CN113326759A - Uncertainty estimation method for remote sensing image building identification model - Google Patents
Uncertainty estimation method for remote sensing image building identification model Download PDFInfo
- Publication number
- CN113326759A CN113326759A CN202110577026.2A CN202110577026A CN113326759A CN 113326759 A CN113326759 A CN 113326759A CN 202110577026 A CN202110577026 A CN 202110577026A CN 113326759 A CN113326759 A CN 113326759A
- Authority
- CN
- China
- Prior art keywords
- model
- uncertainty
- building
- node
- dropout
- 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
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
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
- G06F18/24155—Bayesian classification
-
- 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
- G06N3/047—Probabilistic or stochastic 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/12—Edge-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10032—Satellite or aerial image; Remote sensing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20076—Probabilistic image processing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Biomedical Technology (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Probability & Statistics with Applications (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses an uncertainty estimation method of a remote sensing image building identification model, which provides a building identification uncertainty estimation method considering both semantic and shape characteristics on the basis of a Bayesian approximation reasoning theory. The method tries to perform reliability quantitative self-evaluation on the recognition result from two aspects of semantics and shape characteristics, thereby helping a user to establish a result confidence-gathering standard and having important significance for promoting the deep application of a deep learning technology in mapping practice and promoting the intelligent development of related industries.
Description
Technical Field
The invention relates to the technical field of surveying and mapping science, in particular to an uncertainty estimation method for a remote sensing image building identification model.
Background
At present, the deep learning model still has limitations on the problem of building identification of high-resolution remote sensing images, and it can be expected that a building identification algorithm based on deep learning still faces a certain precision bottleneck in a long time, which means that a result automatically generated by a machine can be submitted as a product only after manual quality inspection and revision. The workload of the quality inspection process is very laborious, and the quality inspection personnel cannot know the reliability of the automatically generated result, so that all buildings must be inspected one by one. If the algorithm can carry out reliability quantitative description when the result is predicted, namely the confidence of the model to the result is high, the algorithm is helpful for a user to establish a result confidence standard, and a low-reliability result is preferentially captured and repaired in a quality inspection link.
Meanwhile, the area of the urban area in China currently exceeds 20 ten thousand square kilometers, and the corresponding tasks are huge and fussy to make and update vector maps. The method for automatically extracting the high-resolution remote sensing image building vector, which is more intelligent in development and research, has important significance for improving the production efficiency of mapping industry and promoting the intelligent development of the field of remote sensing image interpretation research in the big data era. Therefore, it is necessary to research and develop the uncertainty self-evaluation capability of the existing building identification model, and further improve the model applicability.
Uncertainty estimation is one of the popular research directions in the field of artificial intelligence in recent years, and a Bayesian Neural Network (BNN) is the basis of most deep learning uncertainty estimation methods at present. A bayesian approximation inference method which has gained much attention in recent years is a Monte Carlo drop (MC-drop) algorithm proposed by Gal and the like, and the algorithm has been widely used in the field of computer vision. However, in general, uncertainty estimation of a deep learning model has not been sufficiently focused in the field of remote sensing, and such researches aiming at building identification problems are still rare, and particularly, related researches for performing reliability self-evaluation on building identification results at the vector shape level are lacked.
Disclosure of Invention
In view of the above, the present invention provides an uncertainty estimation method for a remote sensing image building identification model, which is based on bayesian approximate inference theory and method, researches and explores a BNN parameter sampling and approximate inference method suitable for a high-resolution remote sensing image building identification task, further realizes quantitative self-evaluation of reliability of a building identification result from two aspects of pixel segmentation and vector generation, and helps a user to establish a result confidence standard, and specifically includes the following contents:
s1, obtaining approximate posterior distribution of the building identification model parameters by adopting an MC-dropout method;
s2, applying the MC-dropout method to a building segmentation model to estimate the uncertainty of the building segmentation;
s3, applying the MC-dropout method to the shape optimization model, providing a building vector optimization and uncertainty estimation learning framework based on shape modeling, and quantitatively representing uncertainty of the vector contour while optimizing the vector contour.
Further, for the CNN model which completes training by applying dropout, the dropout operation is still reserved in the model prediction stage to realize random sampling of approximate posterior distribution of model parameters; therefore, after the same data is repeatedly input into the CNN model and model prediction is performed for multiple times, the variance of the prediction result quantitatively reflects the uncertainty of the model on the current prediction result, and the average value is regarded as the final prediction result.
Furthermore, the MC-dropout building segmentation model is structurally characterized in that a plurality of dropout operation layers are added to a full convolution network under the current mainstream ResNet and characteristic pyramid structure, and the added dropout operation layers are mainly concentrated in a coding and decoding module in the middle section of the model.
Furthermore, the training mode of the building segmentation model is the same as that of the classical FCN model, namely a cross entropy loss function is constructed by using training data and labels, and the back propagation and random gradient descent algorithm is adopted for optimization solution;
in the testing stage, a certain random dropout proportion is set, N times of prediction are repeatedly carried out on the same input image, N pixel classification probability map results are obtained, a final probability map is obtained by averaging according to pixels, and a segmentation result can be generated after threshold processing is carried out on the final probability map; for each segmentation class, the variance of N probability maps is calculated according to pixels and is used as the uncertainty of each segmentation class, and the total uncertainty self-evaluation result of the model for the current input image can be obtained after the classes are crossed, the number of the pixels is counted and the average is taken.
Further, the process of optimizing the initial contour generated by the building segmentation model S2 by the building vector optimization and uncertainty estimation learning framework based on shape modeling is as follows:
(1) after the repeated prediction semantic feature set of the segmentation model is averaged element by element, performing pooling processing based on the initial contour node coordinates to obtain an image feature vector corresponding to each node;
(2) cascading the two-dimensional coordinates of each node with the corresponding image characteristics, and encrypting the cascading characteristics of each node by a one-dimensional convolution kernel to form a tensor [ f1,f2,…,fn];
(3) Continuously performing convolution processing to improve feature dimensionality, and generating a global feature G through maximum pooling operation;
(4) g and node local feature [ f ]1,f2,…,fn]Respectively cascading, continuously encrypting the local and global fusion features into final point features, and predicting the coordinate correction value of each contour node;
(5) and adding the corrected value and the input coordinate to obtain the optimized contour node coordinate.
Further, the building vector optimization and uncertainty estimation learning framework based on shape modeling establishes a loss function for the profile optimization effect from both node bias and angle bias, namely:
Lpolygon=Lpoint+Lline
adjusting the number of the nodes outputting the optimized contour and the corresponding truth value to be consistent by additionally increasing the sampling points to form a pairing point set S, and then defining a node deviation loss function L by calculating the average distance of the pairing pointspoint: considering that the side lengths of each polygon are different, the same node deviation distance may bring different degrees of shape deformation, so that the angle deviation loss function L is definedlineDriving the output polygon to be parallel to the homonymous side of the true value of the output polygon as much as possible, wherein the loss is defined by the average value of the cosine of the included angle of the homonymous side; wherein the homonymous edge is determined by the pairing point set S;
and respectively calculating the mean value and the variance of the optimized contour one by one according to the nodes, obtaining the final optimized contour result and the uncertainty representation of each node, and obtaining the building contour uncertainty self-evaluation result based on the shape characteristics by taking the uncertainty mean value of all the nodes.
The technical scheme provided by the invention has the beneficial effects that: (1) the uncertainty estimation method of the remote sensing image building identification model can perform reliability quantitative description when predicting the result, can honestly know the data scene beyond the identification capability of the model, and can realize the quantitative self-evaluation of the reliability of the segmentation result and the vector extraction result by the model;
(2) the uncertainty estimation method of the remote sensing image building identification model provided by the invention is beneficial to the establishment of a result confidence standard by production personnel, preferentially captures and repairs and measures a low-reliability result when responding to an emergency demand, and simultaneously improves the industrial situation that the production personnel lack the confidence standard when screening or repairing the automatic extraction result of a machine model.
Drawings
The invention will be further described with reference to the accompanying drawings and examples, in which:
FIG. 1 is a diagram of a MC-dropout-based building segmentation model according to an embodiment of the invention;
FIG. 2 is a diagram of a building vector optimization and uncertainty estimation learning framework according to an embodiment of the invention.
Detailed Description
For a more clear understanding of the technical features, objects and effects of the present invention, embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
Referring to fig. 1, the present invention discloses an uncertainty estimation method for a remote sensing image building identification model, which includes the following steps:
s1, obtaining approximate posterior distribution of CNN parameters of the building identification model by adopting an MC-dropout method, and realizing Bayesian approximation of model parameter distribution;
the main principle of BNN-based uncertainty estimation is to obtain posterior distribution p (ω | X, Y) of neural network model parameters to measure uncertainty of the model on the premise of knowing a training data set X and a corresponding label Y, and since the huge parameters of the neural network determine prior probabilities that different parameter combinations cannot be obtained by conventional means, only approximate inference can be sought. The invention adopts MC-dropout method to obtain approximate posterior distribution of building identification model parameters, and the method realizes Bayesian approximation of model parameter distribution based on mathematical relationship between dropout and Bayesian variational inference.
The main idea of bayesian variational inference is to measure the similarity between the approximate distribution of model parameters and the actual posterior distribution, and further to solve the approximate distribution by minimizing the KL divergence between the approximate variational distribution q (ω) and p (ω | X, Y) of the model parameters, that is:
KL(q(ω)||p(ω∣X,Y))
relevant research on MC-dropout suggests that dropout operations can be mathematically interpreted as Bayesian variational approximation reasoning for Gaussian processes with model parameters satisfying Bernoulli distribution, taking CNN as an example, for the ith convolutional layer in the network, its approximate variational distribution q (ω)i) Can be defined as:
wherein wi,jRepresents the jth parameter in the convolutional layer,expressing the approximate variation parameter of the corresponding Gaussian process, and K expressing the total number of the parameters of the convolution layer; p is a radical ofiThe probability of occurrence of a bernoulli distribution corresponds to the proportion of randomly discarded neurons in a dropout operation.
Based on the conclusion, for the CNN model which completes training by applying dropout, the dropout operation is still reserved in the model prediction stage, and random sampling of approximate posterior distribution of model parameters can be realized; therefore, after the same data is repeatedly input into the model and model prediction is performed for a plurality of times, the variance of the prediction result quantitatively reflects the uncertainty of the model on the current prediction result, and the mean value of the prediction result can be regarded as the final prediction result.
S2, applying the MC-dropout method to the building segmentation model, and carrying out uncertainty estimation on the building segmentation model;
in order to focus on the performance of the model on the segmentation problem, a frame selection result of a known building can be preset (the uncertainty can also be evaluated by adopting a similar scheme), the classification uncertainty of each pixel of an input image is evaluated, please refer to fig. 2, the MC-dropout-based building segmentation model has the main structural characteristics that a plurality of dropout operation layers are added to a Full Convolution Network (FCN) under the current mainstream ResNet and feature pyramid structure; as the existing research shows that the performance cannot be improved by performing weighted random sampling on lower-dimensional features, the dropout operation layer to be added is mainly concentrated on an encoding and decoding module in the middle section of the model.
The training mode of the model is the same as that of a classical FCN model, namely a cross entropy loss function is constructed by using training data and labels, and the model is optimized and solved by adopting a back propagation and random gradient descent algorithm. In the testing stage, the same input image is repeatedly predicted for N times by setting a certain random dropout proportion (such as 50%), N pixel classification probability graph results are obtained, a final probability graph is obtained by averaging the N pixel classification probability graph results according to pixels, and a segmentation result can be generated after threshold processing is carried out on the final probability graph; for each segmentation class (roof/wall/background), calculating the variance of N probability maps thereof by pixel as respective uncertainty representation, and carrying out cross-class statistics and pixel number averaging to obtain the total uncertainty self-evaluation result of the model for the current input image.
S3, applying the MC-dropout method to a shape optimization model, providing a building vector optimization and uncertainty estimation learning framework based on shape modeling, and quantitatively representing uncertainty while optimizing a vector outline;
the uncertainty of the building identification result can be measured from another aspect through the contour shape feature, and a building vector optimization and uncertainty estimation learning framework based on shape modeling is provided, and the uncertainty of the vector contour is quantitatively expressed while the vector contour is optimized. The framework optimizes the initial contour (contour tracking result of the segmentation graph) generated by the building segmentation model according to the following flow:
1) after the repeated prediction semantic feature set of the segmentation model is averaged element by element, performing pooling processing based on the initial contour node coordinates to obtain an image feature vector corresponding to each node;
2) cascading the two-dimensional coordinates of each node with the corresponding image characteristics, and encrypting the cascading characteristics of each node by a one-dimensional convolution kernel to form a tensor [ f1,f2,…,fn];
3) Continuously performing convolution processing to improve feature dimensionality, and generating a global feature G through maximum pooling operation;
4) g and node local feature [ f ]1,f2,…,fn]Respectively cascading, continuously encrypting the local and global fusion features into final point features, and predicting the coordinate correction value of each contour node;
5) and adding the corrected value and the input coordinate to obtain the optimized contour node coordinate.
In order to train the learning framework effectively, a loss function is established for the profile optimization effect from the two aspects of node deviation and angle deviation, namely:
Lpolygon=Lpoint+Lline
passing amountAdjusting the number of output optimized contour (polygon) and its corresponding true value nodes to be consistent by adding sampling points to form a matched point set S, and defining a node deviation loss function L by calculating the average distance of matched pointspoint: considering that the length of each side of the polygon is different, the same node deviation distance may bring about different degrees of shape deformation, so the angle deviation loss function L is definedlineThe output polygon is driven to be as parallel as possible to the homonymous side of its true value, and the loss can be defined by the average of the cosines of the homonymous sides (determined by the set of matched points S).
By adding dropout operation before convolution operation of the model and repeatedly predicting input data after model training is completed, a prediction set for optimizing the contour result for multiple times is generated. And respectively calculating the mean value and the variance of the optimized contour one by one according to the nodes, obtaining the final optimized contour result and the uncertainty representation of each node, and obtaining the building contour uncertainty self-evaluation result based on the shape characteristics by taking the uncertainty mean value of all the nodes.
Based on the embodiment, the invention provides a building segmentation uncertainty estimation method based on an MC-dropout method by researching a Bayesian approximation reasoning method suitable for a building identification task, wherein the classification uncertainty of each pixel of an input image is estimated to generate a pixel-level-by-pixel reliability self-estimation result; meanwhile, the invention provides a building vector optimization and uncertainty estimation learning framework based on modeling, which optimizes the building extraction result and quantitatively self-evaluates the reliability of the output result; on the whole, the building identification model uncertainty estimation method based on the Bayesian neural network theory can realize the quantitative self-evaluation of the reliability of the segmentation result and the vector extraction result by the model, and has important significance for promoting the deep application of the deep learning technology in the mapping practice and promoting the intelligent development of related industries.
The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims (6)
1. An uncertainty estimation method for a remote sensing image building identification model is characterized by comprising the following steps:
s1, obtaining approximate posterior distribution of CNN parameters of the building identification model by adopting an MC-dropout method, and realizing Bayesian approximation of model parameter distribution;
s2, applying the MC-dropout method to the building segmentation model, and carrying out uncertainty estimation on the building segmentation model;
s3, applying the MC-dropout method to the shape optimization model, providing a building vector optimization and uncertainty estimation learning framework based on shape modeling, and quantitatively representing uncertainty of the vector contour while optimizing the vector contour.
2. The uncertainty estimation method of the remote sensing image building identification model according to claim 1, characterized in that for the CNN model which completes training by applying dropout, the dropout operation is still retained in the model prediction stage to realize random sampling of the model parameter approximation posterior distribution; therefore, after the same data is repeatedly input into the CNN model and model prediction is performed for multiple times, the variance of the prediction result quantitatively reflects the uncertainty of the model on the current prediction result, and the average value is regarded as the final prediction result.
3. The uncertainty estimation method for the remote sensing image building identification model according to claim 1, characterized in that the MC-dropout building segmentation model is structurally characterized in that a plurality of dropout operation layers are added to a full convolution network under the current mainstream ResNet and feature pyramid structure, and the added dropout operation layers are mainly concentrated on a coding and decoding module in the middle section of the model.
4. The uncertainty estimation method of the remote sensing image building identification model according to claim 1, characterized in that the training mode of the building segmentation model is the same as that of a classical FCN model, namely, a cross entropy loss function is constructed by using training data and labels, and the solution is optimized by using back propagation and a random gradient descent algorithm;
in the testing stage, a certain random dropout proportion is set, N times of prediction are repeatedly carried out on the same input image, N pixel classification probability map results are obtained, a final probability map is obtained by averaging according to pixels, and a segmentation result can be generated after threshold processing is carried out on the final probability map; for each segmentation class, the variance of N probability maps is calculated according to pixels and is used as the uncertainty of each segmentation class, and the total uncertainty self-evaluation result of the model for the current input image can be obtained after the classes are crossed, the number of the pixels is counted and the average is taken.
5. The method for estimating the uncertainty of the building identification model based on the remote sensing images as claimed in claim 1, wherein the process of optimizing the initial contour generated by the building segmentation model S2 by the learning framework of building vector optimization and uncertainty estimation based on shape modeling is as follows:
(1) after the repeated prediction semantic feature set of the segmentation model is averaged element by element, performing pooling processing based on the initial contour node coordinates to obtain an image feature vector corresponding to each node;
(2) cascading the two-dimensional coordinates of each node with the corresponding image characteristics, and encrypting the cascading characteristics of each node by a one-dimensional convolution kernel to form a tensor [ f1,f2,…,fn];
(3) Continuously performing convolution processing to improve feature dimensionality, and generating a global feature G through maximum pooling operation;
(4) g and node local feature [ f ]1,f2,…,fn]Respectively cascaded, and the local and global fusion features are encrypted to the final featurePredicting the coordinate correction value of each contour node;
(5) and adding the corrected value and the input coordinate to obtain the optimized contour node coordinate.
6. The method for estimating the uncertainty of the building identification model based on the remote sensing images as claimed in claim 1, wherein the building vector optimization and uncertainty estimation learning framework based on the shape modeling establishes a loss function for the profile optimization effect from both node deviation and angle deviation, namely:
Lpolygon=Lpoint+Lline
adjusting the number of the nodes outputting the optimized contour and the corresponding truth value to be consistent by additionally increasing the sampling points to form a pairing point set S, and then defining a node deviation loss function L by calculating the average distance of the pairing pointspoint: considering that the length of each side of the polygon is different, the same node deviation distance may bring about different degrees of shape deformation, so the angle deviation loss function L is definedlineDriving the output polygon to be parallel to the homonymous side of the true value of the output polygon as much as possible, wherein the loss is defined by the average value of the cosine of the included angle of the homonymous side; wherein the homonymous edge is determined by the pairing point set S;
and respectively calculating the mean value and the variance of the optimized contour one by one according to the nodes, obtaining the final optimized contour result and the uncertainty representation of each node, and obtaining the building contour uncertainty self-evaluation result based on the shape characteristics by taking the uncertainty mean value of all the nodes.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110577026.2A CN113326759A (en) | 2021-05-26 | 2021-05-26 | Uncertainty estimation method for remote sensing image building identification model |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110577026.2A CN113326759A (en) | 2021-05-26 | 2021-05-26 | Uncertainty estimation method for remote sensing image building identification model |
Publications (1)
Publication Number | Publication Date |
---|---|
CN113326759A true CN113326759A (en) | 2021-08-31 |
Family
ID=77416944
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202110577026.2A Pending CN113326759A (en) | 2021-05-26 | 2021-05-26 | Uncertainty estimation method for remote sensing image building identification model |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN113326759A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116805351A (en) * | 2023-06-14 | 2023-09-26 | 壹品慧数字科技(上海)有限公司 | Intelligent building management system and method based on Internet of things |
CN117496309A (en) * | 2024-01-03 | 2024-02-02 | 华中科技大学 | Building scene point cloud segmentation uncertainty evaluation method and system and electronic equipment |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20180341876A1 (en) * | 2017-05-25 | 2018-11-29 | Hitachi, Ltd. | Deep learning network architecture optimization for uncertainty estimation in regression |
CN110415280A (en) * | 2019-05-06 | 2019-11-05 | 中国地质大学(武汉) | Remote sensing image and building vector method for registering and system under multitask CNN model |
CN110490061A (en) * | 2019-07-11 | 2019-11-22 | 武汉大学 | A kind of uncertainties model and measure of characteristics of remote sensing image |
CN111652892A (en) * | 2020-05-02 | 2020-09-11 | 王磊 | Remote sensing image building vector extraction and optimization method based on deep learning |
CN111723674A (en) * | 2020-05-26 | 2020-09-29 | 河海大学 | Remote sensing image scene classification method based on Markov chain Monte Carlo and variation deduction and semi-Bayesian deep learning |
CN112733945A (en) * | 2021-01-13 | 2021-04-30 | 中国科学技术大学 | Gaussian dropout-based deep learning model uncertainty calculation method |
-
2021
- 2021-05-26 CN CN202110577026.2A patent/CN113326759A/en active Pending
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20180341876A1 (en) * | 2017-05-25 | 2018-11-29 | Hitachi, Ltd. | Deep learning network architecture optimization for uncertainty estimation in regression |
CN110415280A (en) * | 2019-05-06 | 2019-11-05 | 中国地质大学(武汉) | Remote sensing image and building vector method for registering and system under multitask CNN model |
CN110490061A (en) * | 2019-07-11 | 2019-11-22 | 武汉大学 | A kind of uncertainties model and measure of characteristics of remote sensing image |
CN111652892A (en) * | 2020-05-02 | 2020-09-11 | 王磊 | Remote sensing image building vector extraction and optimization method based on deep learning |
CN111723674A (en) * | 2020-05-26 | 2020-09-29 | 河海大学 | Remote sensing image scene classification method based on Markov chain Monte Carlo and variation deduction and semi-Bayesian deep learning |
CN112733945A (en) * | 2021-01-13 | 2021-04-30 | 中国科学技术大学 | Gaussian dropout-based deep learning model uncertainty calculation method |
Non-Patent Citations (2)
Title |
---|
MICHAEL KAMPFFMEYER ET AL.: "Semantic segmentation of small objects and modeling of uncertainty in urban remote sensing images using deep convolutional neural networks", 《2016 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION WORKSHOPS (CVPRW)》 * |
YARIN GAL ET AL.: "Dropout as a Bayesian Approximation Representing Model Uncertainty in Deep Learning", 《PROCEEDINGS OF THE 33"RD INTERNATIONAL CONFERENCE ON MACHINE LEARNING (PMLR)》 * |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116805351A (en) * | 2023-06-14 | 2023-09-26 | 壹品慧数字科技(上海)有限公司 | Intelligent building management system and method based on Internet of things |
CN117496309A (en) * | 2024-01-03 | 2024-02-02 | 华中科技大学 | Building scene point cloud segmentation uncertainty evaluation method and system and electronic equipment |
CN117496309B (en) * | 2024-01-03 | 2024-03-26 | 华中科技大学 | Building scene point cloud segmentation uncertainty evaluation method and system and electronic equipment |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Yu et al. | A real-time detection approach for bridge cracks based on YOLOv4-FPM | |
CN111242208B (en) | Point cloud classification method, segmentation method and related equipment | |
CN110569901B (en) | Channel selection-based countermeasure elimination weak supervision target detection method | |
CN111612066B (en) | Remote sensing image classification method based on depth fusion convolutional neural network | |
CN111192270A (en) | Point cloud semantic segmentation method based on point global context reasoning | |
CN111462210B (en) | Monocular line feature map construction method based on epipolar constraint | |
CN110929848B (en) | Training and tracking method based on multi-challenge perception learning model | |
CN111583263A (en) | Point cloud segmentation method based on joint dynamic graph convolution | |
CN113326759A (en) | Uncertainty estimation method for remote sensing image building identification model | |
CN105809672A (en) | Super pixels and structure constraint based image's multiple targets synchronous segmentation method | |
CN112766186B (en) | Real-time face detection and head posture estimation method based on multitask learning | |
CN114638954B (en) | Training method of point cloud segmentation model, point cloud data segmentation method and related device | |
CN114092697B (en) | Building facade semantic segmentation method with attention fused with global and local depth features | |
CN110599443A (en) | Visual saliency detection method using bidirectional long-term and short-term memory network | |
CN115410059B (en) | Remote sensing image part supervision change detection method and device based on contrast loss | |
CN111553296B (en) | Two-value neural network stereo vision matching method based on FPGA | |
CN115375672A (en) | Coal gangue detection method based on improved YOLOv4 algorithm | |
CN114202035B (en) | Multi-feature fusion large-scale network community detection algorithm | |
CN114358246A (en) | Graph convolution neural network module of attention mechanism of three-dimensional point cloud scene | |
CN113139432B (en) | Industrial packaging behavior identification method based on human skeleton and partial image | |
CN117765258A (en) | Large-scale point cloud semantic segmentation method based on density self-adaption and attention mechanism | |
CN114463340A (en) | Edge information guided agile remote sensing image semantic segmentation method | |
Naufal et al. | Weather image classification using convolutional neural network with transfer learning | |
CN115861713A (en) | Carotid plaque ultrasonic image processing method based on multitask learning | |
CN116110074A (en) | Dynamic small-strand pedestrian recognition method based on graph neural network |
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 | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20210831 |
|
WD01 | Invention patent application deemed withdrawn after publication |