in electronic devices, such as a computer, a mobile phone, a tablet computer, and the like, fig. 1 is a flowchart of the training method of the false positive recognition model according to the embodiment of the present invention, and as shown in fig. 1, the flowchart includes the following steps:
s11, obtaining the target recognition model and obtaining at least one disturbance target recognition model.
The disturbed target recognition model is obtained by training after random disturbance processing is carried out on output characteristic information of a preset convolution layer in the target recognition model.
The target recognition model is a model that can be used for target detection or target segmentation, and the specific structure of the model is not limited at all, and may be set according to actual conditions. For example, the object recognition model is an object detection or segmentation model based on a convolutional neural network. The target recognition model may be previously trained and stored in the electronic device, may be obtained by the electronic device through real-time training when training the false positive recognition model, or may be obtained by the electronic device from the outside, and the like.
The object recognition model is used for recognizing objects in the image and the probability that the objects correspond to each category. For example, for object detection, the output of the object recognition model may be A bounding boxes Bi(i ═ 1, 2, …, a), each bounding box BiThe corresponding probability value is PBi(c) Where C is the class of the target (C ═ 1, 2, …, C); for object segmentation, the output of the object recognition model may result in C segmentation maps Sc(c＝1，2，…，C)，ScEach pixel represents the pixel genusIn the probability of the c-th object class, the pixels larger than the preset threshold value T are judged to belong to the object, so as to generate the segmentation result of the object, namely an area R consisting of the pixels of the same typec。
The number of the disturbance recognition models acquired by the electronic device may be one, two or more, and the specific number is not limited at all, and may be set according to the actual situation. The disturbance recognition model is obtained by training after random disturbance is added to the output of the convolution layer of the target recognition model on the basis of the target recognition model. Specifically, the output information of the convolutional layer is the extracted feature information, and the accuracy of the feature information directly affects the accuracy of the recognition result, so that random disturbance is added to the output information of the convolutional layer, and the disturbance degree is quantified by using an uncertainty feature vector subsequently, so that the trained false positive recognition model has higher recognition accuracy. As can be seen, the input of the disturbance recognition model and the target recognition model are both images, and the output is the probability that each target in the image belongs to each category.
The preset convolutional layer may be one or some convolutional layers in the target recognition model (for example, a convolutional layer in the middle of the target recognition model), or may be all convolutional layers, which is not limited herein. And outputting the characteristic information output from the preset convolution layer to a next module of the preset convolution layer for processing after random disturbance processing. For example, the preset convolutional layer is the 2 nd convolutional layer in the target recognition model, and in the target recognition model, the output of the 2 nd convolutional layer is connected with the input of the 3 rd convolutional layer; in the disturbance target identification model, the output of the 2 nd convolution layer is subjected to random disturbance processing and then connected with the input of the 3 rd convolution layer.
The random disturbance may be randomly modifying the output characteristic information, for example, randomly modifying a characteristic value in the output characteristic information, and the like, which will be described in detail below.
The random disturbance target recognition model may be obtained by training the electronic device in the process of training the false positive recognition model, may also be stored in the electronic device after being trained in advance, or may be obtained from the outside by the electronic device, and the like.
And S12, determining a first uncertainty characteristic vector based on the relationship between the first recognition result of each disturbed target recognition model to the target image and the second recognition result of the target recognition model to the target image.
When the electronic equipment acquires the target recognition model and the disturbance target recognition model, the target images are respectively input into the models to obtain a second recognition result corresponding to each disturbance target recognition model and a second recognition result corresponding to the target recognition model.
As described above, the recognition result may be represented by a bounding box, or a segmentation map, and probabilities corresponding to the respective categories. If the same target is identified, the identification result obtained by using the perturbed target identification model and the target identification model is that there is an intersecting part, for example, the intersection of bounding boxes. If the attribution of each pixel value in the segmentation map is determined.
The electronic device can determine the first uncertainty feature vector using the relationship. For example, the relationship between the first recognition result and the second recognition result may measure the characterization of the stochastic disturbance by the disturbance target recognition model, and the characterization represents the uncertainty. Taking target detection as an example, if the intersection area is large, the identification result of the disturbed target identification model is reliable; if the intersection area is small, the identification result of the disturbed target identification model is inaccurate, and the identification result can be ignored in the calculation process of the uncertainty characteristic vector.
The step will be described in detail below, and is not described herein.
And S13, training the false positive recognition model according to the first uncertainty characteristic vector and the label of the target image, and determining the target false positive recognition model.
The label of the target image is used to indicate whether each target in the target image is in each category, and the specific labeling manner is not limited herein.
The electronic equipment inputs the first uncertainty characteristic vector into a false positive identification model to obtain whether each identification result belongs to each category; and calculating a loss function by using the label of the target image, further training the false positive identification model, and determining the target false positive identification model. Therefore, the input of the target false positive recognition model is uncertainty characteristic vectors, and the output is whether the target false positive recognition model belongs to each category. Therefore, the false positive recognition model may be a classifier model, for example, constructed based on a neural network, a support vector machine, a random forest, or the like, and the specific network structure is not limited at all, and may be set according to actual situations.
According to the training method for the false positive identification model, on the basis of not changing the structure of the target identification model, random disturbance processing is performed on output feature information generated after convolution, a first identification result under a random disturbance condition is obtained, uncertainty feature vectors of the identification result can be quantitatively evaluated based on the first identification result, then the false positive identification model is trained by using the uncertainty feature vectors, the false positive identification model which is identified more accurately can be obtained, the identification result is subsequently screened by using the false positive identification model, and the accuracy of the identification result can be improved.
In this embodiment, a training method of a false positive recognition model is provided, which can be used in electronic devices, such as a computer, a mobile phone, a tablet computer, and the like, fig. 2 is a flowchart of the training method of the false positive recognition model according to the embodiment of the present invention, and as shown in fig. 2, the flowchart includes the following steps:
s21, obtaining the target recognition model and obtaining at least one disturbance target recognition model.
The disturbed target recognition model is obtained by training after random disturbance processing is carried out on output characteristic information of a preset convolution layer in the target recognition model.
Specifically, the above S21 may include the following steps:
and S211, acquiring a target recognition model.
Please refer to the related description of the target recognition model in S11 in the embodiment shown in fig. 1, which is not repeated herein.
S212, inputting the sample image into the target recognition model, and extracting the output characteristic information of the preset convolution layer in the target recognition model.
The electronic equipment inputs the sample image into a target recognition model, and the convolution layers in the target recognition model perform feature extraction on the sample image and have corresponding output feature information corresponding to each convolution layer. The electronic equipment extracts the output characteristic information of the preset convolution layer to carry out random disturbance processing on the preset convolution layer.
For example, the target recognition model has Z convolutional layers, and the electronic device selects N (1 ≦ N ≦ Z) convolutional layers from the Z convolutional layers, wherein the selected convolutional layer is denoted as Convs(s ═ 1, 2, …, N). The output characteristic information of each convolutional layer is represented in the form of a characteristic diagram, convolutional layer ConvsThe generated feature map is Fs(s＝1，2，…，N)。
And S213, randomly selecting a plurality of disturbance coordinate points.
The electronic device may generate K (K)>0) A plurality of random numbers, and taking the K random numbers as a plurality of selected disturbance coordinate points Pk(K ═ 1, 2, …, K). The specific value of K may be set according to an actual situation, and is not limited herein, and may be set according to an actual situation.
And S214, processing the pixel value corresponding to the disturbance coordinate point in the output characteristic information by using the disturbance coordinate point to obtain a disturbance characteristic diagram.
After obtaining the disturbance coordinate point, the electronic device may map the disturbance coordinate point to the output characteristic information, determine a pixel value of a pixel point of the disturbance coordinate point in the output characteristic information, and after determining the pixel value, the electronic device may set the pixel value to a preset value, for example, to 0, or to another value. After setting, the disturbance characteristic diagram can be obtained.
In some optional implementations of this embodiment, the step S214 may include the following steps:
(1) and determining a disturbance area by taking each disturbance coordinate point as a center.
Electronic equipment perturbs coordinate point PkFor the center, a disturbance area is defined. The perturbation region may be a template region m of L × L size, or a template region with R as a radius, and the like, and the shape of the perturbation region is not limited herein.
(2) And setting the pixel value corresponding to the disturbance area in the output characteristic information as a preset value to obtain a disturbance characteristic diagram.
After the disturbance area is determined, the electronic equipment maps the disturbance area to the output characteristic information, determines the position of the disturbance area in the output characteristic information, and sets the pixel value corresponding to the disturbance area in the output characteristic information as a preset value, so that a disturbance characteristic diagram can be obtained. For example, the pixel value in the output characteristic information corresponding to the disturbance area is set to 0.
And determining a disturbance area by taking the disturbance coordinate point as a center, and processing the pixel value corresponding to the disturbance area in the output characteristic information so as to adjust the disturbance size and further ensure the reliability of the false positive identification model.
S215, training the target recognition model based on the disturbance characteristic diagram to obtain the disturbance target recognition model.
And the electronic equipment trains the target recognition model on the basis of the disturbance characteristic diagram, and adjusts the parameters of the target recognition model to obtain the disturbance target recognition model.
It should be noted that the number of the obtained disturbance target identification models may be set according to actual situations. The training process for multiple disturbance target recognition models may be repeated by repeating the above steps S212 to S215. For example, the electronic device may set q random numbers to control the stochastic process, and retrain the target recognition model using S212-S215 described above to obtain q perturbed target recognition models.
And S22, determining a first uncertainty characteristic vector based on the relationship between the first recognition result of each disturbed target recognition model to the target image and the second recognition result of the target recognition model to the target image.
Please refer to S12 in fig. 1, which is not described herein again.
And S23, training the false positive recognition model according to the first uncertainty characteristic vector and the label of the target image, and determining the target false positive recognition model.
Please refer to S13 in fig. 1, which is not described herein again.
According to the training method for the false positive identification model, the output characteristic information is disturbed in a mode of randomly selecting the disturbance coordinate point, on one hand, the randomness of disturbance can be guaranteed, on the other hand, the processing mode is simple, and the training efficiency of the false positive identification model is improved.
In this embodiment, a training method of a false positive recognition model is provided, which can be used in electronic devices, such as a computer, a mobile phone, a tablet computer, and the like, fig. 3 is a flowchart of the training method of the false positive recognition model according to the embodiment of the present invention, and as shown in fig. 3, the flowchart includes the following steps:
s31, obtaining the target recognition model and obtaining at least one disturbance target recognition model.
The disturbed target recognition model is obtained by training after random disturbance processing is carried out on output characteristic information of a preset convolution layer in the target recognition model.
Please refer to S21 in fig. 2 for details, which are not described herein.
And S32, determining a first uncertainty characteristic vector based on the relationship between the first recognition result of each disturbed target recognition model to the target image and the second recognition result of the target recognition model to the target image.
Specifically, when the target recognition model is used for target detection, the above S32 may include the following steps:
s321, corresponding to each disturbed target recognition model, calculating the overlapping rate of each first target recognition area in the disturbed target recognition model and the corresponding second target recognition area in the target recognition model.
After the electronic equipment obtains the disturbed target identification models, respectively utilizing each disturbed target identification model to carry out target identification on the target image to obtain a first target identification area and the probability of each category corresponding to the area; and carrying out target recognition on the target image by using the target recognition model to obtain a second target recognition area and the probability of the area corresponding to each category.
After obtaining the first target recognition area and the second target recognition area, the electronic device may determine the overlapping ratio by calculating an intersection ratio of the two areas.
Taking the example that the target identification model M _ D is used for target detection and q disturbance target identification models are obtained in the above S31, the disturbance target identification model is represented as M _ Do(o ═ 1, 2, …, q). The second target recognition area output by the target recognition model M _ D adopts a bounding box BiRepresenting, perturbing the recognition model M _ DoThe output first target identification area adopts a bounding box BoIndicating, electronic device statistics M _ DoAnd the overlapping rate IOU of the output bounding box and Bi.
S322, the identification probability of the first target identification area is adjusted by using the overlapping rate, and the target identification probability of each first target identification area corresponding to each target category is obtained.
The electronic device compares the overlap ratio calculated in the above step S321 with the threshold IOUthresholdAnd comparing, and adjusting the recognition probability of the first target recognition area based on the comparison result. Specifically, when the calculated overlap ratio is greater than the threshold IOUthresholdWhen, M _ DoBounding box B of the outputoThe corresponding recognition probability value is denoted as PBo(c) Wherein, PB iso(c) And outputting the identification probability corresponding to the category c for the disturbance target identification model. When with BiThe overlapping rate IOU is less than the threshold value IOUthresholdThen M _ D will beoBounding box B of the outputoThe corresponding recognition probability value is denoted as pbo (c) ═ 0, and PB is obtained therefromo(c)(o＝1，2，…，q)。
Here, the target recognition probabilities are each that each first target recognition area corresponds to each target category.
S323, carrying out statistical analysis on the target identification probability corresponding to each disturbed target identification model, and determining a first uncertainty characteristic vector.
The statistical analysis may be to calculate statistical information such as a mean value or a variance of the target recognition probabilities corresponding to each disturbed target recognition model in all target categories. And determining a first uncertainty characteristic vector by using the statistical information obtained by calculation, wherein each element in the vector is a specific numerical value of each statistical information.
As an optional implementation manner of this embodiment, the foregoing S323 may include the following steps:
(1) calculating a statistical value of the target identification probability corresponding to each target category by using the target identification probability corresponding to each disturbed target identification model, wherein the statistical value of the target identification probability comprises at least one of entropy, standard deviation, root mean square error, range and average absolute deviation;
for example, the statistical value of the target recognition probability may be any combination of the above 5 statistical values, and is not limited herein. In the following description, a detailed description is given taking an example in which the first uncertain feature vector includes the statistical value in 5 above.
Taking target detection as an example, the first uncertain feature vector is Vd ═ Vd1, Vd2, Vd3, Vd4, Vd5 ]:
Figure BDA0002920008820000141
Figure BDA0002920008820000142
Figure BDA0002920008820000143
vd4＝max(PBo(c))-min(PBo(c))
Figure BDA0002920008820000144
wherein the content of the first and second substances,
Figure BDA0002920008820000145
is PBo(c) (o ═ 1, 2, …, q).
(2) And determining a first uncertainty characteristic vector corresponding to each target class based on the statistic value of the target identification probability corresponding to each target class.
The electronic device obtains, according to the calculated statistical value, a first uncertainty feature vector Vd ═ Vd [ Vd1, Vd2, Vd3, Vd4, Vd5] corresponding to each target class.
Because the statistical value of the target identification probability can accurately measure the heterogeneity or the non-uniformity of the data, the statistical value of the target identification probability is used for calculating the first uncertainty characteristic vector corresponding to each target category, and the reliability of the calculation result can be ensured.
In some optional implementations of this embodiment, when the target recognition model is used for target segmentation, the above S32 may include the following steps:
(1) and calculating the average value of the probabilities that all the pixels in the second target identification region belong to each target class to obtain the target identification probability of the first target identification region corresponding to each target class.
Taking target segmentation as an example, q retrained disturbed target recognition models M _ S can be obtainedo(o ═ 1, 2, …, q). Obtaining M _ SoThe output segmentation graph is used for finding the segmentation result R output by the target segmentation model M _ S on the segmentation graphcRegion, calculating R in the segmentation mapcThe average value PS of the probabilities that all pixels in the region belong to the c-th object classo(c)(o＝1，2，…，q)。
(2) And carrying out statistical analysis on the target recognition probability corresponding to each disturbed target recognition model to determine a first uncertainty characteristic vector.
The specific statistical analysis method is similar to the above statistical analysis method for target detection, and specifically, the first uncertain feature vector may be expressed as: vs ═ Vs [ Vs1, Vs2, Vs3, Vs4, Vs5 ]:
Figure BDA0002920008820000151
Figure BDA0002920008820000152
Figure BDA0002920008820000153
vs4＝max(PSo(c))-min(PSo(c))
Figure BDA0002920008820000154
wherein the content of the first and second substances,
Figure BDA0002920008820000155
is the mean value of pso (c) (o ═ 1, 2, …, q).
And S33, training the false positive recognition model according to the first uncertainty characteristic vector and the label of the target image, and determining the target false positive recognition model.
The reason for uncertainty is adopted in the inventive embodiments: for an object which is easy to identify correctly, even if some disturbance exists, the probability of the network predicting the object does not change greatly, but for an object which is identified incorrectly (namely false positive), under the condition of disturbance, the object identification result of the network is unstable, and the probability value is reflected to have large fluctuation. And (3) adopting a statistical index representing the non-uniformity degree of the data to quantify the fluctuation degree, defining the group of statistical indexes as uncertainty characteristic vectors, and then training a model based on the uncertainty characteristic vectors so as to identify false positive results.
Specifically, the above S33 may include the following steps:
and S331, inputting the first uncertainty characteristic vector into a false positive identification model to obtain a prediction classification result.
The electronic device receives the first uncertainty feature vector calculated in S32 as an input to the false positive recognition model, and inputs the first uncertainty feature vector to the false positive recognition model to obtain a prediction classification result. As described above, if the false positive recognition model is the classifier model, the predicted classification result is whether the recognition result of the target recognition model is a false positive.
S332, adjusting parameters of the false positive identification model by using the label of the target image and the prediction classification result to determine the target false positive identification model.
For object detection, the label is bounding box BiWhether it overlaps with object c; for object segmentation, the label is the segmentation result RcWhether the region belongs to object c. After training is complete, a recognition bounding box B can be constructediModel Md of whether it is false positiveFPOr identifying the segmentation result RcModel Ms of whether a region is false positiveFP。
It should be noted here that, corresponding to each target class, the training may be performed by using the first uncertainty feature vector corresponding to each target class; or combining all the first uncertainty feature vectors corresponding to each target class into an uncertain feature matrix, training a false positive recognition model by using the uncertain feature matrix, and the like, and specifically performing corresponding setting according to actual conditions.
According to the training method for the false positive identification model, the identification probability of the identification result corresponding to the disturbance identification model is adjusted by using the disturbance target identification model and the overlapping rate corresponding to the identification result of the target identification model, so that the identification probability can be changed along with the change of the overlapping rate, and the reliability of the false positive identification model is further ensured. Furthermore, the uncertain feature vectors are used for training the false positive recognition model, so that the false positive rate can be reduced, and the accuracy rate of target recognition is finally improved.
In accordance with an embodiment of the present invention, there is provided an object recognition method embodiment, it is noted that the steps illustrated in the flowchart of the drawings may be performed in a computer system such as a set of computer-executable instructions and that, although a logical order is illustrated in the flowchart, in some cases, the steps illustrated or described may be performed in an order different than here.
In this embodiment, a target recognition method is provided, which can be used in electronic devices, such as a computer, a mobile phone, a tablet computer, and the like, fig. 4 is a flowchart of a training method of a false positive recognition model according to an embodiment of the present invention, and as shown in fig. 4, the flowchart includes the following steps:
and S41, acquiring the image to be recognized.
The image to be recognized may be obtained by the electronic device from the outside, or may be stored in the electronic device, and the manner in which the electronic device obtains the image to be recognized is not limited.
And S42, respectively inputting the image to be recognized into the target recognition model and the at least one disturbed target recognition model to respectively obtain a third recognition result and a fourth recognition result of the image to be recognized.
The disturbed target recognition model is obtained by training after random disturbance processing is carried out on output characteristic information of a preset convolution layer in the target recognition model.
The step is similar to the obtaining manner of the first recognition result and the second recognition result in the above embodiments, and please refer to the above description for details, which is not repeated herein.
S43, a second uncertainty feature vector is determined based on the relationship between the third recognition result and the fourth recognition result.
This step is similar to the determination method of the first uncertainty feature vector in the above embodiment, and please refer to the above description for details, which is not repeated herein.
And S44, inputting the second uncertainty characteristic vector into the false positive recognition model, screening the third recognition result, and determining the target recognition result.
The false positive recognition model is obtained by training according to the training method of the false positive recognition model in the above embodiment of the invention.
And the electronic equipment inputs the second uncertainty characteristic vector corresponding to the image to be recognized into the false positive recognition model, and screens the third recognition result to determine a target recognition result.
Specifically, the second uncertainty feature vector corresponds to the recognition result corresponding to each target category in the image to be recognized, so that each recognition result of different target categories can be represented by using the second uncertainty feature vector, and the false positive recognition model is used for screening the false positive recognition result, so that the recognition result can be screened to obtain the target recognition result.
For example, the electronic device inputs the second uncertainty feature vector of each bounding box or segmented result region into the established false positive recognition model Md after obtaining the second uncertainty feature vectorFPOr MsFPJudgment of bounding Box BiOr the segmentation result RcAnd if the area belongs to false positive, removing the result, and improving the detection or segmentation accuracy after removing the false positive result.
According to the target identification method provided by the embodiment, the false positive identification model is used for screening the target identification result, so that the false positive rate can be reduced, the identification accuracy is improved, and further optimization of the identification result of the target identification model is realized; the method has strong universality and can be applied to all target identification based on the convolutional neural network.
In this embodiment, a training apparatus for a false positive recognition model or a target recognition apparatus is further provided, and the apparatus is used to implement the foregoing embodiments and preferred embodiments, which have already been described and are not described again. As used below, the term "module" may be a combination of software and/or hardware that implements a predetermined function. Although the means described in the embodiments below are preferably implemented in software, an implementation in hardware, or a combination of software and hardware is also possible and contemplated.
The present embodiment provides a training apparatus for a false positive recognition model, as shown in fig. 5, including:
a first obtaining module 51, configured to obtain a target recognition model and obtain at least one disturbed target recognition model, where the disturbed target recognition model is obtained by training output characteristic information of a preset convolutional layer in the target recognition model after random disturbance processing;
a first determining module 52, configured to determine a first uncertainty feature vector based on a relationship between a first recognition result of each disturbed target recognition model on a target image and a second recognition result of the target recognition model on the target image;
and the training module 53 is configured to train the false positive recognition model according to the first uncertainty feature vector and the label of the target image, so as to determine the target false positive recognition model.
According to the training device for the false positive identification model, on the basis that the structure of the target identification model is not changed, random disturbance processing is performed on output feature information generated after convolution, a first identification result under a random disturbance condition is obtained, uncertainty feature vectors of the identification result can be quantitatively evaluated based on the first identification result, the false positive identification model is trained by using the uncertainty feature vectors, the false positive identification model which is identified more accurately can be obtained, the identification result is subsequently screened by using the false positive identification model, and the accuracy of the identification result can be improved.
The present embodiment further provides an object recognition apparatus, as shown in fig. 6, including:
the second obtaining module 61 is used for obtaining an image to be identified;
the recognition module 62 is configured to input the image to be recognized into a target recognition model and at least one disturbed target recognition model respectively, so as to obtain a third recognition result and a fourth recognition result of the image to be recognized respectively, where the disturbed target recognition model is obtained by training after performing random disturbance processing on output characteristic information of a preset convolutional layer in the target recognition model;
a second determining module 63, configured to determine a second uncertainty feature vector based on a relationship between the third recognition result and the fourth recognition result;
a screening module 64, configured to input the second uncertainty feature vector into the false positive recognition model, screen the third recognition result, and determine a target recognition result, where the false positive recognition model is obtained by training according to the training method of the false positive recognition model in any of the above embodiments of the present invention.
The target identification device provided by the embodiment utilizes the false positive identification model to screen the target identification result, so that the false positive rate can be reduced, the identification accuracy is improved, and the further optimization of the identification result of the target identification model is realized; the method has strong universality and can be applied to all target identification based on the convolutional neural network.
The training device of the false positive recognition model, or the target recognition device in this embodiment, is presented in the form of a functional unit, where the unit refers to an ASIC circuit, a processor and a memory executing one or more software or fixed programs, and/or other devices that can provide the above-mentioned functions.
Further functional descriptions of the modules are the same as those of the corresponding embodiments, and are not repeated herein.
An embodiment of the present invention further provides an electronic device, which has the training apparatus for the false positive recognition model shown in fig. 5 or the target recognition apparatus shown in fig. 6.
Referring to fig. 7, fig. 7 is a schematic structural diagram of an electronic device according to an alternative embodiment of the present invention, and as shown in fig. 7, the electronic device may include: at least one processor 71, such as a CPU (Central Processing Unit), at least one communication interface 73, memory 74, at least one communication bus 72. Wherein a communication bus 72 is used to enable the connection communication between these components. The communication interface 73 may include a Display (Display) and a Keyboard (Keyboard), and the optional communication interface 73 may also include a standard wired interface and a standard wireless interface. The Memory 74 may be a high-speed RAM Memory (volatile Random Access Memory) or a non-volatile Memory (non-volatile Memory), such as at least one disk Memory. The memory 74 may alternatively be at least one memory device located remotely from the processor 71. Wherein the processor 71 may be in connection with the apparatus described in fig. 5 or 6, an application program is stored in the memory 74, and the processor 71 calls the program code stored in the memory 74 for performing any of the above-mentioned method steps.
The communication bus 72 may be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus. The communication bus 72 may be divided into an address bus, a data bus, a control bus, and the like. For ease of illustration, only one thick line is shown in FIG. 7, but this is not intended to represent only one bus or type of bus.
The memory 74 may include a volatile memory (RAM), such as a random-access memory (RAM); the memory may also include a non-volatile memory (english: non-volatile memory), such as a flash memory (english: flash memory), a hard disk (english: hard disk drive, abbreviated: HDD) or a solid-state drive (english: SSD); the memory 74 may also comprise a combination of memories of the kind described above.
The processor 71 may be a Central Processing Unit (CPU), a Network Processor (NP), or a combination of CPU and NP.
The processor 71 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a Programmable Logic Device (PLD), or a combination thereof. The PLD may be a Complex Programmable Logic Device (CPLD), a field-programmable gate array (FPGA), a General Array Logic (GAL), or any combination thereof.
Optionally, the memory 74 is also used for storing program instructions. The processor 71 may call program instructions to implement a training method for a false positive recognition model as shown in the embodiments of fig. 1 to 3 of the present application, or a target recognition method as shown in the embodiment of fig. 4.
Embodiments of the present invention further provide a non-transitory computer storage medium, where the computer storage medium stores computer-executable instructions, and the computer-executable instructions may execute the training method of the false positive recognition model or the target recognition method in any of the above method embodiments. The storage medium may be a magnetic Disk, an optical Disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a Flash Memory (Flash Memory), a Hard Disk (Hard Disk Drive, abbreviated as HDD), a Solid State Drive (SSD), or the like; the storage medium may also comprise a combination of memories of the kind described above.
Although the embodiments of the present invention have been described in conjunction with the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations fall within the scope defined by the appended claims.

Claims (12)

1. A training method for a false positive recognition model is characterized by comprising the following steps:
acquiring a target recognition model and at least one disturbed target recognition model, wherein the disturbed target recognition model is obtained by training after random disturbance processing is carried out on output characteristic information of a preset convolution layer in the target recognition model;
determining a first uncertainty characteristic vector based on the relation between a first recognition result of each disturbed target recognition model to a target image and a second recognition result of the target recognition model to the target image;
and training the false positive recognition model according to the first uncertainty characteristic vector and the label of the target image to determine the target false positive recognition model.
2. The training method of claim 1, wherein the obtaining at least one perturbed target recognition model comprises:
inputting a sample image into the target identification model, and extracting output characteristic information of a preset convolution layer in the target identification model;
randomly selecting a plurality of disturbance coordinate points;
processing the pixel value corresponding to the disturbance coordinate point in the output characteristic information by using the disturbance coordinate point to obtain a disturbance characteristic diagram;
and training the target recognition model based on the disturbance characteristic diagram to obtain the disturbance target recognition model.
3. The training method according to claim 2, wherein the processing, by using the disturbance coordinate point, a pixel value corresponding to the disturbance coordinate point in the output feature information to obtain a disturbance feature map includes:
determining a disturbance area by taking each disturbance coordinate point as a center;
and setting the pixel value corresponding to the disturbance area in the output characteristic information as a preset value to obtain the disturbance characteristic graph.
4. The training method according to claim 1, wherein the determining a first uncertainty feature vector based on a relationship between a first recognition result of the perturbed target recognition model on a target image and a second recognition result of the target recognition model on the target image comprises:
corresponding to each disturbed target recognition model, calculating the overlapping rate of each first target recognition area in the disturbed target recognition model and the corresponding second target recognition area in the target recognition model;
adjusting the recognition probability of the first target recognition area by using the overlapping rate to obtain the target recognition probability of each first target recognition area corresponding to each target category;
and carrying out statistical analysis on the target identification probability corresponding to each disturbance target identification model to determine the first uncertainty characteristic vector.
5. The training method according to claim 1, wherein the determining a first uncertainty feature vector based on a relationship between a first recognition result of the perturbed target recognition model on a target image and a second recognition result of the target recognition model on the target image comprises:
calculating the average value of the probabilities that all pixels in the second target identification region belong to each target category to obtain the target identification probability of the first target identification region corresponding to each target category;
and carrying out statistical analysis on the target identification probability corresponding to each disturbance target identification model to determine the first uncertainty characteristic vector.
6. The training method according to claim 4 or 5, wherein the performing a statistical analysis on the target recognition probability corresponding to each of the disturbed target recognition models to determine the first uncertainty feature vector comprises:
calculating a statistical value of the target identification probability corresponding to each target category by using the target identification probability corresponding to each disturbed target identification model, wherein the statistical value of the target identification probability comprises at least one of entropy, standard deviation, root mean square error, range and average absolute deviation;
and determining a first uncertainty characteristic vector corresponding to each target class based on the statistic value of the target identification probability corresponding to each target class.
7. The training method according to claim 1, wherein the training the false positive recognition model according to the first uncertainty feature vector and the label of the target image to determine a target false positive recognition model comprises:
inputting the first uncertainty characteristic vector into the false positive identification model to obtain a prediction classification result;
and adjusting parameters of the false positive identification model by using the label of the target image and the prediction classification result to determine the target false positive identification model.
8. An object recognition method, characterized in that the recognition method comprises:
acquiring an image to be identified;
respectively inputting the images to be recognized into a target recognition model and at least one disturbed target recognition model to respectively obtain a third recognition result and a fourth recognition result of the images to be recognized, wherein the disturbed target recognition model is obtained by training after random disturbance processing is carried out on output characteristic information of a preset convolution layer in the target recognition model;
determining a second uncertainty feature vector based on a relationship of the third recognition result and the fourth recognition result;
inputting the second uncertainty feature vector into the false positive recognition model, screening the third recognition result, and determining a target recognition result, wherein the false positive recognition model is obtained by training according to the training method of the false positive recognition model as claimed in any one of claims 1 to 7.
9. A training apparatus for a false positive recognition model, the training apparatus comprising:
the system comprises a first acquisition module, a second acquisition module and a third acquisition module, wherein the first acquisition module is used for acquiring a target recognition model and acquiring at least one disturbed target recognition model, and the disturbed target recognition model is obtained by training after random disturbance processing is carried out on output characteristic information of a preset convolution layer in the target recognition model;
the first determining module is used for determining a first uncertainty characteristic vector based on the relation between a first recognition result of each disturbed target recognition model to a target image and a second recognition result of the target recognition model to the target image;
and the training module is used for training the false positive recognition model according to the first uncertainty characteristic vector and the label of the target image to determine the target false positive recognition model.
10. An object recognition apparatus, characterized in that the recognition apparatus comprises:
the second acquisition module is used for acquiring an image to be identified;
the recognition module is used for respectively inputting the images to be recognized into a target recognition model and at least one disturbed target recognition model to respectively obtain a third recognition result and a fourth recognition result of the images to be recognized, and the disturbed target recognition model is obtained by training after random disturbance processing is carried out on output characteristic information of a preset convolution layer in the target recognition model;
a second determining module, configured to determine a second uncertainty feature vector based on a relationship between the third recognition result and the fourth recognition result;
a screening module, configured to input the second uncertainty feature vector into the false positive recognition model, screen the third recognition result, and determine a target recognition result, where the false positive recognition model is obtained by training according to the training method of the false positive recognition model according to any one of claims 1 to 7.
11. An electronic device, comprising:
a memory and a processor, the memory and the processor being communicatively connected to each other, the memory having stored therein computer instructions, the processor executing the computer instructions to perform the training method of the false positive recognition model according to any one of claims 1 to 7 or the target recognition method according to claim 8.
12. A computer-readable storage medium storing computer instructions for causing a computer to perform the training method of the false positive recognition model according to any one of claims 1 to 7 or the target recognition method according to claim 8.
CN202110113779.8A 2021-01-27 2021-01-27 Training method of false positive recognition model, target recognition method and device Active CN112784494B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110113779.8A CN112784494B (en) 2021-01-27 2021-01-27 Training method of false positive recognition model, target recognition method and device

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110113779.8A CN112784494B (en) 2021-01-27 2021-01-27 Training method of false positive recognition model, target recognition method and device

Publications (2)

Publication Number Publication Date
CN112784494A true CN112784494A (en) 2021-05-11
CN112784494B CN112784494B (en) 2024-02-06

Family

ID=75759170

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110113779.8A Active CN112784494B (en) 2021-01-27 2021-01-27 Training method of false positive recognition model, target recognition method and device

Country Status (1)

Country Link
CN (1) CN112784494B (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113408650A (en) * 2021-07-12 2021-09-17 厦门大学 Semi-supervised three-dimensional shape recognition method based on consistency training
CN114638829A (en) * 2022-05-18 2022-06-17 安徽数智建造研究院有限公司 Anti-interference training method of tunnel lining detection model and tunnel lining detection method

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170061246A1 (en) * 2015-09-02 2017-03-02 Fujitsu Limited Training method and apparatus for neutral network for image recognition
CN110070034A (en) * 2019-04-19 2019-07-30 北京朗镜科技有限责任公司 Model training method, section recognition methods, device, equipment and medium
CN111062871A (en) * 2019-12-17 2020-04-24 腾讯科技（深圳）有限公司 Image processing method and device, computer equipment and readable storage medium
CN111401396A (en) * 2019-01-03 2020-07-10 阿里巴巴集团控股有限公司 Image recognition method and device
CN111523597A (en) * 2020-04-23 2020-08-11 北京百度网讯科技有限公司 Target recognition model training method, device, equipment and storage medium
WO2020215557A1 (en) * 2019-04-24 2020-10-29 平安科技（深圳）有限公司 Medical image interpretation method and apparatus, computer device and storage medium
CN111914939A (en) * 2020-08-06 2020-11-10 平安科技（深圳）有限公司 Method, device and equipment for identifying blurred image and computer readable storage medium
CN112116613A (en) * 2020-09-25 2020-12-22 贝壳技术有限公司 Model training method, image segmentation method, image vectorization method and system thereof

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170061246A1 (en) * 2015-09-02 2017-03-02 Fujitsu Limited Training method and apparatus for neutral network for image recognition
CN111401396A (en) * 2019-01-03 2020-07-10 阿里巴巴集团控股有限公司 Image recognition method and device
CN110070034A (en) * 2019-04-19 2019-07-30 北京朗镜科技有限责任公司 Model training method, section recognition methods, device, equipment and medium
WO2020215557A1 (en) * 2019-04-24 2020-10-29 平安科技（深圳）有限公司 Medical image interpretation method and apparatus, computer device and storage medium
CN111062871A (en) * 2019-12-17 2020-04-24 腾讯科技（深圳）有限公司 Image processing method and device, computer equipment and readable storage medium
CN111523597A (en) * 2020-04-23 2020-08-11 北京百度网讯科技有限公司 Target recognition model training method, device, equipment and storage medium
CN111914939A (en) * 2020-08-06 2020-11-10 平安科技（深圳）有限公司 Method, device and equipment for identifying blurred image and computer readable storage medium
CN112116613A (en) * 2020-09-25 2020-12-22 贝壳技术有限公司 Model training method, image segmentation method, image vectorization method and system thereof

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
陈伟;何家欢;裴喜平;: "基于相空间重构和卷积神经网络的电能质量扰动分类", 电力系统保护与控制, no. 14 *

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113408650A (en) * 2021-07-12 2021-09-17 厦门大学 Semi-supervised three-dimensional shape recognition method based on consistency training
CN113408650B (en) * 2021-07-12 2023-07-18 厦门大学 Semi-supervised three-dimensional shape recognition method based on consistency training
CN114638829A (en) * 2022-05-18 2022-06-17 安徽数智建造研究院有限公司 Anti-interference training method of tunnel lining detection model and tunnel lining detection method

Also Published As

Publication number Publication date
CN112784494B (en) 2024-02-06

Similar Documents

Publication Publication Date Title
CN110363182B (en) Deep learning-based lane line detection method
CN108470172B (en) Text information identification method and device
CN110060237B (en) Fault detection method, device, equipment and system
WO2020164282A1 (en) Yolo-based image target recognition method and apparatus, electronic device, and storage medium
KR20210110823A (en) Image recognition method, training method of recognition model, and related devices and devices
CN111652869B (en) Slab void identification method, system, medium and terminal based on deep learning
CN111860439A (en) Unmanned aerial vehicle inspection image defect detection method, system and equipment
CN111274926B (en) Image data screening method, device, computer equipment and storage medium
CN112132206A (en) Image recognition method, training method of related model, related device and equipment
CN112419202B (en) Automatic wild animal image recognition system based on big data and deep learning
CN112784494B (en) Training method of false positive recognition model, target recognition method and device
CN111553438A (en) Image identification method based on convolutional neural network
CN109598298B (en) Image object recognition method and system
CN111368636A (en) Object classification method and device, computer equipment and storage medium
CN111310826A (en) Method and device for detecting labeling abnormity of sample set and electronic equipment
CN112861917A (en) Weak supervision target detection method based on image attribute learning
CN116030237A (en) Industrial defect detection method and device, electronic equipment and storage medium
CN111753929A (en) Artificial intelligence based classification method, device, terminal and storage medium
CN115880260A (en) Method, device and equipment for detecting base station construction and computer readable storage medium
CN111597875A (en) Traffic sign identification method, device, equipment and storage medium
CN113487610A (en) Herpes image recognition method and device, computer equipment and storage medium
CN117036737A (en) Feature extraction and matching method based on information entropy, GMS and LC significant detection
WO2023160666A1 (en) Target detection method and apparatus, and target detection model training method and apparatus
CN114973300B (en) Component type identification method and device, electronic equipment and storage medium
CN114663731B (en) Training method and system of license plate detection model, and license plate detection 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