CN110826478A - Aerial photography illegal building identification method based on countermeasure network - Google Patents

Aerial photography illegal building identification method based on countermeasure network Download PDF

Info

Publication number
CN110826478A
CN110826478A CN201911063291.8A CN201911063291A CN110826478A CN 110826478 A CN110826478 A CN 110826478A CN 201911063291 A CN201911063291 A CN 201911063291A CN 110826478 A CN110826478 A CN 110826478A
Authority
CN
China
Prior art keywords
network
target
detector
feature
transformation
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.)
Pending
Application number
CN201911063291.8A
Other languages
Chinese (zh)
Inventor
宫法明
徐晨曦
李昕
杨天濠
刘芳华
袁咪咪
唐昱润
司朋举
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.)
China University of Petroleum East China
Original Assignee
China University of Petroleum East China
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 China University of Petroleum East China filed Critical China University of Petroleum East China
Priority to CN201911063291.8A priority Critical patent/CN110826478A/en
Publication of CN110826478A publication Critical patent/CN110826478A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/176—Urban or other man-made structures
    • 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
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/07—Target detection

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)

Abstract

The invention relates to an aerial photography default construction identification method based on a countermeasure network, and belongs to the technical field of target detection and identification in deep learning. The method aims to solve the problems that the ground deformation field training samples caused by aerial photography are few, and the deformation field is difficult to identify due to poor robustness of a target detector. While general antagonistic networks are usually trained to produce a good generator for realistic images by antagonistic learning, our networks are doing the opposite task, and by competing against the network, we hopefully train a better detector robust to deformation. The capacity of identifying deformation samples is improved through antagonistic learning, and the detection accuracy is improved through the antagonistic learning strategy, so that the performance is substantially improved compared with that of a standard Fast-RCNN network.

Description

Aerial photography illegal building identification method based on countermeasure network
Technical Field
The invention belongs to the field of computer graphics and image processing, and particularly relates to an aerial photography default identification method based on a countermeasure network.
Background
The speed of city construction is increasing day by day, and a lot of problems are brought to city planning and management work: the driving and management of benefits generate a large amount of illegal construction, the large-scale land occupation of the urban and rural junctions is violated, the construction of cultural relics is destroyed, the large-scale demolition and large construction of cultural relics are destroyed, the invisible planning is unauthorized to construct, and the like, and the existence of the illegal construction has a plurality of hazards to urban planning and construction. Aiming at the illegal construction detection, the traditional aerial photography mode (such as satellite remote sensing and common aerial remote sensing) has the problems of high information acquisition cost, long data acquisition period, lack of maneuvering flexibility and the like, and is not suitable for short-term and high-frequency urban dynamic monitoring research. The flight mode of the unmanned aerial vehicle has the advantages of short data acquisition period, flexible acquisition of high-resolution image data, free and flexible flight time, relatively low one-time investment cost and use cost, easy operation and maintenance of equipment and the like. Gradually becoming the current default building detection preferred equipment.
With the continuous development of image processing technology and deep learning technology, the target detector is trained by utilizing a deep learning framework, and the purpose identification of the aerial video is very meaningful research. At the initial stage of model training, the problem of field deformation caused by high-altitude oblique shooting is found out when relevant personnel work and the camera shooting is oblique shooting in order to improve the working efficiency. That for learning a target detector that is invariant to deformation, our current solution is to use a data-driven strategy, i.e. to collect large-scale data sets with target instances under different conditions. However, in practical situations, the data set cannot cover all situations, which causes a problem of difficult recognition of the deformation field.
Disclosure of Invention
In order to solve the problems, the invention provides an aerial photography default identification method based on a countermeasure network, which improves the identification accuracy of a target detector on a deformation target by means of a countermeasure learning strategy.
The method comprises the following specific steps:
s1, dividing a target video into pictures according to frames by screening videos shot by the unmanned aerial vehicle in work, manually labeling the target, and properly reducing the resolution of the pictures to improve the calculation speed;
s2, taking the training set as the input of the convolution network of Fast-RCNN, obtaining a basic target detector as a discriminator in the countermeasure network through training;
s3, using a warped network, the key idea is to create a warp on the target feature and make the target recognition of the detector difficult, we use an STN spatial transform network, which contains three parts: a local network, a grid generator and a sampler;
s4, the input feature map, STN spatial transform network will estimate the amount to be deformed (e.g., rotation degree, translation distance and affine transform), these variables will be used as input to the grid generator and sampler on the feature map, the output is the deformed feature map;
s5, inputting the deformed characteristic diagram into a discriminator, namely a pre-trained basic target detector for discrimination, competing with the networks and overcoming obstacles, and processing deformation by Fast-RCNN learning in a robust mode;
s6, the discriminator feeds back the recognition result to the generator, and updates the generator parameters, so that the generator can generate a more useful deformation characteristic diagram;
s7, a more robust target detector is obtained through the antagonistic learning.
The technical scheme of the invention is characterized by comprising the following steps:
for step S2, the convolution network of Fast-RCNN takes the whole image as input and produces as output a convolution feature map, since the operation is mainly convolution and max pooling, the spatial size of the output feature map will vary according to the size of the input image, given a feature map, the RoI-posing layer is used to project candidate regions onto the feature space, the RoI pooling layer will crop and resize each target candidate region to generate a fixed-size feature vector, and then pass these feature vectors through the fully-connected layer whose output is the probability and bounding box coordinates of each target class including the background class, for training, SoftMax Loss and Regression Loss are applied to these two outputs, respectively, and back-propagated to perform end-to-end learning, the total Loss function is as follows:
L(p,u,tu,v)＝Lcls(p,u)+λ[u≥1]Lcls(tu,v) (1)
in formula (1), the total loss function is divided into a weighted sum of classification loss and regression positioning loss function, and p represents the discrete probability of each RoI; u represents the category of true correct tags; t represents the regression compensation of the true correct label; v represents the bounding box regression target for the true correct label; λ is a hyper-parameter, which is specified before training, indicating the degree of importance of classification and regression;
for step S4, the operation mechanism of the space transformer adopted in the present invention can be divided into three parts, namely, local Network (localization Network), Grid generator (Grid generator), and Sampler (Sampler). The local network is a network used for regression transformation parameter theta, the local network inputs the characteristic image and outputs spatial transformation parameters through a series of hidden network layers, the form of theta can be various, if 2D affine transformation is required to be realized, the theta is the output of a 6-dimensional (2x3) vector; the size of θ depends on the type of transform; a Grid Generator (Grid Generator) constructs a sampling Grid according to predicted transformation parameters, calculates the coordinate position of each coordinate in the original feature map U in the transformed target feature V in a matrix operation mode, namely generates T (G), and the corresponding relation can be written as an expression (2), namely the affine transformation of the picture can be written as an expression (3); the sampler uses the sampling grid and the input characteristic diagram as input to generate output at the same time, and obtains the result of the characteristic diagram after conversion;
Figure BDA0002257446210000031
in the formula (2)
Figure BDA0002257446210000032
Is the coordinate corresponding to each pixel in the feature map U after transformation,
Figure BDA0002257446210000033
is the coordinate, T, corresponding to each pixel in the original target feature VθAs a two-dimensional affine transformation function, matrix AθA parameter matrix output for the local network;
Figure BDA0002257446210000034
in the formula (3)Is the coordinate corresponding to each pixel in the feature map U after transformation,
Figure BDA0002257446210000036
the coordinate is corresponding to each pixel in the original target feature V, and the coefficient matrix theta is an affine transformation coefficient.
For step S5, the challenge network must learn features that are predicted to be misdetected by the detector, and we train this challenge network by the following loss function:
L＝-Lsoftmax(Fc(A(X)),C) (4)
in formula (4), A (X) represents STN network, X is the calculated feature on the image, FcThe output representing class probability, C is the true class of X, so if features generated by the challenge network are easily classified by the detector, they will get high loss for the STN network, on the other hand, if after the challenge feature generation, it is difficult for the detector to classify, we get high loss for the detector, and the loss for the STN network is low.
For step S7, our network is built on a space transformation network, which was proposed to deform features to make classification easier in the original work, while our network is doing the opposite task, by competing with our Fast-RCNN network, we can train a better detector robust to deformation.
The invention has the beneficial effects that:
(1) the invention provides an automatic detection and identification method for the identification work of the illegal site by using the aerial photography of the unmanned aerial vehicle, and can improve the working efficiency and liberate manpower.
(2) Aiming at the inherent deformation problem existing in unmanned aerial vehicle shooting, the invention provides a method for generating deformation samples by using an antagonistic network for training, and solves the problem of less deformation samples in data concentration, thereby obtaining a detector with robustness.
Drawings
FIG. 1 is a flow chart of an aerial photography violation identification method based on a countermeasure network according to the present invention.
FIG. 2 is a network structure diagram of an aerial photography violation identification method based on a countermeasure network.
Detailed Description
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 only a part of the embodiments of the present invention, and not all of the embodiments. 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.
As shown in the first figure, the present invention provides an aerial photography default identification method based on countermeasure network, which is an implementation flowchart of the present invention, and the method includes:
s1, dividing a target video into pictures according to frames by screening videos shot by the unmanned aerial vehicle in work, manually labeling the target, and properly reducing the resolution of the pictures to improve the calculation speed;
s2, the training set is used as the input of the convolution network of Fast-RCNN, the basic target detector is obtained through training and is used as the discriminator in the countermeasure network, the convolution network of Fast-RCNN takes the whole image as the input, and the convolution characteristic graph is generated as the output. Since the operations are mainly convolution and maximum pooling, the spatial size of the output feature map will vary depending on the input image size. Given the feature map, the RoI-posing layer is used to project the candidate region onto the feature space. The RoI pooling layer will crop and resize each target candidate region to generate a fixed-size feature vector. These feature vectors are then passed through the fully connected layers. The output of the fully connected layer is the probability and bounding box coordinates of each object class including the background class. For training, SoftMax Loss and Regression Loss are applied to the two outputs, respectively, and propagated backwards to perform end-to-end learning. The overall loss function is as follows:
L(p,u,tu,v)＝Lcls(p,u)+λ[u≥1]Lcls(tu,v) (5)
in the formula (5), the total loss function is divided into a weighted sum of classification loss and regression positioning loss function, and p represents the discrete possibility of each RoI; u represents the category of true correct tags; t represents the regression compensation of the true correct label; v represents the bounding box regression target for the true correct label; λ is a hyper-parameter, which is specified before training, indicating the degree of importance of classification and regression;
s3, using deformation Network, the key idea is to create deformation on the target feature and make the target identification of the detector difficult, we adopt STN space transformation Network, the operation mechanism of a space transformer can be divided into three parts, local Network (localization Network), Grid generator (Grid generator), Sampler (Sampler). (ii) a
S4, the input feature map, STN spatial transformation network will estimate the amount to be deformed (e.g., degree of rotation, translation distance, and affine transformation). These variables will be used as inputs to the grid generator and sampler on the feature map, the output being the warped feature map. The local network is a network for regressing a transformation parameter theta, which is input into a characteristic image and then outputs a spatial transformation parameter through a series of hidden network layers, wherein the form of theta can be various, and if 2D affine transformation is required to be realized, the theta is output of a 6-dimensional (2x3) vector. The size of θ depends on the type of transform; a Grid Generator (Grid Generator) constructs a sampling Grid according to predicted transformation parameters, calculates the coordinate position of each coordinate in the original feature map U in the transformed target feature V in a matrix operation mode, namely generates T (G), and the corresponding relation can be written as an expression (2), namely the affine transformation of the picture can be written as an expression (3); the sampler uses the sampling grid and the input characteristic diagram as input to generate output at the same time, and obtains the result after the characteristic diagram is transformed.
In the formula (6)
Figure BDA0002257446210000052
Is the coordinate corresponding to each pixel in the feature map U after transformation,is the coordinate, T, corresponding to each pixel in the original target feature VθAs a two-dimensional affine transformation function, matrix AθAnd (4) parameter matrix output by the local network.
Figure BDA0002257446210000054
In the formula (7)
Figure BDA0002257446210000055
Is the coordinate corresponding to each pixel in the feature map U after transformation,
Figure BDA0002257446210000056
the coordinate is corresponding to each pixel in the original target feature V, and the coefficient matrix theta is an affine transformation coefficient.
S5, inputting the deformed feature map into a discriminator, namely a pre-trained basic target detector for discrimination, competing with the networks and overcoming obstacles, and processing deformation in a robust mode by the Fast-RCNN society. The countermeasure network must learn features that are predicted to be misdetected by the detector. We train this countermeasure network by the following loss function:
L＝-Lsoftmax(Fc(A(X)),C) (8)
in formula (8), A (X) represents STN network, X is the calculated feature on the image, FcThe output representing class probability, C is the true class of X, so if features generated by the countermeasure network are easily classified by the detector, they will get high loss for the STN network. On the other hand, if after the antagonistic features are generated, it is difficult for the detector to classify, then for the detector we get a high loss, and the loss for the STN network is low.
S6, the discriminator feeds back the recognition result to the generator, and updates the generator parameters, so that the generator generates a more useful deformation feature map, as shown in fig. 2.
S7, a more robust target detector is obtained through the antagonistic learning. As shown in fig. 2, our network is built on a spatial transformation network. In the original work, a spatial transformation network is proposed to deform the features and make the classification easier. While our network is accomplishing the opposite task. By competing with our Fast-RCNN network, we can train a better detector robust to deformation.
The invention discloses an aerial photography default identification method based on a countermeasure network. This antagonism creates examples of different deformations in the training process, making them difficult for the original target detector to recognize. Therefore, the model with invariance to deformation is obtained, and the problem that the recognition accuracy of the ground deformation field is low due to the fact that a training sample is concentrated and a large number of deformation samples are not few is solved.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents, improvements and the like that fall within the spirit and principle of the present invention are intended to be included therein.

Claims (5)

1. An aerial photography default identification method based on a countermeasure network is characterized by comprising the following specific steps:
s1, dividing a target video into pictures according to frames by screening videos shot by the unmanned aerial vehicle in work, manually labeling the target, and properly reducing the resolution of the pictures to improve the calculation speed;
s2, taking the training set as the input of the convolution network of Fast-RCNN, obtaining a basic target detector as a discriminator in the countermeasure network through training;
s3, using a warped network, the key idea is to create a warp on the target feature and make the target recognition of the detector difficult, we use an STN spatial transform network, which contains three parts: a local network, a grid generator and a sampler;
s4, inputting the feature map, the STN space transform network will estimate the amount to be deformed, these variables will be used as input to the grid generator and sampler on the feature map, the output is the deformed feature map;
s5, inputting the deformed characteristic diagram into a discriminator, namely a pre-trained basic target detector for discrimination, competing with the networks and overcoming obstacles, and processing deformation by Fast-RCNN learning in a robust mode;
s6, the discriminator feeds back the recognition result to the generator, and updates the generator parameters, so that the generator can generate a more useful deformation characteristic diagram;
s7, a more robust target detector is obtained through the antagonistic learning.
2. The method for identifying aerial photography violation based on countermeasure network as claimed in claim 1, wherein for step S2, the convolution network of Fast-RCNN of the present invention takes the whole image as input and generates convolution feature map as output, since the operation is mainly convolution and maximum pooling, the spatial size of the output feature map will change according to the size of the input image, given the feature map, the RoI-posing layer is used to project the candidate region onto the feature space; the RoI pooling layer will crop and resize each target candidate region to generate a fixed-size feature vector, and then pass these feature vectors through a fully connected layer whose outputs are the probability and bounding box coordinates of each target class including the background class, to which outputs SoftMax Loss and Regression Loss are applied, respectively, for training, back-propagating to perform end-to-end learning, the total Loss function is as follows:
L(p,u,tu,v)＝Lcls(p,u)+λ[u≥1]Lcls(tu,v) (1)
in formula (1), the total loss function is divided into a weighted sum of classification loss and regression positioning loss function, and p represents the discrete probability of each RoI; u represents the category of true correct tags; t represents the regression compensation of the true correct label; v represents the bounding box regression target for the true correct label; λ is a hyperparameter, which is specified prior to training, indicating the degree of importance of classification and regression.
3. The method for identifying the aerial photography default based on the countermeasure network according to claim 1, wherein for step S4, the operation mechanism of the space transformer adopted by the invention can be divided into three parts, namely a local network (localization network), a Grid generator (Grid generator), and a Sampler (Sampler); the local network is a network used for regression transformation parameter theta, the local network inputs the characteristic image and outputs spatial transformation parameters through a series of hidden network layers, the form of theta can be various, if 2D affine transformation is required to be realized, the theta is the output of a 6-dimensional (2x3) vector; the size of θ depends on the type of transform; a Grid Generator (Grid Generator) constructs a sampling Grid according to predicted transformation parameters, calculates the coordinate position of each coordinate in the original feature map U in the transformed target feature V in a matrix operation mode, namely generates T (G), and the corresponding relation can be written as an expression (2), namely the affine transformation of the picture can be written as an expression (3); the sampler uses the sampling grid and the input characteristic diagram as input to generate output at the same time, and obtains the result of the characteristic diagram after conversion;
Figure FDA0002257446200000021
in the formula (2)
Figure FDA0002257446200000022
Is the coordinate corresponding to each pixel in the feature map U after transformation,is the coordinate, T, corresponding to each pixel in the original target feature VθAs a two-dimensional affine transformation function, matrix AθA parameter matrix output for the local network;
Figure FDA0002257446200000024
in the formula (3)
Figure FDA0002257446200000025
Is the coordinate corresponding to each pixel in the feature map U after transformation,
Figure FDA0002257446200000026
the coordinate is corresponding to each pixel in the original target feature V, and the coefficient matrix theta is an affine transformation coefficient.
4. The method for identifying aerial photography default based on countermeasure network according to claim 1, wherein for step S5, the STN network adopted in the present invention must learn the characteristics predicted to be misdetected by the detector, and we train this countermeasure network by the following loss function:
L＝-Lsoftmax(Fc(A(X)),C) (4)
in the formula (4), A (X) represents STN network, and X is figureLike the feature being calculated, FcThe output representing class probability, C is the true class of X, so if features generated by the countermeasure network are easily classified by the detector, they will get high loss for the STN network; on the other hand, if after the antagonistic features are generated, it is difficult for the detector to classify, then for the detector we get a high loss, and the loss for the STN network is low.
5. The method for recognizing the aerial photography violation based on the countermeasure network as claimed in claim 1, wherein for step S7, the network of the present invention is built on a space transformation network, and in the original work, the space transformation network is proposed to deform the features, so that the classification is easier, while our network is performing the opposite task, and by competing with our Fast-RCNN network, we can train a better detector with robustness to deformation.
CN201911063291.8A 2019-11-01 2019-11-01 Aerial photography illegal building identification method based on countermeasure network Pending CN110826478A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201911063291.8A CN110826478A (en) 2019-11-01 2019-11-01 Aerial photography illegal building identification method based on countermeasure network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201911063291.8A CN110826478A (en) 2019-11-01 2019-11-01 Aerial photography illegal building identification method based on countermeasure network

Publications (1)

Publication Number Publication Date
CN110826478A true CN110826478A (en) 2020-02-21

Family

ID=69552311

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201911063291.8A Pending CN110826478A (en) 2019-11-01 2019-11-01 Aerial photography illegal building identification method based on countermeasure network

Country Status (1)

Country Link
CN (1) CN110826478A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111353548A (en) * 2020-03-11 2020-06-30 中国人民解放军军事科学院国防科技创新研究院 Robust feature deep learning method based on confrontation space transformation network
CN113420749A (en) * 2021-05-25 2021-09-21 广州铁路职业技术学院（广州铁路机械学校） Container number positioning and identifying method

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111353548A (en) * 2020-03-11 2020-06-30 中国人民解放军军事科学院国防科技创新研究院 Robust feature deep learning method based on confrontation space transformation network
CN113420749A (en) * 2021-05-25 2021-09-21 广州铁路职业技术学院（广州铁路机械学校） Container number positioning and identifying method

Similar Documents

Publication Publication Date Title
CN109949316A (en) A kind of Weakly supervised example dividing method of grid equipment image based on RGB-T fusion
CN106909902B (en) Remote sensing target detection method based on improved hierarchical significant model
CN110555420B (en) Fusion model network and method based on pedestrian regional feature extraction and re-identification
CN111368690A (en) Deep learning-based video image ship detection method and system under influence of sea waves
CN109191255B (en) Commodity alignment method based on unsupervised feature point detection
CN112712052A (en) Method for detecting and identifying weak target in airport panoramic video
CN114820655A (en) Weak supervision building segmentation method taking reliable area as attention mechanism supervision
CN111046756A (en) Convolutional neural network detection method for high-resolution remote sensing image target scale features
CN111368637A (en) Multi-mask convolution neural network-based object recognition method for transfer robot
CN114373009A (en) Building shadow height measurement intelligent calculation method based on high-resolution remote sensing image
CN110826478A (en) Aerial photography illegal building identification method based on countermeasure network
CN115100542A (en) Power transmission tower remote sensing target detection method based on semi-supervised learning and deformable convolution
CN117496158A (en) Semi-supervised scene fusion improved MBI contrast learning and semantic segmentation method
Chen et al. Memory-contrastive unsupervised domain adaptation for building extraction of high-resolution remote sensing imagery
Nurunnabi et al. Investigation of pointnet for semantic segmentation of large-scale outdoor point clouds
CN112668662B (en) Outdoor mountain forest environment target detection method based on improved YOLOv3 network
CN114581307A (en) Multi-image stitching method, system, device and medium for target tracking identification
CN114187506A (en) Remote sensing image scene classification method of viewpoint-aware dynamic routing capsule network
CN113052121A (en) Multi-level network map intelligent generation method based on remote sensing image
CN112785629A (en) Aurora motion characterization method based on unsupervised deep optical flow network
CN112465836A (en) Thermal infrared semantic segmentation unsupervised field self-adaption method based on contour information
CN111401203A (en) Target identification method based on multi-dimensional image fusion
CN114882139A (en) End-to-end intelligent generation method and system for multi-level map
Wang et al. Extraction of main urban roads from high resolution satellite images by machine learning
CN115272846A (en) Improved Orientdrcnn-based rotating target detection method

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
WD01 Invention patent application deemed withdrawn after publication
WD01 Invention patent application deemed withdrawn after publication

Application publication date: 20200221