CN113029318A - Satellite platform tremor detection and analysis method based on deep learning - Google Patents
Satellite platform tremor detection and analysis method based on deep learning Download PDFInfo
- Publication number
- CN113029318A CN113029318A CN202110137673.1A CN202110137673A CN113029318A CN 113029318 A CN113029318 A CN 113029318A CN 202110137673 A CN202110137673 A CN 202110137673A CN 113029318 A CN113029318 A CN 113029318A
- Authority
- CN
- China
- Prior art keywords
- tremor
- feature matching
- registration error
- network
- line
- 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
- G01—MEASURING; TESTING
- G01H—MEASUREMENT OF MECHANICAL VIBRATIONS OR ULTRASONIC, SONIC OR INFRASONIC WAVES
- G01H9/00—Measuring mechanical vibrations or ultrasonic, sonic or infrasonic waves by using radiation-sensitive means, e.g. optical means
Abstract
The invention relates to a satellite platform tremor detection and analysis method based on deep learning, which belongs to the technical field of satellite remote sensing and comprises the steps of preparing data and preprocessing an original multispectral image; the pretreatment comprises the following steps: relative radiation correction and line frequency normalization; selecting wave bands, and determining two wave bands for detection; image matching, namely extracting feature matching point pairs by adopting a deep learning matching algorithm based on a convolutional neural network, and generating feature matching point coordinates, feature matching point confidence coefficients and feature matching point descriptors; performing tremor analysis, and calculating registration error for each feature matching point; analyzing line by line to obtain the change of the registration error along with the imaging line; and carrying out spectrum analysis on the satellite platform tremor time sequence data through Fourier transform to obtain the spectrum characteristics of the registration error caused by tremor, wherein the obtained platform tremor analysis data has high precision and high confidence level and meets the requirements of a satellite remote sensing system.
Description
Technical Field
The invention belongs to the technical field of satellite remote sensing, and particularly relates to a satellite platform tremor detection and analysis method based on deep learning.
Background
The satellite platform tremor is a tiny tremor response generated by the influence of internal and external factors during the in-orbit operation of the satellite, and along with the improvement of the resolution of the satellite image, the platform tremor has more and more prominent influence on the geometric accuracy of the high-resolution optical satellite remote sensing image.
Treatment measures for the remote sensing satellite platform tremor are as follows: on one hand, vibration isolation and suppression measures are required to be adopted through a dynamic method, and the imaging image is avoided and reduced; on the other hand, the platform tremor measurement needs to be carried out in a direct or indirect mode, and the image quality reduction and the geometric deformation caused by the platform tremor measurement are compensated. However, the platform vibration is unavoidable and difficult to control, and the vibration suppression and isolation are difficult to achieve completely, so that the high-quality and high-precision processing of the high-resolution remote sensing image can be guaranteed only by accurately measuring the platform vibration.
The multispectral image is a multispectral image which is acquired at the same imaging time and is applied to the same area on the ground, theoretically, different spectral bands at the same time are located at the same geometric position relative to image points acquired at the same ground point, but are influenced by platform tremor, and the geometric consistency among different spectral bands is damaged, so that the detection and analysis of the platform tremor of the remote sensing satellite can be realized by detecting the geometric precision consistency of adjacent scanning lines of the multispectral image.
In the prior art, chinese patent application No. 201410234816.0 entitled "a method and system for detecting satellite platform tremor based on multispectral images" discloses a method for detecting satellite platform tremor based on multispectral images, which adopts intensive point-by-point sampling, combines image correlation coefficients with a least square method for matching, and performs tremor analysis based on the method. The method can realize tremor detection of the satellite platform along the push-broom direction and the vertical direction, can acquire more accurate tremor frequency through frequency spectrum analysis, and provides a basis for further improving the image processing quality and the geometric precision. However, because the method adopts a dense multipoint sampling mode, the precision of the method is not enough to meet the tremor analysis precision required by modern technology due to uncertainty, low precision and weak representativeness of the sampling, and the efficiency of the whole system is low due to the large amount of data required to be processed due to the dense multipoint sampling.
Disclosure of Invention
The invention aims to provide a satellite platform tremor detection and analysis method based on deep learning, which is used for solving the problem of low tremor analysis precision caused by low matching precision in the existing method.
Based on the above purpose, a technical scheme of the method is as follows:
s01: data preparation, namely preprocessing an original multispectral image; the pretreatment comprises the following steps: relative radiation correction and line frequency normalization;
s02: selecting wave bands, and determining two wave bands for detection;
s03: image matching, namely extracting feature matching point pairs by adopting a deep learning matching algorithm based on a convolutional neural network, and generating feature matching point coordinates, feature matching point confidence coefficients and feature matching point descriptors;
s04: performing tremor analysis, and calculating registration error for each feature matching point; analyzing line by line to obtain the change of the registration error along with the imaging line; and carrying out spectrum analysis on the satellite platform tremor time sequence data through Fourier transform to obtain the spectrum characteristics of the registration error caused by tremor.
The beneficial effects of the above technical scheme are:
the method is based on a deep learning method to determine the relevant information of the image feature matching points and carry out tremor analysis; the image feature matching points have high stability, high accuracy and strong representativeness, the tremor analysis is carried out by determining the relevant information of the feature matching points, the feature matching points obtained by the method have high accuracy, strong typicality and high reliability, and the platform tremor analysis data obtained by the method have high accuracy and high confidence level and meet the requirements of a satellite remote sensing system.
Preferably, in the present invention, in the step S02, the two wavelength bands are adjacent wavelength bands.
Preferably, in the present invention, the step of S03 includes: constructing a VGG-16 network framework based on a deep learning matching algorithm;
dividing the VGG-16 network framework into different blocks according to convolutional layers and pooling layers included in the VGG-16 network framework, wherein each block comprises a plurality of convolutional layers and one pooling layer; the front end of the network frame is connected with the blocks, and the rear end of the network frame is sequentially provided with a full connection layer and a classification function.
Preferably, in the present invention, the step S03 further includes:
setting an input end network, wherein the input end network is a full convolution neural network, and the full convolution neural network comprises two input convolution layers; wherein each of the input convolutional layers is subjected to two convolutions and one pooling.
Preferably, in the present invention, the step S03 further includes:
setting an output end network, wherein the output end network comprises a coordinate sub-network, a confidence sub-network and a description sub-network corresponding to the feature matching points;
after the confidence sub-network convolves the input feature map, the confidence of each feature matching point is obtained through an activation function;
the numerical value corresponding to each coordinate (r, c) in the coordinate subnetwork is converted into the coordinate formula of the feature matching point in the image as follows:
Pmap,x(r,c)＝(c+Prelative,x(r,c))·fdownsample
Pmap,y(r,c)＝(r+Prelative,y(r,c))·fdownsample
wherein, the image down-sampling multiplying power fdownsample＝8，Pmap,x(r,c)、Pmap,y(r, c) respectively representing coordinate values of the feature matching points in the image on the x and y axes; prelative,x(r,c)、Prelative,y(r, c) respectively representing the offset of the feature matching points in the image relative to the row and column coordinates of c and r on the x and y axes;
the descriptor subnetwork is used for providing a multi-dimensional correlation vector for each feature matching point.
Preferably, in the present invention: in step S04, the calculating the registration error for each feature matching point includes:
assuming that there are V feature matching point pairs in the u-th scanning line, the registration error calculation formula of the k-th homologous image point of the u-th scanning line is as follows:
wherein M, N are two wave bands in the step S02; Δ xu,k、Δyu,kRepresenting the registration errors in the CCD direction and in the push-scan direction, respectively;
respectively representing the coordinates of the kth homonymous image point of the u scanning line on M, N two wave band images; Δ x, Δ y represent M, N designed values for the translation between the CCDs of the two wavebands in the CCD direction and in the push-scan direction.
Preferably, in the present invention, in step S04, the analyzing line by line to obtain the variation of the registration error with the imaging line includes:
taking the registration error of all the feature matching points of each scanning line to obtain an average value, and taking the average value as the optimal estimation of the registration error corresponding to the scanning line; and obtaining the time sequence data of the registration errors along the CCD direction and the push-scanning direction along the change of the imaging lines through line-by-line statistics.
Preferably, in the present invention, in step S04, the performing spectral analysis on the satellite platform tremor time series data through fourier transform to obtain the spectral characteristics of the registration error caused by tremor includes:
performing spectrum analysis on the time series data corresponding to the registration error by using the Fourier transform, wherein the integration time of each scanning line of the multispectral image is delta T, the total imaging time T of the image comprising U imaging lines is delta T-U, and the imaging time T corresponding to the U-th imaging line is Tu＝t0+ Δ t · U (U ═ 0,1, …, U-1), where t is0Is the initial imaging time;
performing the fourier transform on a discrete sequence of Δ x (U) (U-0, 1, …, U-1):
wherein K represents the decomposed serial number and takes the value of 0-U/2; a isKAnd bKRespectively representing the amplitudes of the K-th harmonic cosine and sine functions; c. CKRepresents the amplitude of the K harmonic;
frequency f corresponding to K harmonicKComprises the following steps:
and obtaining a spectrum curve corresponding to the spectrum characteristic of the registration error along the CCD direction.
Preferably, in the present invention, the frequency and amplitude obtained by the fourier transform are used as initial values of a sine function, and a least square estimation is used to determine the registration error frequency, the registration error amplitude, and the registration error initial phase of the registration error.
Drawings
FIG. 1 is a schematic flow chart of an implementation of a method for detecting and analyzing tremor of a satellite platform based on deep learning according to an embodiment of the present invention;
FIG. 2 is a schematic structural diagram of a VGG-16 network framework in an embodiment of the invention;
fig. 3 is a schematic structural diagram of an unsurpoint network frame in the embodiment of the present invention.
Detailed Description
The following further describes embodiments of the present invention with reference to the drawings.
The embodiment provides a method for detecting and analyzing tremor of a satellite platform based on deep learning, and the overall process is shown in fig. 1, and the method for detecting and analyzing tremor specifically comprises the following steps:
s01: data preparation, namely preprocessing an original multispectral image; the pretreatment specifically comprises: and (4) performing relative radiation correction and line frequency normalization, and processing to obtain an unregistered multispectral image.
The purpose of relative radiation correction is to eliminate the influence of stripe noise on subsequent processing and improve the matching precision. In this step, the line frequency normalization means that the imaging frequency and time of each line in the multispectral spectral band images are unified, and the imaging time of each line of the multispectral spectral band images is normalized according to the imaging frequency of each multispectral spectral band image by using the imaging line integral time information of each line of the multispectral spectral band images. Therefore, through line frequency normalization, the influence of integral time jump can be removed, and the images are ensured to be uniform and continuous in space and time.
S02: selecting wave bands, and determining two wave bands for detection;
the band selection includes two principles:
1) selecting the wave bands with similar spectral characteristics as far as possible, and avoiding the spectral characteristic difference from reducing the matching precision;
2) the adjacent wave bands are selected as much as possible, and because the tremor amplitude is usually small and the frequency is high, the adjacent wave bands are adopted for detection, and the result is finer. For the high-resolution multispectral camera, the CCDs of the blue, green, red and infrared spectral bands are sequentially arranged on a focal plane in parallel at certain intervals. Because the spectral characteristics of the ground objects in different wave bands are different, the wave bands with similar spectral characteristics are selected as far as possible, the imaging characteristics of the infrared wave band and the visible light wave band have great difference, and whether the matching error of the infrared wave band and the visible light wave band meets the requirement needs to be checked firstly; meanwhile, the influence of the topographic relief is considered, in the step, the images of adjacent wave bands are selected, the difference of the field angle is small, the influence of the topographic relief is small, and the detection result is finer.
Preferably, in this step, two adjacent visible light bands are used for detection.
S03: image matching, namely extracting feature matching point pairs by adopting a deep learning matching algorithm based on a convolutional neural network to generate feature matching point coordinates, feature matching point confidence coefficients and feature matching point descriptors;
among them, the deep learning algorithm based on the convolutional neural network is a deep learning method which is developed on the basis of a multilayer neural network and is designed specifically for an image, and an excellent effect is obtained in image processing. In the image feature extraction, in a convolutional neural network, convolution and pooling operation are performed on an image to extract features, and a feature matrix extracted by the convolution and pooling operation is input into a full-link layer or a global mean pooling layer to generate feature vectors of the image.
Preferably, in the present invention, the step S03 includes: constructing a VGG-16 network framework based on a deep learning matching algorithm;
dividing the VGG-16 network framework into different blocks according to convolutional layers and pooling layers included in the VGG-16 network framework, wherein each block comprises a plurality of convolutional layers and a pooling layer; the front end of the network frame is connected by different blocks, and the rear end of the network frame is sequentially provided with a full connection layer and a classification function. Specifically, a schematic structural diagram of the VGG-16 network framework is shown in fig. 2.
And extracting the feature matching points and generating the feature matching point descriptors by adopting an UnSuperPoint deep learning matching algorithm. The backsbone part in the UnSuperPoint deep learning matching algorithm (figure 3) is constructed by adopting a VGG-16 network framework (figure 2), wherein the VGG-16 network framework comprises 13 convolutional layers, 3 fully-connected layers and 5 pooling layers, the convolutional cores of all convolutional layers are 3 x 3, and the parameters of the pooling layers are 2 x 2. According to the convolution layer and the pooling layer, the VGG can be divided into different blocks, each block comprises a plurality of convolution layers and one pooling layer, the VGG-16 can be divided into 5 blocks, the number of channels of each convolution layer in each block is the same, and the number of channels of adjacent blocks is doubled while the width and the height of a feature diagram are reduced by half until the number of channels reaches 512. The back end of the network framework is three fully connected layers and a classification function.
In fig. 2, context denotes convolutional layer, ReLU denotes activation function, max _ power denotes maximum pooling, full _ connected denotes fully-connected layer, and softmax denotes prediction layer, i.e., classification function.
Therefore, by adopting the VGG-16 network framework, the convolution layers all adopt the same convolution kernel parameters, the pooling layers all adopt the same pooling kernel parameters, and the model is formed by stacking a plurality of convolution layers and pooling layers, so that a deeper network structure can be easily formed.
Preferably, in the present invention, in the step S03, the network structure of the UnSuperPoint (point of interest detector) specifically includes:
(1) setting an input end network, wherein the input end network is a full convolution neural network, and the full convolution neural network comprises two input convolution layers; wherein each of the input convolutional layers is subjected to two convolutions and one pooling.
The input end network in the UnSuperPoint is a full convolutional neural network similar to the VGG, but the full connection layer at the rear end of the input end network is cancelled, and two input convolutional layers are used instead (fig. 3). Fig. 3 is a schematic structural diagram of the network framework of the UnSuperPoint. Input represents the Input image and the backhaul is a different layer image structure formed using convolutional layers and pooling layers, each layer performing two convolution and one pooling operation.
Taking three-channel RGB images as input images, outputting a down-sampling feature map, and performing convolution twice and pooling once on each input convolution layer for three layers. And (3) forming a new image layer by convolution and pooling of each layer of the original image, wherein the width and the height of the new image layer are reduced by half relative to the image of the previous layer, and the number of channels is doubled at the same time until the number of channels reaches 256. The number of channels corresponding to each layer of image structure in the Backbone is as follows: 32-32-64-64-128-256, one pixel in the generated feature map corresponds to 8 × 8 area in the input image.
The Feature map is a processing result obtained by convolving the image layers of 256 channels and 256 channels, and is also called an incoming Feature map. And (3) performing convolution of three channels on the incoming Feature map to obtain a confidence coefficient sub-network (Score), a Feature point coordinate sub-network (Position XY) and a description point sub-network (Descriptor).
(2) And setting an output end network, wherein the output end network comprises a coordinate sub-network, a confidence sub-network and a description sub-network corresponding to the feature matching points. Each network is illustrated below:
for confidence sub-network (Score):
after the confidence sub-network (Score) convolves the incoming feature map, the confidence of each feature matching point is obtained by activating a function; wherein, the output result of the confidence coefficient of each feature matching point is a single channel, and the numerical value is in the interval [0,1 ]. The size is 1/8 of the input image, which means that there is at most one feature matching point in the 8 × 8 neighborhood of the input image, so that the neighborhood can be screened once directly in the feature matching point positioning stage, thereby achieving the similar effect to NMS without the need of processing in the subsequent process.
The confidence of each feature point in the confidence sub-network indicates the probability that the point is a feature point, the value is 1, the point is a feature point, the value is 0, the point cannot be a feature point, and the closer the value is to 1, the higher the probability that the point is a feature point.
For the feature point coordinate sub-network:
the sub-network of feature point coordinates records the position information of each feature point. In the Position sub-network (Position XY), the output result is two channels, the value is also between the interval [0,1], and the size is 1/8 of the input image. And, the numerical value corresponding to each coordinate (r, c) of the output result is converted into the coordinate formula of the feature matching point in the image as follows:
Pmap,x(r,c)＝(c+Prelative,x(r,c))·fdownsample
Pmap,y(r,c)＝(r+Prelative,y(r,c))·fdownsample
wherein, the image down-sampling multiplying power fdownsample＝8，Pmap,x(r,c)、Pmap,y(r, c) respectively representing coordinate values of the feature matching points in the image on the x and y axes; prelative,x(r,c)、Prelative,y(r, c) respectively representing the offset of the feature matching point in the image relative to the row and column coordinates of c and r on the x and y axes; (r, c) is a known quantity, Prelative,x(r,c)、Prelative,yAnd (r, c) are unknown quantities and are obtained by feature extraction.
For a description sub-network:
the descriptor subnetwork is used for providing a multi-dimensional correlation vector for each feature matching point. Moreover, the output result of the descriptor subnetwork is 256 channels, a 256-dimensional vector is provided for each feature matching point, the feature information of each feature point in the 256-dimensional channel is described, the vector can be directly used as a rough descriptor of the corresponding feature matching point, and the descriptor can be obtained by interpolating according to the feature matching point coordinates calculated by the coordinate subnetwork.
In this step, the feature points extracted finally are determined according to the confidence of each feature point in the confidence sub-network and the set standard.
As shown in fig. 3, the output Outputs the confidence (Smap), the coordinate value (Pmap) and the descriptor (Fmap) of the feature point extracted last. And reordering all input feature points according to the confidence (Score) to obtain feature points ordered according to the confidence, and only the first N feature points with the most obvious confidence can be reserved as the final output result, namely the final extracted N feature points (Top N points).
And then, matching all the extracted feature points according to the feature point confidence coefficient (Smap), the coordinate value (Pmap) and the descriptor (Fmap) information in Outputs, implementing a matching process, and searching the feature points with equivalent confidence coefficient and closest descriptors as matching point pairs.
In order to determine the network parameters of the UnSuperPoint, a multispectral image A needs to be input first, an image B is generated after the multispectral image A rotates for a plurality of angles, feature point extraction and matching processing are carried out between the image A and the image B, self-adaptive learning and training are achieved, and supervision conditions are not needed.
After the network parameters of the UnSuperPoint are determined, matching the multispectral images M and N with the two wave bands determined by wave band selection in S02 by using a trained network.
S04: performing tremor analysis, and calculating registration error for each feature matching point; analyzing line by line to obtain the change of the registration error along with the imaging line; and carrying out spectrum analysis on the satellite platform tremor time sequence data through Fourier transform to obtain the spectrum characteristics of the registration error caused by tremor. The spectral features of the registration error specifically include spectral features of the registration error, such as period, frequency, amplitude, and the like.
Preferably, in this step, calculating the registration error for each feature matching point includes:
and if V characteristic matching point pairs exist in the u scanning line, the registration error calculation formula of the kth homonymous image point of the u scanning line is as follows:
wherein M, N are two wavebands in step S02; Δ xu,k、Δyu,kRespectively representing registration errors in the CCD direction and in the push-scan direction;
respectively shows that the k-th homonymous image point of the u-th scanning line is at M,Coordinates on the N two band images; Δ x, Δ y represent the designed values of the translation between the CCDs of the M, N two wavebands in the CCD direction and in the push-scan direction.
Preferably, in this step, performing line-by-line analysis, and obtaining the change of the registration error with the imaging line includes:
taking the registration error of all feature matching points of each scanning line to obtain an average value as the optimal estimation of the registration error corresponding to the scanning line; and obtaining the time sequence data of the registration error along the CCD direction and the push-scan direction along the change of the imaging line through line-by-line statistics.
Preferably, in this step, performing spectrum analysis on the satellite platform tremor time series data through fourier transform, and acquiring the spectrum characteristics of the registration error caused by tremor includes:
performing spectrum analysis on time series data corresponding to the registration error by adopting Fourier transform, wherein the integral time of each scanning line of the multispectral image is delta T, the total imaging time T of the image containing U imaging lines is delta t.U, and the imaging time T corresponding to the U-th imaging line is Tu＝t0+ Δ t · U (U ═ 0,1, …, U-1), where t is0Is the initial imaging time;
fourier transforming a discrete sequence of Δ x (U) (U-0, 1, …, U-1):
wherein K represents the decomposed serial number and takes the value of 0-U/2; a isKAnd bKRespectively representing the amplitudes of the K-th harmonic cosine and sine functions; c. CKRepresents the amplitude of the K harmonic;
frequency f corresponding to K harmonicKIs composed of
And obtaining a spectrum curve corresponding to the spectrum characteristic of the registration error along the CCD direction. Changing Δ x (u) into Δ y (u), and calculating according to a formula to obtain a spectrum curve of the registration error in the push-broom direction.
In the invention, since the platform tremor is represented by periodic motion, the registration error caused by the platform tremor is also periodic variation, and the non-zero frequency at the peak of the spectrum curve of the registration error is the satellite tremor frequency.
Preferably, in the present invention, the frequency and amplitude obtained by fourier transform are used as initial values of a sine function, and the registration error frequency, amplitude and initial phase of the registration error are determined by using least square estimation.
The method is based on a deep learning method to determine the relevant information of the image feature matching points and carry out tremor analysis; the image feature matching points have high stability, high precision and strong representativeness, the tremor analysis is carried out by determining the relevant information of the feature matching points, and the data obtained by the method has high precision and high confidence level and meets the requirements of a satellite remote sensing system.
The embodiment of the invention also provides a deep learning-based satellite platform tremor detection and analysis system, and the tremor detection and analysis system is used for executing the deep learning-based satellite platform tremor detection and analysis method.
Specifically, the tremor detection and analysis system comprises the following modules:
the data preparation module is used for preprocessing the original multispectral image; the pretreatment comprises the following steps: relative radiation correction and line frequency normalization;
the device comprises a wave band selection module, a detection module and a control module, wherein the wave band selection module is used for determining two wave bands for detection;
and the image matching module extracts the feature matching point pairs and generates feature matching point coordinates, feature matching point confidence degrees and feature matching point descriptors by adopting a deep learning matching algorithm based on a convolutional neural network.
A tremor analysis module for calculating a registration error for each feature matching point; analyzing line by line to obtain the change of the registration error along with the imaging line; and carrying out spectrum analysis on the satellite platform tremor time sequence data through Fourier transform to obtain the spectrum characteristics of the registration error caused by the tremor.
Finally, it should be noted that: the above embodiments are only for illustrating the technical solutions of the present invention and not for limiting the same, and although the present invention is described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: modifications and equivalents may be made to the embodiments of the invention without departing from the spirit and scope of the invention, which is to be covered by the claims.
Claims (9)
1. The satellite platform tremor detection and analysis method based on deep learning is characterized by comprising the following steps:
s01: data preparation, namely preprocessing an original multispectral image; the pretreatment comprises the following steps: relative radiation correction and line frequency normalization;
s02: selecting wave bands, and determining two wave bands for detection;
s03: image matching, namely extracting feature matching point pairs by adopting a deep learning matching algorithm based on a convolutional neural network, and generating feature matching point coordinates, feature matching point confidence coefficients and feature matching point descriptors;
s04: performing tremor analysis, and calculating registration error for each feature matching point; analyzing line by line to obtain the change of the registration error along with the imaging line; and carrying out spectrum analysis on the satellite platform tremor time sequence data through Fourier transform to obtain the spectrum characteristics of the registration error caused by tremor.
2. The deep learning-based satellite plateau tremor detection analysis method of claim 1, wherein: in step S02, the two bands are adjacent bands.
3. The deep learning-based satellite plateau tremor detection analysis method of claim 1, wherein: the step of S03 includes: constructing a VGG-16 network framework based on a deep learning matching algorithm;
dividing the VGG-16 network framework into different blocks according to convolutional layers and pooling layers included in the VGG-16 network framework, wherein each block comprises a plurality of convolutional layers and one pooling layer; the front end of the network frame is connected with the blocks, and the rear end of the network frame is sequentially provided with a full connection layer and a classification function.
4. The deep learning-based satellite plateau tremor detection analysis method of claim 3, wherein: the step of S03 further includes:
setting an input end network, wherein the input end network is a full convolution neural network, and the full convolution neural network comprises two input convolution layers; wherein each of the input convolutional layers is subjected to two convolutions and one pooling.
5. The deep learning-based satellite plateau tremor detection analysis method of claim 4, wherein: the step of S03 further includes:
setting an output end network, wherein the output end network comprises a coordinate sub-network, a confidence sub-network and a description sub-network corresponding to the feature matching points;
after the confidence sub-network convolves the input feature map, the confidence of each feature matching point is obtained through an activation function;
the numerical value corresponding to each coordinate (r, c) in the coordinate subnetwork is converted into the coordinate formula of the feature matching point in the image as follows:
Pmap,x(r,c)＝(c+Prelative,x(r,c))·fdownsample
Pmap,y(r,c)＝(r+Prelative,y(r,c))·fdownsample
wherein, the image down-sampling multiplying power fdownsample＝8，Pmap,x(r,c)、Pmap,y(r, c) respectively representing coordinate values of the feature matching points in the image on the x and y axes; prelative,x(r,c)、Prelative,y(r, c) respectively representing the offset of the feature matching points in the image relative to the row and column coordinates of c and r on the x and y axes;
the descriptor subnetwork is used for providing a multi-dimensional correlation vector for each feature matching point.
6. The deep learning-based satellite plateau tremor detection analysis method of claim 1, wherein: in step S04, the calculating the registration error for each feature matching point includes:
assuming that there are V feature matching point pairs in the u-th scanning line, the registration error calculation formula of the k-th homologous image point of the u-th scanning line is as follows:
wherein M, N are two wave bands in the step S02; Δ xu,k、Δyu,kRepresenting the registration errors in the CCD direction and in the push-scan direction, respectively;
respectively representing the coordinates of the kth homonymous image point of the u scanning line on M, N two wave band images; Δ x, Δ y represent M, N designed values for the translation between the CCDs of the two wavebands in the CCD direction and in the push-scan direction.
7. The deep learning-based satellite plateau tremor detection analysis method of claim 6, wherein: in step S04, the analyzing line by line, and obtaining the variation of the registration error with the imaging line includes:
taking the registration error of all the feature matching points of each scanning line to obtain an average value, and taking the average value as the optimal estimation of the registration error corresponding to the scanning line; and obtaining the time sequence data of the registration errors along the CCD direction and the push-scanning direction along the change of the imaging lines through line-by-line statistics.
8. The deep learning-based satellite plateau tremor detection analysis method of claim 7, wherein: in step S04, the performing spectral analysis on the satellite platform tremor time series data through fourier transform to obtain the spectral characteristics of the registration error caused by tremor includes:
performing spectrum analysis on the time series data corresponding to the registration error by using the Fourier transform, wherein the integration time of each scanning line of the multispectral image is delta T, the total imaging time T of the image comprising U imaging lines is delta T-U, and the imaging time T corresponding to the U-th imaging line is Tu＝t0+ Δ t · U (U ═ 0,1, …, U-1), where t is0Is the initial imaging time;
performing the fourier transform on a discrete sequence of Δ x (U) (U-0, 1, …, U-1):
wherein K represents the decomposed serial number and takes the value of 0-U/2; a isKAnd bKRespectively representing the amplitudes of the K-th harmonic cosine and sine functions; c. CKRepresents the amplitude of the K harmonic;
frequency f corresponding to K harmonicKComprises the following steps:
and obtaining a spectrum curve corresponding to the spectrum characteristic of the registration error along the CCD direction.
9. The deep learning-based satellite plateau tremor detection analysis method of claim 8, wherein: and determining the registration error frequency, the registration error amplitude and the registration error initial phase of the registration error by using least square estimation by taking the frequency and the amplitude obtained according to the Fourier transform as initial values of a sine function.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110137673.1A CN113029318A (en) | 2021-02-01 | 2021-02-01 | Satellite platform tremor detection and analysis method based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110137673.1A CN113029318A (en) | 2021-02-01 | 2021-02-01 | Satellite platform tremor detection and analysis method based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN113029318A true CN113029318A (en) | 2021-06-25 |
Family
ID=76459671
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110137673.1A Pending CN113029318A (en) | 2021-02-01 | 2021-02-01 | Satellite platform tremor detection and analysis method based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113029318A (en) |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1602056A (en) * | 2004-10-18 | 2005-03-30 | 海信集团有限公司 | TV circuit realizing high resolution 1080P/60 format |
| FR2930820A1 (en) * | 2008-04-30 | 2009-11-06 | Centre Nat Etd Spatiales | VIBRATION CHARACTERIZATION METHOD FOR OBSERVATION SATELLITE |
| CN101827223A (en) * | 2010-04-20 | 2010-09-08 | 武汉大学 | Inner field stitching method of non-collinear TDI CCD imaging data based on line frequency normalization |
| CN103983343A (en) * | 2014-05-29 | 2014-08-13 | 武汉大学 | Satellite platform chattering detection method and system based on multispectral image |
| CN110766024A (en) * | 2019-10-08 | 2020-02-07 | 湖北工业大学 | Visual odometer feature point extraction method based on deep learning and visual odometer |
| CN110866448A (en) * | 2019-10-21 | 2020-03-06 | 西北工业大学 | Flutter signal analysis method based on convolutional neural network and short-time Fourier transform |
-
2021
- 2021-02-01 CN CN202110137673.1A patent/CN113029318A/en active Pending
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1602056A (en) * | 2004-10-18 | 2005-03-30 | 海信集团有限公司 | TV circuit realizing high resolution 1080P/60 format |
| FR2930820A1 (en) * | 2008-04-30 | 2009-11-06 | Centre Nat Etd Spatiales | VIBRATION CHARACTERIZATION METHOD FOR OBSERVATION SATELLITE |
| CN101827223A (en) * | 2010-04-20 | 2010-09-08 | 武汉大学 | Inner field stitching method of non-collinear TDI CCD imaging data based on line frequency normalization |
| CN103983343A (en) * | 2014-05-29 | 2014-08-13 | 武汉大学 | Satellite platform chattering detection method and system based on multispectral image |
| CN110766024A (en) * | 2019-10-08 | 2020-02-07 | 湖北工业大学 | Visual odometer feature point extraction method based on deep learning and visual odometer |
| CN110866448A (en) * | 2019-10-21 | 2020-03-06 | 西北工业大学 | Flutter signal analysis method based on convolutional neural network and short-time Fourier transform |
Non-Patent Citations (3)
| Title |
|---|
| PETER HC, ET AL: "UNSUPERPOINT: END-TO-END UNSUPERVISED INTEREST POINT DETECTOR AND DESCRIPTOR", 《ARXIV》 * |
| 杨露菁: "《智能图像处理及应用》", 31 March 2019, 中国铁道出版社有限公司 * |
| 范大昭 等: "卫星影像匹配的深度卷积神经网络方法", 《测绘学报》 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US20110194734A1 (en) | Method for capturing images comprising a measurement of local motions | |
| CN110211044B (en) | Multispectral imaging method and device based on demosaicing algorithm and principal component analysis | |
| CN112946679B (en) | Unmanned aerial vehicle mapping jelly effect detection method and system based on artificial intelligence | |
| CN107274441B (en) | Wave band calibration method and system for hyperspectral image | |
| CN115236655B (en) | Landslide identification method, system, equipment and medium based on fully-polarized SAR | |
| CN112686935A (en) | Airborne depth sounding radar and multispectral satellite image registration method based on feature fusion | |
| CN113570536B (en) | Panchromatic and multispectral image real-time fusion method based on CPU and GPU cooperative processing | |
| CN111583330B (en) | Multi-scale space-time Markov remote sensing image sub-pixel positioning method and system | |
| CN113763271A (en) | High-quality spectrum denoising method based on physical noise model | |
| CN110288050B (en) | Hyperspectral and LiDar image automatic registration method based on clustering and optical flow method | |
| CN110956601B (en) | Infrared image fusion method and device based on multi-sensor mode coefficients and computer readable storage medium | |
| CN111062267A (en) | Time series remote sensing image dimension reduction method | |
| CN110738693A (en) | ground-based imaging radar multi-angle image registration method | |
| CN111932452B (en) | Infrared image convolution neural network super-resolution method based on visible image enhancement | |
| CN113029318A (en) | Satellite platform tremor detection and analysis method based on deep learning | |
| CN116309139A (en) | Demosaicing method suitable for one-to-many infrared multispectral image | |
| CN116109535A (en) | Image fusion method, device and computer readable storage medium | |
| CN114998708A (en) | Tea type identification method and device based on map signals | |
| CN114565653A (en) | Heterogeneous remote sensing image matching method with rotation change and scale difference | |
| CN114494039A (en) | Underwater hyperspectral push-broom image geometric correction method | |
| JP2007066050A (en) | Tree crown circle extraction system | |
| CN113670440A (en) | Compressed spectrum imaging method based on adaptive dictionary | |
| CN112907495A (en) | Spatial resolution enhancement method and device for hyperspectral image and electronic equipment | |
| CN115311363B (en) | High-resolution satellite image assistance-based footprint camera orientation method and system | |
| CN115564808B (en) | Multi-resolution hyperspectral/SAR image registration method based on public space-spectrum subspace |
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 |
Application publication date: 20210625 |
|
| RJ01 | Rejection of invention patent application after publication |