CN106845343B - Automatic detection method for optical remote sensing image offshore platform - Google Patents
Automatic detection method for optical remote sensing image offshore platform Download PDFInfo
- Publication number
- CN106845343B CN106845343B CN201611161835.0A CN201611161835A CN106845343B CN 106845343 B CN106845343 B CN 106845343B CN 201611161835 A CN201611161835 A CN 201611161835A CN 106845343 B CN106845343 B CN 106845343B
- Authority
- CN
- China
- Prior art keywords
- remote sensing
- neural network
- sensing image
- random field
- image
- 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.)
- Active
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
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Multimedia (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Astronomy & Astrophysics (AREA)
- Remote Sensing (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses an automatic detection method for an offshore platform of an optical remote sensing image. And then, the signal-to-noise ratio and the signal-to-clutter ratio of the offshore platform detection are improved through a space-time Markov random field model, and a target saliency image is output. And finally, constructing a large-scale target standard data set, performing multi-node parallel architecture training on the deep convolutional neural network based on a hadoop distributed file storage system, continuously optimizing the deep convolutional neural network, extracting the stability characteristics of a large-scale target, and realizing the rapid and accurate automatic detection of the offshore oil-gas platform in the mass remote sensing images.
Description
Technical Field
The invention belongs to the field of remote sensing image interpretation, and particularly relates to an automatic detection method for an optical remote sensing image offshore platform.
Background
1.1 detection method based on relative fixing of geographic positions of offshore oil and gas platforms
The detection is realized according to the characteristics that the geographic position of the offshore oil and gas platform is relatively fixed and the ship target continuously moves. Detecting an offshore target by adopting a CFAR (computational fluid dynamics) method based on SAR (synthetic aperture radar) image data of a European environment satellite such as Casadio S, and extracting an offshore oil-gas platform according to the position invariance of the target; cheng L, and the like adopt a double-parameter CFAR to detect an offshore target, and then extract an offshore platform according to the principle that the relative triangular position of the fixed target is unchanged; yongxue Liu et al use the multispectral data of Landsat land imager (OLI) to detect offshore oil and gas platforms based on context characteristics and invariance of position and scale.
1.2 detection method based on infrared radiation characteristics of offshore oil and gas platform
The second method realizes target detection through the infrared radiation characteristic of the offshore oil and gas platform. Most offshore oil and gas platforms burn waste gas through emptying torches, and these offshore platforms can be effectively detected through extracting fire points in infrared wave bands. Global natural gas fire detection is performed by OLS (Operational Line-scan System) sensor data of U.S. defense weather satellites such as Elvidge; casadio and the like perform radiation feature extraction on an SAR image detection target through ATSR (altitude tracking Scanning Radiometer) sensor data to detect an offshore platform in the North sea area of the Atlantic sea; anejionu et al, through satellite data of Landsat and MODIS, employ radiation filtering and spatial filtering to detect the offshore platform burning natural gas; the method comprises the steps of extracting the offshore oil and gas platform by adopting multispectral data of a Landsat (TM) sensor through optimal threshold selection and a sliding window method by the aid of Mongolian and the like.
The two types of offshore oil and gas platform detection methods firstly carry out target detection on images in a single time phase, obviously, under the condition, the overall detection rate is directly influenced by the detection rate of the images in the single time phase, and the final missed detection or false alarm can be caused by the detection error of a certain time phase. And for offshore oil and gas platforms, the detection rate under the single-time phase is difficult to guarantee, mainly because: 1) the offshore oil and gas platform has smaller scale, the length and the width of the offshore oil and gas platform are about 100 meters generally, and the offshore oil and gas platform only has dozens of pixels in the remote sensing image with the common resolution; 2) the imaging signal of the offshore oil and gas platform is weak and is difficult to be obviously distinguished from the background; 3) the sea environment is complex, and clutter interference such as sea waves, cloud layers and the like is serious; 4) similar to the imaging characteristics of ships, the imaging system is easily interfered by ship targets.
Disclosure of Invention
The invention aims to solve the technical problems of low target detection accuracy and large workload when detecting an offshore oil and gas platform from massive remote sensing image data, and provides an offshore platform detection method based on a deep convolutional neural network and a three-dimensional space-time Markov random field.
The invention adopts the following technical scheme to solve the technical problems
An optical remote sensing image offshore platform automatic detection method specifically comprises the following steps:
step 1, analyzing the multi-resolution and multi-viewpoint remote sensing imaging characteristics of an offshore oil and gas platform, and fitting a remote sensing image through a characteristic subspace and a principal component;
step 2, improving the signal-to-noise ratio and the signal-to-clutter ratio of offshore platform detection through a space-time Markov random field model, and outputting a target saliency image;
and 3, constructing a target standard data set, performing multi-node parallel architecture training based on a hadoop distributed file storage system on the deep convolutional neural network, continuously optimizing the deep convolutional neural network, extracting the stability characteristic of a large-scale target, and realizing quick and accurate automatic detection of the offshore oil and gas platform in the mass remote sensing images.
As a further preferable scheme of the optical remote sensing image offshore platform automatic detection method of the invention, the step 1 specifically comprises the following steps:
step 1.1, selecting a multi-temporal optical remote sensing image set under different resolutions and multiple viewpoints;
step 1.2, establishing a priori knowledge database of a marine detection identification area;
step 1.3, processing the input image set in the step 1 through the priori knowledge database established in the step 2, and removing non-marine region information in remote sensing data and information of a marine deeper region which cannot be developed by marine oil and gas to obtain a processed image set;
step 1.4, manually calibrating the image set obtained in the step 3 to obtain a marked remote sensing image set;
and 1.5, carrying out simulation analysis on the characteristic subspace and the main component of the marked remote sensing image set to obtain a multi-temporal remote sensing image set.
As a further preferable scheme of the optical remote sensing image offshore platform automatic detection method of the invention, the step 2 specifically comprises the following steps:
step 2.1, carrying out time domain low-pass filtering on the images in the image set, and initializing a Markov random field;
step 2.2, establishing a three-dimensional Markov random field model according to the input multi-temporal remote sensing image, and considering that the gray distribution information of the remote sensing image conforms to a finite Gaussian mixture model to obtain joint condition probability;
step 2.3, carrying out maximum likelihood estimation according to the conditional probability in the step 2.2, and calculating Markov random field model parameters;
step 2.4, updating the Markov random field by using the calculated parameters, and acquiring the Markov random field model parameters again through maximum likelihood estimation;
step 2.5, comparing whether the parameter errors of the Markov random field model calculated twice before and after meet the convergence; if yes, executing step 2.6, otherwise, continuing to execute step 2.4;
and 2.6, calculating and outputting a target saliency image through a three-dimensional space-time Markov random field model.
As a further preferable scheme of the optical remote sensing image offshore platform automatic detection method of the invention, the step 3 specifically comprises the following steps:
step 3.1, a platform sample base is obtained and then stored in a hadoop distributed file storage system, sample base subsets are stored in k nodes, each node trains the sample base independently, and each parameter of the convolutional neural network is initialized;
step 3.2, simultaneously carrying out complete network training on the sample library subsets by each node to obtain a parameter vector of the convolutional neural network, and fitting the parameters of the whole neural network by using the neural network parameters of each node;
3.3, updating the neural network parameters of each node by using the overall neural network parameters;
step 3.4, determining the training effect through manual interaction, continuously adjusting parameters, and repeatedly executing the step 3.2 to the step 3.3 until the training meets the requirements;
step 3.5, adding a target sample, and repeating the step 3.2 to the step 3.4;
and 3.6, detecting the target saliency image output in the step 2.6 by using the convolutional neural network trained in the step 3.5, and outputting a target detection result.
As a further preferable scheme of the optical remote sensing image offshore platform automatic detection method of the invention, in step 1.2, the data database comprises the latitude and longitude of the offshore and the ocean depth.
Compared with the prior art, the invention adopting the technical scheme has the following technical effects:
1) an analysis method of optical remote sensing imaging characteristics of the offshore platform under multiple resolutions and multiple viewpoints is established, and discrimination and distinguishing of targets and false alarm targets are facilitated;
2) establishing a multi-temporal three-dimensional space-time Markov random field model, effectively outputting a target saliency image, and remarkably inhibiting image clutter and noise interference;
3) the deep convolutional neural network is built based on hadoop distributed file storage and trained, training efficiency is greatly improved, and high accuracy of target detection of the offshore platform is achieved.
Drawings
FIG. 1 is a block diagram of an offshore platform inspection method;
FIG. 2 is a three-dimensional Markov random field model;
FIG. 3 is a diagram of a convolutional neural network architecture;
FIG. 4 is a diagram of a convolutional neural network parallel computing architecture;
FIG. 5 is a logic diagram of simulation analysis of multi-temporal remote sensing data;
FIG. 6 is a logic block diagram of three-dimensional Markov random field model modeling;
FIG. 7 is a logic block diagram of a deep convolutional neural network parallel training architecture.
Detailed Description
The technical scheme of the invention is further explained in detail by combining the attached drawings:
the technical scheme adopted by the invention is as follows: firstly, the multi-resolution and multi-viewpoint remote sensing imaging characteristics of the offshore oil and gas platform are analyzed, and the optimal remote sensing image is fitted through a characteristic subspace and a principal component. And then, the signal-to-noise ratio and the signal-to-clutter ratio of the offshore platform detection are improved through a space-time Markov random field model, and a target saliency image is output. And finally, constructing a large-scale target standard data set, performing multi-node parallel architecture training on the deep convolutional neural network based on a hadoop distributed file storage system, continuously optimizing the deep convolutional neural network, extracting the stability characteristics of a large-scale target, and realizing the rapid and accurate automatic detection of the offshore oil-gas platform in the mass remote sensing images.
Step 1, analyzing the multi-resolution and multi-viewpoint remote sensing imaging characteristics of an offshore oil and gas platform, and fitting a remote sensing image through a characteristic subspace and a principal component;
step 2, improving the signal-to-noise ratio and the signal-to-clutter ratio of offshore platform detection through a space-time Markov random field model, and outputting a target saliency image;
and 3, constructing a large-scale target standard data set, performing multi-node parallel architecture training on the deep convolutional neural network based on a hadoop distributed file storage system, continuously optimizing the deep convolutional neural network, extracting the stability characteristics of a large-scale target, and realizing the quick and accurate automatic detection of the offshore oil and gas platform in the mass remote sensing images.
As a further preferable scheme of the optical remote sensing image offshore platform automatic detection method of the invention, the step 1 specifically comprises the following steps:
step 1.1, selecting a certain number of multi-temporal optical remote sensing image sets A1 with different resolutions and under multiple viewpoints and with obvious characteristics;
step 1.2, establishing a prior knowledge database SQ of a large-range ocean detection identification area;
step 1.3, processing the input image set A1 in the step 1 through the prior knowledge database SQ established in the step 2, and removing non-ocean area information in remote sensing data and information of ocean deeper areas which cannot be developed by offshore oil and gas to obtain a processed image set A2;
step 1.4, manually calibrating the image set A2 obtained in the step 3 to obtain a marked remote sensing image set A3;
and step 1.5, carrying out simulation analysis on the characteristic subspace and the main component of the marked remote sensing image set to obtain a multi-temporal remote sensing image set A4.
As a further preferable scheme of the optical remote sensing image offshore platform automatic detection method of the invention, the step 2 specifically comprises the following steps:
step 2.1, carrying out time domain low-pass filtering on the images in the image set A4, and initializing a Markov random field;
step 2.2, establishing a three-dimensional Markov random field model according to the input multi-temporal remote sensing image, and considering that the gray distribution information of the remote sensing image conforms to a finite Gaussian mixture model to obtain joint condition probability;
step 2.3, carrying out maximum likelihood estimation according to the conditional probability in the step 2.2, and calculating Markov random field model parameters;
step 2.4, updating the Markov random field by using the calculated parameters, and acquiring the Markov random field model parameters again through maximum likelihood estimation;
step 2.5, comparing whether the parameter errors of the Markov random field model calculated twice before and after meet the convergence; if yes, executing step 2.6, otherwise, continuing to execute step 2.4;
and 2.6, calculating and outputting a target saliency image through a three-dimensional space-time Markov random field model.
As a further preferable scheme of the optical remote sensing image offshore platform automatic detection method of the invention, the step 3 specifically comprises the following steps:
step 3.1, a platform sample base is obtained and stored in a hadoop distributed file storage system, sample base subsets are stored in k nodes, each node trains the sample base independently, and each parameter of the convolutional neural network is initialized;
step 3.2, simultaneously carrying out complete network training on the sample library subsets by each node to obtain a parameter vector of the convolutional neural network, and fitting the parameters of the whole neural network by using the neural network parameters of each node;
3.3, updating the neural network parameters of each node by using the overall neural network parameters;
step 3.4, determining the training effect through manual interaction, continuously adjusting parameters, and repeatedly executing the step 3.2 to the step 3.3 until the training meets the requirements;
step 3.5, adding a target sample, and repeating the step 3.2 to the step 3.4;
and 3.6, detecting the target saliency image output in the step 2.6 by using the convolutional neural network trained in the step 3.5, and outputting a target detection result.
FIG. 1 is a main structure block diagram of the method of the present invention, and the detection method of the remote sensing image offshore platform based on the deep convolutional neural network and the three-dimensional Markov random field mainly comprises three parts: the method comprises the following steps of firstly, carrying out analysis simulation and fitting on remote sensing imaging characteristics of an offshore platform; a second portion; modeling a three-dimensional space-time Markov random field of the multi-temporal offshore platform; and in the third part, parallel architecture training based on a deep convolutional neural network and automatic extraction of the offshore platform. The output of the first section serves as the input of the second section, and the output of the second section serves as the input of the platform detection section in the third section.
FIG. 2 is a three-dimensional space-time Markov random field model diagram, in which single-frame remote sensing images of a two-dimensional coordinate system are arranged according to space and time to form a three-dimensional image model, and any pixel in the model is a neighboring point with 8 space domains and a neighboring point with 18 time domains.
FIG. 3 is a diagram of the architecture of a convolutional neural network, which is used primarily to identify two-dimensional patterns of displacement, scaling and other forms of distortion invariance, and is a multi-layered neural network, each layer consisting of multiple two-dimensional planes, each plane consisting of multiple independent neurons. In the figure, C1, S2, C3 and S4 are feature images, C1 and C3 are convolutional layer feature images, the emphasis is on feature extraction, and S2 and S4 are sampling layer feature images and the emphasis is on feature calculation.
FIG. 4 is a diagram of a convolutional neural network parallel computing architecture. A large number of training samples are stored on different storage nodes, the samples are independent from each other, and each storage node is provided with a complete convolutional neural network, so that the neural network on each node is only responsible for training a part of sample sets, and the nodes are updated once after being processed. In the training process of the samples, the samples are processed in a unidirectional layer-by-layer mode among all layers of the network, the feature maps, the neurons and the neuron weights of the same layer are independent, and therefore a plurality of samples can be calculated in a convolutional neural network in a parallel mode.
FIG. 5 is a logic diagram of simulation analysis of multi-temporal remote sensing data, which is implemented by first determining the geographical location of the remote sensing data according to the input remote sensing data, establishing a geographic information database of the remote sensing data according to prior knowledge, and removing non-ocean parts in the input remote sensing data. And secondly, manually calibrating a platform target, non-targets such as ships, reefs and the like on the processed image. And then, carrying out characteristic subspace and principal component analysis on the remote sensing image, and selecting the optimal waveband image.
FIG. 6 is a logic block diagram of a three-dimensional spatiotemporal Markov random field model modeling to initialize a Markov random field by performing a temporal low pass filtering on an input image. The gray information of the remote sensing image accords with the finite mixed Gaussian distribution, the conditional probability of the three-dimensional Markov random field model is obtained, the model parameters are solved through maximum likelihood estimation, the model is updated, and when the maximum likelihood estimation is converged, the target saliency image is output.
Fig. 7 is a logic block diagram of a deep convolutional neural network parallel training architecture, in which a sample library is stored in a distributed manner, training of the deep convolutional neural network is performed on each storage node at the same time, model parameters are continuously optimized through manual interactive confirmation, and a sample library is added for training. When an ideal convolutional neural network is achieved, the input image is rapidly detected, and a detection result is output.
The technical solution of the present invention and the scientific principles underlying it are explained in detail below.
1. Principal component analysis principle in remote sensing imaging characteristic analysis of offshore platform:
①, assuming that n frames of remote sensing image data are input, converting an ith frame of image into a vector with the length of L, and further establishing an L multiplied by n image two-dimensional matrix C ═ C1,c2,…cn}。
secondly, considering that the data of the optical remote sensing image is large, assuming that n is less than or equal to L, the mean value of the two-dimensional matrix image CIs composed of
For convenience of description, will be
Is marked as
To pair
Singular value decomposition is carried out to obtain
Where Σ is a diagonal matrix whose diagonal elements σ1,…,σnIs that
And the element values are to be ordered, i.e. σ1≥σ2≥σ3≥,…,≥σn(ii) a U is an orthogonal matrix, and the column vector in U is
The eigenvector of (a), and sigma in sigma1,…,σnWant to correspond; v is an orthogonal matrix, and the column vector in V is
The feature vector of (2).
for diagonal matrix sigma, when the value of i is greater than a certain value, the value of sigma i is very small, at this time, the first m columns of U matrix are retained to obtain new matrix UmI.e. the principal component data of the remote sensing image.
2. The principle of the space-time three-dimensional Markov random field model of the remote sensing image is as follows:
(1) markov model principle
The Markov random field contains two layers of meaning of Markov property and random field. The markov property means that future states are only related to the current state, and when each position is randomly assigned a value of the phase space according to some distribution, the totality is called a random field.
let S { (i, j) |1 ≦ i ≦ M,1 ≦ j ≦ N } represent a finite set of points for MN locations, i.e., locations in the random field, Λ ═ {1,2,3, …, L } represents a state space, i.e., a phase space in the random field, X ═ { X ≦ MsThe expression, | S ∈ S } is defined in
Random field of (a), xsHidden state random variables with state space Λ represented on random field X, i.e. XsE Λ. In an image, a set of grid points S represents the location of a pixel, X is called a label field, usually a set of pixel values, and Λ is a labeled random variable XsL denotes the number of different regions into which the image is divided.
let δ ═ { δ (S) | S ∈ S } be the set of general neighborhood systems defined on S, which satisfy the following characteristics:
the position r ∈ δ(s) is called the neighborhood of s, δ(s) is called the set of neighborhoods of s. In the invention, a neighborhood system is defined according to the Euclidean distance to the pixel:
δ(n)(s)＝{r|d(s,r)≤n,r≠s} (3)
in the formula (3), n is the order of the neighborhood system, and d (·) is expressed by the euclidean distance. For the
Satisfy the characteristic
There are different neighborhood structures in S, on which there is a single pixel or a subset of pixels and its neighbors
Called a sub-cluster, the set of sub-clusters C is denoted by C.
let δ be neighborhood system on S, when random field X ═ XsI S belongs to S and satisfies the following conditions:
x is called a markov random field with δ as the neighborhood system.
(2) Iterative estimation of markov random field parameters:
establishing a distribution relation between a Markov random field and Gibbs in order to determine the prior probability of a label field and the neighborhood local relation of the label field, wherein the Gibbs distribution satisfies the following joint probability distribution form:
P(X＝x)＝(1/Z)exp[-U(x)](5)
wherein the content of the first and second substances,
called energy function, Vc(x) Is the sub-cluster potential function that is only related to the pixel values within sub-cluster c,
referred to as the allocation function. The equivalent conditions for distribution of Gibbs in MRF are:
secondly, segmenting the image by using the Markov random field is to correctly express potential image labels so as to reach the maximum posterior probability (MAP):
and thirdly, using the gray information of the image as the characteristic of the remote sensing image, wherein the gray information of the remote sensing image conforms to the following limited Gaussian mixture model.
Wherein the gray scale of the image is divided into I areas, ciIs a component weight, muiIs taken as the mean value of the average value,
for variance, Γ (·) is a Gamma function.
and carrying out maximum likelihood estimation on the finite mixed Gaussian distribution probability function of the image gray level, and carrying out iterative computation until the maximum likelihood estimation error of the previous time and the next time is less than α (required precision value), so as to obtain the maximum posterior probability, further obtain the correct image label and obtain the target saliency image.
(3) The principle of the three-dimensional space-time Markov random field model is as follows:
and (4) the elevation information is not considered, and a time-space domain Markov random field model of the multi-time phase remote sensing image is established by combining the remote sensing image data in the space domain and the multi-time phase data in the time domain. The three-dimensional space-time Markov model is shown in FIG. 2: for points on the two-bit image, there are 8 spatial neighbors and 18 temporal neighbors, and the markov field distribution is described by the Gibbs random field.
In the method, the sub-cluster potential function of the space-time neighborhood system is defined as follows:
β is the current airspace radical parameter, βt-1and betat+1Radical parameters. Establishing a joint probability distribution function:
P(f|xs)＝Pt-1(f|xs)Pt(f|xs)Pt+1(f|xs) (10)
and (3) carrying out iterative computation by using the maximum likelihood estimation method in the step (2) to obtain the maximum posterior probability, and further obtaining the correct image label.
3. The parallel architecture rapid target extraction principle of the deep convolutional neural network is as follows:
(1) convolutional Neural Networks (CNN):
the convolutional neural network is mainly used for identifying two-dimensional graphs with displacement, scaling and other forms of distortion invariance, and is a multi-layer neural network, each layer is composed of a plurality of two-dimensional planes, and each plane is composed of a plurality of independent neurons. The general structure of a convolutional neural network is shown in fig. 3.
The output node of the volume base layer may be represented as:
wherein:
respectively representing the characteristic diagrams of the current and previous layers,
represents the convolution kernel from the m-th feature map of the previous layer to the n-th feature map of the current layer,
for neuron biasing, f (-) is the signal activation function.
The sub-sampling layer node output may be expressed as:
where s x s is the sub-sampling template scale,
is the template weight.
The result of the convolutional neural network CNN fully connected output layer can be expressed as:
(2) the parallel architecture principle of the deep convolutional neural network is as follows:
the model training method based on deep learning generally needs to consume huge time overhead, and in order to improve the training speed, a multi-node parallel computing architecture based on hadoop distributed file storage is adopted, as shown in fig. 4. A large number of training samples are stored on different storage nodes, the samples are independent from each other, and each storage node is provided with a complete convolutional neural network, so that the neural network on each node is only responsible for training a part of sample sets, and the nodes are updated once after being processed. In the training process of the samples, the samples are processed in a unidirectional layer-by-layer mode among all layers of the network, the feature maps, the neurons and the neuron weights of the same layer are independent, and therefore a plurality of samples can be calculated in a convolutional neural network in a parallel mode.
Claims (3)
1. An optical remote sensing image offshore platform automatic detection method is characterized in that: the method specifically comprises the following steps:
step 1, analyzing the multi-resolution and multi-viewpoint remote sensing imaging characteristics of an offshore oil and gas platform, and fitting a remote sensing image through a characteristic subspace and a principal component;
step 2, improving the signal-to-noise ratio and the signal-to-clutter ratio of offshore platform detection through a space-time Markov random field model, and outputting a target saliency image;
step 3, constructing a large-scale target standard data set, performing multi-node parallel architecture training based on a hadoop distributed file storage system on the deep convolutional neural network, continuously optimizing the deep convolutional neural network, extracting the stability characteristics of a large-scale target, and realizing quick and accurate automatic detection of the offshore oil and gas platform in a mass remote sensing image;
the step 1 specifically comprises the following steps:
step 1.1, selecting a multi-temporal optical remote sensing image set under different resolutions and multiple viewpoints;
step 1.2, establishing a priori knowledge database of a large-range ocean detection identification area;
step 1.3, processing the input image set in the step 1 through the priori knowledge database established in the step 2, and removing non-marine region information in remote sensing data and information of a marine deeper region which cannot be developed by marine oil and gas to obtain a processed image set;
step 1.4, manually calibrating the image set obtained in the step 3 to obtain a marked remote sensing image set;
step 1.5, carrying out simulation analysis on the characteristic subspace and the main component of the marked remote sensing image set to obtain a multi-temporal remote sensing image set;
the step 2 specifically comprises the following steps:
step 2.1, carrying out time domain low-pass filtering on the images in the image set, and initializing a Markov random field;
step 2.2, establishing a three-dimensional Markov random field model according to the input multi-temporal remote sensing image, and considering that the gray distribution information of the remote sensing image conforms to a finite Gaussian mixture model to obtain joint condition probability;
step 2.3, carrying out maximum likelihood estimation according to the conditional probability in the step 2.2, and calculating Markov random field model parameters;
step 2.4, updating the Markov random field by using the calculated parameters, and acquiring the Markov random field model parameters again through maximum likelihood estimation;
step 2.5, comparing whether the parameter errors of the Markov random field model calculated twice before and after meet the convergence; if yes, executing step 2.6, otherwise, continuing to execute step 2.4;
and 2.6, calculating and outputting a target saliency image through a three-dimensional space-time Markov random field model.
2. The automatic detection method for the optical remote sensing image offshore platform according to claim 1, characterized in that: the step 3 specifically comprises the following steps:
step 3.1, a platform sample base is obtained and then stored in a hadoop distributed file storage system, sample base subsets are stored in k nodes, each node trains the sample base independently, and each parameter of the convolutional neural network is initialized;
step 3.2, simultaneously carrying out complete network training on the sample library subsets by each node to obtain a parameter vector of the convolutional neural network, and fitting the parameters of the whole neural network by using the neural network parameters of each node;
3.3, updating the neural network parameters of each node by using the overall neural network parameters;
step 3.4, determining the training effect through manual interaction, continuously adjusting parameters, and repeatedly executing the step 3.2 to the step 3.3 until the training meets the requirements;
step 3.5, adding a target sample, and repeating the step 3.2 to the step 3.4;
and 3.6, detecting the target saliency image output in the step 2.6 by using the convolutional neural network trained in the step 3.5, and outputting a target detection result.
3. The automatic detection method for the optical remote sensing image offshore platform according to claim 1, characterized in that: in step 1.2, the database contains the latitude and longitude of the offshore, the sea depth.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611161835.0A CN106845343B (en) | 2016-12-15 | 2016-12-15 | Automatic detection method for optical remote sensing image offshore platform |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611161835.0A CN106845343B (en) | 2016-12-15 | 2016-12-15 | Automatic detection method for optical remote sensing image offshore platform |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106845343A CN106845343A (en) | 2017-06-13 |
| CN106845343B true CN106845343B (en) | 2020-05-19 |
Family
ID=59139361
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201611161835.0A Active CN106845343B (en) | 2016-12-15 | 2016-12-15 | Automatic detection method for optical remote sensing image offshore platform |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106845343B (en) |
Families Citing this family (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108009469B (en) * | 2017-10-24 | 2020-11-24 | 中国科学院电子学研究所苏州研究院 | Offshore oil and gas platform detection method based on structure recurrent neural network |
| CN108335222A (en) * | 2017-12-21 | 2018-07-27 | 中国石油天然气股份有限公司 | Marine site oil-gas exploration on-site investigation method and device based on multispectral remote sensing |
| CN111028255B (en) * | 2018-10-10 | 2023-07-21 | 千寻位置网络有限公司 | Farmland area pre-screening method and device based on priori information and deep learning |
| CN110823837A (en) * | 2019-11-19 | 2020-02-21 | 中国科学院遥感与数字地球研究所 | Method and device for simulating water leaving radiant quantity of ocean water body |
| CN112070062A (en) * | 2020-09-23 | 2020-12-11 | 南京工业职业技术大学 | Hadoop-based crop waterlogging image classification detection and implementation method |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102768356A (en) * | 2012-08-03 | 2012-11-07 | 南京大学 | Method for detecting ocean stationary targets by multi-temporal radar data |
| CN103544528A (en) * | 2013-11-15 | 2014-01-29 | 南京大学 | BP neural-network classification method based on Hadoop |
| US8787638B2 (en) * | 2011-04-07 | 2014-07-22 | The Chinese University Of Hong Kong | Method and device for retinal image analysis |
-
2016
- 2016-12-15 CN CN201611161835.0A patent/CN106845343B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8787638B2 (en) * | 2011-04-07 | 2014-07-22 | The Chinese University Of Hong Kong | Method and device for retinal image analysis |
| CN102768356A (en) * | 2012-08-03 | 2012-11-07 | 南京大学 | Method for detecting ocean stationary targets by multi-temporal radar data |
| CN103544528A (en) * | 2013-11-15 | 2014-01-29 | 南京大学 | BP neural-network classification method based on Hadoop |
Non-Patent Citations (1)
| Title |
|---|
| Exploiting Global Priors for RGB-D Saliency Detection;Jiangqiang Ren等;《2015 IEEE Conference on Computer Vision and Pattern Recognition Workshops》;20151026;第1-8页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106845343A (en) | 2017-06-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Li et al. | A deep learning method for change detection in synthetic aperture radar images | |
| CN106845343B (en) | Automatic detection method for optical remote sensing image offshore platform | |
| CN110472627B (en) | End-to-end SAR image recognition method, device and storage medium | |
| Wang et al. | New hierarchical saliency filtering for fast ship detection in high-resolution SAR images | |
| CN108460382B (en) | Optical remote sensing image ship detection method based on deep learning single-step detector | |
| Dong et al. | Object detection in high resolution remote sensing imagery based on convolutional neural networks with suitable object scale features | |
| Wu et al. | Inshore ship detection based on convolutional neural network in optical satellite images | |
| Ma et al. | Light-YOLOv4: An edge-device oriented target detection method for remote sensing images | |
| Kremezi et al. | Pansharpening PRISMA data for marine plastic litter detection using plastic indexes | |
| Chen et al. | Adaptive effective receptive field convolution for semantic segmentation of VHR remote sensing images | |
| CN107025654B (en) | SAR image self-adaptive ship detection method based on global iterative inspection | |
| Ai et al. | Convolutional neural network to retrieve water depth in marine shallow water area from remote sensing images | |
| CN108428220A (en) | Satellite sequence remote sensing image sea island reef region automatic geometric correction method | |
| Wang et al. | Study on the combined application of CFAR and deep learning in ship detection | |
| CN115147731A (en) | SAR image target detection method based on full-space coding attention module | |
| Hou et al. | End-to-end automatic ship detection and recognition in high-resolution Gaofen-3 spaceborne SAR images | |
| Akshay et al. | Satellite image classification for detecting unused landscape using CNN | |
| CN112906564B (en) | Intelligent decision support system design and implementation method for automatic target recognition of unmanned airborne SAR (synthetic aperture radar) image | |
| Hu et al. | A sample enhancement method based on simple linear iterative clustering superpixel segmentation applied to multibeam seabed classification | |
| CN117056807A (en) | Integrated cloud detection method based on semantic segmentation network and long-term and short-term memory network, readable storage medium and computing device | |
| Xie et al. | Machine-Learning-Method-Based Inversion of Shallow Bathymetric Maps Using ICESat-2 ATL03 Data | |
| Guerrero-Font et al. | Sparse Gaussian process for online seagrass semantic mapping | |
| Xi et al. | Multitarget Detection Algorithms for Multitemporal Remote Sensing Data | |
| Ruiwen et al. | Segmentation of remote sensing images based on U-net multi-task learning | |
| Yu et al. | A lightweight ship detection method in optical remote sensing image under cloud interference |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |