CN112966605A - Depth learning based positioning model fitting similar to RPC - Google Patents

Depth learning based positioning model fitting similar to RPC Download PDF

Info

Publication number
CN112966605A
CN112966605A CN202110245438.6A CN202110245438A CN112966605A CN 112966605 A CN112966605 A CN 112966605A CN 202110245438 A CN202110245438 A CN 202110245438A CN 112966605 A CN112966605 A CN 112966605A
Authority
CN
China
Prior art keywords
model
fitting
network
deep learning
rpc
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
CN202110245438.6A
Other languages
Chinese (zh)
Inventor
邓明军
刘鑫
汪忠
王迪
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.)
Xiangtan University
Original Assignee
Xiangtan University
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 Xiangtan University filed Critical Xiangtan University
Priority to CN202110245438.6A priority Critical patent/CN112966605A/en
Publication of CN112966605A publication Critical patent/CN112966605A/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
    • G06V20/13—Satellite images
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • 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

Landscapes

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

Abstract

The embodiment of the invention discloses a positioning model similar to RPC (remote procedure call) based on deep learning fitting, which comprises the following overall steps: acquiring deep learning training samples which respectively comprise ground target point space information and image coordinate information; constructing a deep learning network model; fitting the constructed deep learning model; and (5) applying the model. The invention mainly aims at the problems that the fitting accuracy of an RPC model is not high enough and a strict imaging model cannot be replaced under the conditions of an oblique view mode and non-zero Doppler frequency, and adopts a positioning model similar to the RPC based on deep learning fitting, so that the application range and the positioning accuracy of the positioning model are effectively expanded, and convenience is provided for high-resolution satellite positioning.

Description

Depth learning based positioning model fitting similar to RPC
Technical Field
The invention belongs to the field of remote sensing science, and particularly relates to a method for fitting a positioning model similar to RPC in a satellite-borne SAR positioning model by adopting deep learning.
Background
The satellite-borne SAR geometric positioning model describes a mathematical relationship between the image point position and the corresponding ground target point. The method is generally divided into two types of rigorous imaging models and general imaging models [1 ]. The general imaging model is usually a Rational Polynomial Coefficient (RPC) positioning model, has strong applicability, does not need to know parameters such as imaging process, system characteristics and the like of a sensor, and therefore becomes the most widely used imaging model.
The RPC model approximates a rigorous imaging model under the conditions of a front side view mode and zero Doppler. However, in strabismus mode, non-zero doppler frequency, the doppler center frequency varies greatly with distance, and the accuracy of the geometric positioning model is affected [2 ]. In order to ensure the universality of the geometric model, the research on the model which is not influenced by the imaging view angle mode of the image and can obtain more accurate image coordinate information according to the spatial information of the ground target point is a valuable scientific research task.
Disclosure of Invention
In order to solve the problem that the fitting accuracy of an RPC model is not high due to the fact that the change of the azimuth resolution is caused by the fact that the Doppler center frequency is greatly changed along with the distance under the satellite-borne SAR image under the squint mode and the non-zero Doppler condition, the invention provides a fitting model which is similar to an RPC positioning model and is built by using a deep learning technology. The method combines the characteristics of deep learning 'black boxes', does not consider the physical meaning of sensor imaging and the mathematical meaning of mapping relation, and fits image coordinate information by training a network and using ground target space information.
In order to solve the technical problems, the technical scheme adopted by the invention is to construct a fitting model similar to an RPC positioning model by using a deep learning method, and the method comprises the following overall steps:
step 1: and acquiring deep learning training samples which respectively comprise ground target point space information and image coordinate information.
Step 2: building deep learning network model
And step 3: fitting deep learning model
And 4, step 4: and (5) applying the model. After the network is trained, a plurality of groups of input ground coordinate information are given, and output image coordinate information is obtained through the full-connection network.
The step 1 further comprises the following substeps:
step S11: and acquiring data information. The coordinate information of a plurality of groups of surface space points and the corresponding image coordinate information are obtained from the image by utilizing the prior art.
Step S12: a vector polynomial is constructed. Based on the space coordinates (P, L, H) of the object space ground target point, a polynomial form is formed by utilizing the operational relation of the matrix so as to understand the meaning of the data fitting in the deep learning model:
Yj＝Xi+Wi
where Wi represents the bias term, Xi represents the object ground coordinates (P, L, H), and Yj represents the pixel ground coordinates (r, c).
Step S13: and (6) normalizing the data. To enhance the fitting ability of the simulation, the ground and image coordinates were normalized to between-1 and 1.
Step S14: the sample data is divided into two parts of training data and test data.
And 2, a full-connection layer network is adopted, and a deep learning fitting framework of a 3-layer neural network is designed. The network is composed of an input layer, a 1-layer hidden layer and an output layer. Xi is an input value of the network and Yj is an output value of the network in step S12.
Preferably, the ReLU function is adopted in step 2 as an activation function of the network to improve the convergence speed of the network.
The RMSE loss function was chosen as the evaluation index of the model.
The step 3 comprises the following substeps:
and step S31, training the network model. And (4) training by combining the sample data in the step (1) with the network model built in the step (2). The fitting effect of the network deep learning is improved by optimizing parameters, and the fitting effect comprises changing the input dimension of the network and adjusting the learning rate. And judging whether the model is trained or not through the change of the loss function value.
Step S32: and testing the network model. After the model is trained, the prepared test sample data is selected, Xi is substituted into the network model obtained by training in the step S31 for calculation, and the corresponding output Yj is obtained.
Step S33: and (5) processing the prediction data. And performing inverse normalization on the prediction data by adopting the standard of the previous normalization, comparing the inverse normalization with the original data, and calculating to obtain a fitting error so as to judge the effect of the fitting model. If the fitting accuracy is not ideal, the process returns to step S31 to retrain.
And 4, packaging the depth network model obtained by fitting in the step 3 into software, and applying the software to actual data to predict the pixel coordinate value with high precision. Selecting a plurality of groups of object space ground coordinate information, normalizing the selected object space ground coordinate information, and substituting the normalized object space ground coordinate information into the fitted network model in the step 3 to calculate corresponding pixel ground coordinate information.
Compared with the prior art, the invention has the advantages and beneficial effects that:
(1) the invention has a stricter imaging positioning model, has the characteristics of not depending on the structural information of the sensor and being suitable for various high-resolution satellite images.
(2) By adopting a deep learning technology and utilizing the 'black box' characteristic of a deep network, the expected output data can be obtained by fitting as long as the input data is known, even if the mathematical relationship between object elements and pixel data and the geometric meanings of an object space and an image space are not known. Breaks through the limit of modeling in the mathematical sense.
(3) The method can avoid the problem that the fitting accuracy of the RPC model is not as good as that of a strict imaging model under the conditions of a side-view mode and non-zero Doppler, and the model applicability is strong.
Drawings
Fig. 1 is a general technical flow chart of the present invention.
Fig. 2 is a schematic view of a fully connected layer.
Detailed Description
The following further describes a specific embodiment of the present invention with reference to the accompanying drawings, and fig. 1 shows a flowchart of the present invention, which can be mainly divided into four steps of obtaining a sample, constructing a model, fitting the model, and applying the model. Details of the implementation of these four stages will be described below in conjunction with the following figures.
How to realize an imaging model similar to an RPC model with high-precision positioning under the condition of a front-side view mode and a zero Doppler has important value for high-resolution satellite image positioning and application.
In the application of high-resolution satellite positioning, the mapping relation exists between the image-side ground coordinate information obtained by the imaging model and the coordinate information of the object-side ground target point, and the mapping relation can be regarded as a complex nonlinear function, so that the deep learning neural network is used for mapping the relation.
The method utilizes the RPC geometric positioning model which is widely used at present to obtain a plurality of object space and image space ground coordinate information, and the data is processed to be used as sample data. Based on a deep neural network, any probability model theoretical basis can be fitted, and a deep learning model is constructed by utilizing the 'black box' characteristic of a deep learning technology, so that the image space ground coordinate information can be fitted according to the known object space ground coordinate information.
The invention provides a method for constructing an RPC-like positioning model based on deep learning, which comprises the following steps:
step 1: and obtaining deep learning training samples.
Step S11: and acquiring sample data. The longitude (P), the latitude (L) and the elevation value (H) of a batch of object ground target points and the longitude value (r) and the latitude value (c) of corresponding image ground coordinates are read from the images through a widely used RPC model, and are stored into a txt document in sequence in a determinant mode.
Step S12: a vector polynomial is constructed. Based on the space coordinates (P, L, H) of the object space ground target point, a polynomial form is formed by using the operational relation of a vector matrix so as to understand the meaning of the data fitting in the deep learning model:
Yj＝AiXi+Wiequation 1
In formula 1, Wi represents a bias term, Ai represents a coefficient, Xi only comprises P, L, H terms when the first order is taken, and Xi takes the third order if fitting training is performed by simulating an RPC form, wherein the third order comprises 19 terms of P, L, H, PL, PH, LH, PLH, PL2, PH2, LP2, LH2, HP2, HL2, H2, L2, P2, H3, L3 and P3. Yj denotes the pixel ground coordinates (r, c), for example as follows:
Figure BDA0002963923290000051
as in formula 2, X1, X2, and X3 represent three values P, L, H, and Y1 and Y2 represent r and c, respectively.
Step S13: and (6) normalizing the data. Due to the large difference of the latitude, longitude and altitude values, gradient explosion and gradient disappearance can occur during model fitting. In order to enhance the fitting ability of the simulation and reduce the influence caused by the magnitude difference, the ground coordinate and the image coordinate are normalized to be between-1 and 1.
The normalization formula used is:
Y＝(X-Xmean)/(Xmax-Xmin) Equation 3
In the expression, Xmean represents the mean value of each line of data in the txt document, and Xmax and Xmin represent the maximum value and the minimum value of each line of data in the txt document.
Step S14: dividing sample data into two parts of training data and testing data, and respectively storing the two parts of training data and testing data as' train.
Step 2: and constructing a deep learning network model.
The invention introduces a simple full-connection layer network and designs a deep learning fitting framework of a 3-layer neural network. The network consists of an input layer, a hidden layer and an output layer. Xi is input data to the network and Yj is output data to the network in step S12.
Fig. 2 is a very common network structure after a network diagram of a fully-connected layer, and all layers of the network structure are in a nonlinear mapping relationship of input and output, and the relationship is as follows:
yi＝gi(wixi+bi) Equation 4
In the formula, yi represents the output of the ith layer and is also the input of the (i + 1) th layer; xi represents the input of the ith layer; wi represents the weight of the ith layer; gi is a non-linear function of the ith layer; bi represents the deviation of the ith layer.
For a full-link network, the weights and deviations of each layer are variables that can be trained, and when the values need to be updated, the goal can be achieved by minimizing the loss function.
In the invention, the input dimension of the input layer is selected according to the order of Xi, when Xi is the first order, the input dimension of the network is defined as 3 dimensions, and the dimension of the output layer is fixed as 2 dimensions.
Further, an official ReLU function is called in step 2 as an activation function of the network to improve the convergence speed of the network.
The ReLU function formula is as follows:
(x) max (0, x) formula 5
Furthermore, because the sample training data is abundant and tens of thousands of magnitude of values exist in the sample data, the improved RMSE loss function is selected as the evaluation index of the model.
During training, the output of the predicted pixel coordinate value after passing through the fully-connected neural network is close to the test value of the sample by minimizing a loss function, and the loss function expression is as follows:
Figure BDA0002963923290000061
and step 3: and fitting the deep learning model.
And step S31, training the network model. And (3) feeding the training sample data divided in the step (1) into the network model built in the step (2). The fitting effect of the network deep learning is improved by optimizing parameters, and the fitting effect comprises changing the input dimension of the network and adjusting the learning rate.
Further, the quality of model training is judged through the change of the loss function value, when the loss function value generally decreases to a certain small value and subsequently tends to be stable, namely converges to a certain value, network training is completed, and the model weight with the best training is saved for subsequent testing.
Preferably, the training process of the network is essentially a process of continuously updating the weight parameters, and the training process is continuously updated by a large number of samples and a proper optimization method until the optimal weight parameters are found, and the group of parameters is obtained by an optimization algorithm. The invention introduces a common random batch gradient descent method (SGD), and the expression is as follows:
xt+1＝xt-ηtgtformula (II)7
Where xt represents the position of the t-th step, η t represents the step size, and gt represents the random gradient.
Step S32: and testing the network model. After the network model is trained based on the step S32, the prepared test sample data is selected, Xi is fed into the network model for weight parameter calculation, and finally, the corresponding output Yj is obtained.
Step S33: and (5) processing the prediction data. Since the fitting error of the model calculation is the comparison of the raw data, an inverse normalization process is required. And (3) performing inverse normalization on the data predicted in the step (S32) by adopting the standard of the normalized sample data in the step (1), wherein an inverse normalization formula is as follows:
X＝Xpre(Xmax-Xmin+Xmean) Equation 8
Where Xpre represents the value predicted in step S32, and Xmax, Xmin, Xmean are the values in step S13.
And comparing with the original test data, and calculating to obtain the final fitting error of the model so as to judge the fitting effect of the model. The fitting error calculation formula is:
Figure BDA0002963923290000071
wherein y represents the original true value, and ypre represents the predicted value after reverse normalization.
If the fitting precision is less than 5%, the fitting effect of the model is good, and the model can be used for application. Otherwise, if the fitting accuracy is greater than 5%, returning to step S31 for retraining.
And 4, step 4: and (5) applying the model.
Following step 3, the model can be applied directly as needed. In order to facilitate the realization of giving a plurality of groups of input ground coordinate information, such as a plurality of space coordinate values (P, L, H), output image coordinate information (r, c) is obtained through a full-connection network, and the weight of the stored optimal network model is packaged into software.
Specifically, the depth network model obtained by fitting in the step 3 is applied to actual data to predict the pixel coordinate value with high precision. Based on a python language, a software interface is written by utilizing a python version PyQt of a Qt library, and a PyInstaller library is used for packaging the trained neural network model program, the relevant weights and the relevant resource packages into an exe file in a virtual environment. Packaging in a virtual environment can facilitate stable operation of software on different computers.
Further, selecting a plurality of groups of object space ground coordinate information to be fitted, wherein the values are (P, L, H), normalizing the data by using the normalization standard in the step S13, reading the data by using software packaged in the previous step, calculating to obtain the optimal parameters of the data, and outputting the final pixel ground coordinate information (r, c).
The above-described embodiments of the present invention are merely for illustrative purposes and do not represent the quality of the embodiments.
The present invention is described in detail by the above embodiments, and the embodiment method can be clearly understood by those skilled in the art
The above embodiments are not intended to limit the present invention, and the present invention is not limited to the above embodiments, but various modifications, additions and substitutions are possible for those skilled in the art without departing from the spirit of the invention or exceeding the scope of the claims.
Reference to the literature
[1]T.Toutin(2004)Review article:Geometric processing of remote sensing images:models,algorithms and methods,International Journal of Remote Sensing,25:10,1893-1924
[2] Zhao rui shan, spaceborne SAR geometric calibration model and method research [ D ]. Liaoning engineering technology university, 2017.

Claims (6)

1. Fitting a RPC-like positioning model based on deep learning, characterized in that the method comprises the following steps:
step 1: and acquiring deep learning training samples which respectively comprise ground target point space information and image coordinate information.
Step 2: building deep learning network model
And step 3: fitting deep learning model
And 4, step 4: and (5) applying the model. After the network is trained, a plurality of groups of input ground coordinate information are given, and output image coordinate information is obtained through the full-connection network.
2. The deep learning based fitting of an RPC-like localization model according to claim 1, wherein the step 1 further comprises the sub-steps of:
step S11: and acquiring data information.
Step S12: a vector polynomial is constructed. Based on the space coordinates (P, L, H) of the object space ground target point, a polynomial form is formed by utilizing the operational relation of the matrix so as to understand the meaning of the data fitting in the deep learning model:
Yj＝Xi+Wi
wherein WiRepresenting an offset term, XiRepresenting object space ground coordinates (P, L, H), YjRepresenting pixel ground coordinates (r, c).
Step S13: and (6) normalizing the data. To enhance the fitting ability of the simulation, the ground and image coordinates were normalized to between-1 and 1.
Step S14: the sample data is divided into two parts of training data and test data.
3. The RPC-like positioning model based on deep learning fitting of claim 1, wherein step 2 employs a full-connection layer network, and a deep learning fitting framework of a 3-layer neural network is designed. The network is composed of an input layer, a 1-layer hidden layer and an output layer. Xi is an input value of the network and Yj is an output value of the network in step S12.
4. The fitting of an RPC-like location model based on deep learning of claim 3, wherein the ReLU function is used as the activation function of the network in step 2 to increase the convergence rate of the network. The RMSE loss function was chosen as the evaluation index of the model.
5. The deep learning based fitting of an RPC-like localization model according to claim 1 or 3, wherein step 3 comprises the sub-steps of:
and step S31, training the network model. And (4) training by combining the sample data in the step (1) with the network model built in the step (2). The fitting effect of the network deep learning is improved by optimizing parameters, and the fitting effect comprises changing the input dimension of the network and adjusting the learning rate. And judging whether the model is trained or not through the change of the loss function value.
Step S32: and testing the network model. After training the model, selecting the prepared test sample data, and adding XiSubstituting the obtained result into the network model trained in the step S31 to calculate to obtain corresponding output Yj。
Step S33: and (5) processing the prediction data. And performing inverse normalization on the prediction data by adopting the standard of the previous normalization, comparing the inverse normalization with the original data, and calculating to obtain a fitting error so as to judge the effect of the fitting model. If the fitting accuracy is not ideal, the process returns to step S31 to retrain.
6. The RPC-like positioning model based on deep learning fitting of claim 1 or 5, wherein step 4 is to pack the depth network model obtained by fitting in step 3 into software, and apply the software to actual data to predict pixel coordinate values with high precision. Selecting a plurality of groups of object space ground coordinate information, normalizing the selected object space ground coordinate information, and substituting the normalized object space ground coordinate information into the fitted network model in the step 3 to calculate corresponding pixel ground coordinate information.
CN202110245438.6A 2021-03-05 2021-03-05 Depth learning based positioning model fitting similar to RPC Pending CN112966605A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110245438.6A CN112966605A (en) 2021-03-05 2021-03-05 Depth learning based positioning model fitting similar to RPC

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110245438.6A CN112966605A (en) 2021-03-05 2021-03-05 Depth learning based positioning model fitting similar to RPC

Publications (1)

Publication Number Publication Date
CN112966605A true CN112966605A (en) 2021-06-15

Family

ID=76276642

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110245438.6A Pending CN112966605A (en) 2021-03-05 2021-03-05 Depth learning based positioning model fitting similar to RPC

Country Status (1)

Country Link
CN (1) CN112966605A (en)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104978761A (en) * 2015-06-01 2015-10-14 北京航空航天大学 RPC model correction coefficient acquisition method
CN110660099A (en) * 2019-03-22 2020-01-07 西安电子科技大学 Rational function model fitting method for remote sensing image processing based on neural network
US20210027487A1 (en) * 2019-07-26 2021-01-28 Here Global B.V. Method, apparatus, and system for recommending ground control points for image correction

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104978761A (en) * 2015-06-01 2015-10-14 北京航空航天大学 RPC model correction coefficient acquisition method
CN110660099A (en) * 2019-03-22 2020-01-07 西安电子科技大学 Rational function model fitting method for remote sensing image processing based on neural network
US20210027487A1 (en) * 2019-07-26 2021-01-28 Here Global B.V. Method, apparatus, and system for recommending ground control points for image correction

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
赵瑞山: "星载SAR几何定标模型与方法研究", 《中国优秀博硕士学位论文全文数据库(博士)基础科学辑》 *

Similar Documents

Publication Publication Date Title
CN109934282B (en) SAGAN sample expansion and auxiliary information-based SAR target classification method
CN109001736B (en) Radar echo extrapolation method based on deep space-time prediction neural network
CN104268869B (en) Automatic multilevel-resolution remote-sensing image registration method based on particle swarm optimization
CN111174753B (en) Optical image and laser height measurement data adjustment method based on rational function model
CN109685743A (en) Image mixed noise removing method based on noise learning neural network model
CN111144234A (en) Video SAR target detection method based on deep learning
CN105550649A (en) Extremely low resolution human face recognition method and system based on unity coupling local constraint expression
CN111161199A (en) Spatial-spectral fusion hyperspectral image mixed pixel low-rank sparse decomposition method
CN115795683A (en) Wing profile optimization method fusing CNN and Swin transform network
CN115660041A (en) Sea wave height prediction and model training method, electronic device and storage medium
CN112180369A (en) Depth learning-based sea surface wind speed inversion method for one-dimensional synthetic aperture radiometer
CN110660099B (en) Rational function model fitting method for remote sensing image processing based on neural network
CN114964313A (en) RVM-based fiber optic gyroscope temperature compensation method
CN114722900A (en) Airborne radar depth measurement shallow sea sediment classification method
CN116843679B (en) PET image partial volume correction method based on depth image prior frame
CN113569810A (en) Remote sensing image building change detection system and method based on deep learning
CN113899386A (en) Multi-source optical satellite remote sensing image collaborative regional net adjustment method and system based on three-dimensional reference net
CN112966605A (en) Depth learning based positioning model fitting similar to RPC
Daužickaitė et al. Spectral estimates for saddle point matrices arising in weak constraint four‐dimensional variational data assimilation
CN115618749B (en) Error compensation method for real-time positioning of large unmanned aerial vehicle
CN117011668A (en) Weather radar echo extrapolation method based on time sequence prediction neural network
CN115471428A (en) Co-phase error correction method based on all-optical diffraction neural network
CN113487577B (en) Quick Gamma adjustment method, system and application based on GRU-CNN combined model
CN113724325B (en) Multi-scene monocular camera pose regression method based on graph convolution network
CN113435243A (en) Hyperspectral true downsampling fuzzy kernel estimation method

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
RJ01 Rejection of invention patent application after publication
RJ01 Rejection of invention patent application after publication

Application publication date: 20210615