CN109490861A - A kind of blast-furnace line extracting method - Google Patents
A kind of blast-furnace line extracting method Download PDFInfo
- Publication number
- CN109490861A CN109490861A CN201811268646.2A CN201811268646A CN109490861A CN 109490861 A CN109490861 A CN 109490861A CN 201811268646 A CN201811268646 A CN 201811268646A CN 109490861 A CN109490861 A CN 109490861A
- Authority
- CN
- China
- Prior art keywords
- stockline
- blast
- radar
- signal sequence
- furnace
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S7/00—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00
- G01S7/02—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00
- G01S7/41—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section
- G01S7/417—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section involving the use of neural networks
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S13/00—Systems using the reflection or reradiation of radio waves, e.g. radar systems; Analogous systems using reflection or reradiation of waves whose nature or wavelength is irrelevant or unspecified
- G01S13/88—Radar or analogous systems specially adapted for specific applications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Remote Sensing (AREA)
- Radar, Positioning & Navigation (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Data Mining & Analysis (AREA)
- Software Systems (AREA)
- Biomedical Technology (AREA)
- Computer Networks & Wireless Communication (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Electromagnetism (AREA)
- Radar Systems Or Details Thereof (AREA)
Abstract
The present invention provides a kind of blast-furnace line extracting method, is able to ascend accuracy, generalization and the real-time of blast-furnace line extraction.The described method includes: acquisition blast furnace scene radar echo signal, wherein the radar wave energy that radar system issues covers blast furnace radial direction charge level each region;Generate the signal sequence spectrogram of the radar echo signal;Data enhancing is carried out to the signal sequence spectrogram and obtains stockline annotation results, according to the enhanced signal sequence spectrogram of data and stockline annotation results, establishes stockline segmentation database;The full convolutional network model of modularization is constructed, according to the full convolutional network model of modularization described in the stockline of foundation segmentation database training;Blast furnace radar echo signal to be detected is acquired in real time and generates signal sequence spectrogram, according to the full convolutional network model of trained modularization, stockline segmentation is carried out to the signal sequence spectrogram and is extracted.The present invention relates to blast-furnace lines to monitor field.
Description
Technical field
The present invention relates to blast-furnace lines to monitor field, particularly relates to a kind of blast-furnace line extracting method.
Background technique
In iron industry, blast furnace charge level is measured and optimal control to adjust cloth is to improve production efficiency and energy conservation
The key of emission reduction obtains the adjustment that accurate true charge level information is conducive to blast furnace material distribution strategy, guarantees in furnace in material and furnace
The reasonable layout of Gas Flow, therefore real-time and accurately obtain shape of charge level information in blast furnace and have great importance.Currently, high frequency
Microwave radar is widely used in blast furnace charge level detection as measurement sensor ideal under blast furnace complex environment.It is logical
It crosses the collected radar echo signal of quick Fourier transform pairs and carries out time-frequency conversion acquisition signal spectrum, finally by frequency spectrum
Linear relationship between the corresponding peak value spectral line number of charge level echo and radar surveying distance calculates the distance value of stockline to be measured.
But collection in worksite to Energy distribution of the echo-signal on frequency spectrum mainly by distributor chute, cross temperature measurer,
True material echo and strong background noise composition, belong to non-linear, non-stationary signal, true charge level echo is usually blanked it
In, it is difficult to it identifies.The basic model of Radar Signal Processing algorithm is to obtain signal frequency according to discrete Fourier transform at present
Rough estimate recycles discrete spectrum correcting algorithm finely to be estimated.The Model Establishment returns radar on the basis of peak-seeking
Wave signal has very high requirement, and the mode is to carry out independent process to each group radar signal, and it is hidden to lack deeper excavation
It is contained in the regular information of the true charge level consecutive variations between each group radar signal, there are serious information leakages.It faces
The problems such as blast furnace adverse circumstances, existing algorithm is easy to appear missing stockline and part jump.
Summary of the invention
The technical problem to be solved in the present invention is to provide a kind of blast-furnace line extracting methods, to solve present in the prior art
Detection stockline when easily there is the problem of stockline missing and part jump.
In order to solve the above technical problems, the embodiment of the present invention provides a kind of blast-furnace line extracting method, comprising:
Acquire blast furnace scene radar echo signal, wherein the radar wave energy that radar system issues covers blast furnace radial direction charge level
Each region;
Generate the signal sequence spectrogram of the radar echo signal;
Data enhancing is carried out to the signal sequence spectrogram and obtains stockline annotation results, according to the enhanced letter of data
Number timing spectrogram and stockline annotation results establish stockline segmentation database;
The full convolutional network model of modularization is constructed, according to the full convolution of modularization described in the stockline of foundation segmentation database training
Network model;
Blast furnace radar echo signal to be detected is acquired in real time and generates signal sequence spectrogram, according to trained module
Change full convolutional network model, stockline segmentation is carried out to the signal sequence spectrogram and is extracted.
Further, acquisition blast furnace scene radar echo signal includes:
The installation principle that distributing mode and radar system are rotated without clock-type chute used according to blast furnace, in blast furnace roof cloth
If corresponding radar system, meet blast furnace radial direction charge level all standing;
The continuous acquisition radar echo signal in real time in blast furnace production process.
Further, the signal sequence spectrogram for generating the radar echo signal includes:
To collected radar echo signal progress time-frequency transformation, normalization, timing is unfolded and grey scale mapping, described in generation
The signal sequence spectrogram of radar echo signal.
Further, described that time-frequency transformation, normalization, timing expansion and gray scale are carried out to collected radar echo signal
Mapping, the signal sequence spectrogram for generating the radar echo signal include:
Time-frequency convert is carried out to radar echo signal by Fast Fourier Transform (FFT), obtains the amplitude frequency of radar echo signal
Spectrum;
According to the distance between furnace roof radar and charge level, section is carried out to the amplitude frequency spectrum of each group radar echo signal and is cut
It takes, and amplitude is normalized;
According to continuity of the radar echo signal in time series, with the signal after normalized described in continuous N group
Frequency spectrum data is unit, and timing is arranged to make up two-dimensional matrix;
Grey scale mapping is carried out to the two-dimensional matrix, generates signal sequence frequency spectrum grayscale image.
Further, after the generation signal sequence frequency spectrum grayscale image, the method also includes:
Using a cube height for interpolation stretch signal timing frequency spectrum grayscale image.
Further, described to include: to signal sequence spectrogram progress data enhancing
Data enhancing is carried out by the signal sequence spectrogram of the flip horizontal to generation.
Further, the acquisition stockline annotation results include:
Obtain the annotation results pixel-by-pixel of stockline in the enhanced signal sequence spectrogram of data, wherein marked content packet
It includes: the variation tendency of stockline and the floating on stockline, double-deck stockline, continuous tomography;
By preset pixel threshold, it will will mark the pixel value for being more than or equal to preset pixel threshold in picture and be converted into
First identifier, the pixel value less than preset pixel threshold are converted into second identifier, generate a pair of with signal sequence spectrogram one
The first identifier answered-second identifier label picture.
Further, the full convolutional network model of building modularization includes:
Building includes the full convolutional network mould of modularization of characteristic extracting module, Fusion Features module and feature decoder module
Type;Wherein,
The characteristic extracting module is intersected with pond layer by several convolution units and is constituted, for passing through different levels
Convolution unit learns the stockline feature in input signal timing spectrogram, obtains different levels, various sizes of feature
Figure;
The Fusion Features module, different levels, various sizes of characteristic pattern for being extracted to characteristic extracting module
It is merged;
The feature decoder module, for carrying out the net that up-sampling operation obtains size identical as label picture to characteristic pattern
Network output, and changing output channel number not be consistent it Shuo with as several classes of, obtain stockline segmentation result.
Further, the stockline segmentation database includes: training set, verifying collection and test set;
It is described according to the stockline of foundation divide database training described in the full convolutional network model of modularization include:
Utilize the training set training full convolutional network model of modularization;
Optimize, adjust the hyper parameter of the full convolutional network model of the modularization using the verifying collection;
The performance of the full convolutional network model of the modularization is tested, evaluated using the test set.
Further, the radar system includes: distributive array radar, scanning radar and/or phased-array radar.
The advantageous effects of the above technical solutions of the present invention are as follows:
In above scheme, blast furnace scene radar echo signal is acquired, wherein the radar wave energy that radar system issues covers high
Furnace radial direction charge level each region;Generate the signal sequence spectrogram of the radar echo signal, wherein different radars are when different
Between signal sequence spectrogram in section be different;Data enhancing is carried out to the signal sequence spectrogram and obtains stockline mark
As a result, establishing stockline segmentation database according to the enhanced signal sequence spectrogram of data and stockline annotation results;Construct module
Change full convolutional network model, according to the full convolutional network model of modularization described in the stockline of foundation segmentation database training;It adopts in real time
Collect blast furnace radar echo signal to be detected and generate signal sequence spectrogram, according to the full convolutional network mould of trained modularization
Type carries out stockline segmentation to the signal sequence spectrogram and extracts, realizes the reality in production process to blast-furnace line variation tendency
When detect.In this way, dividing database by stockline completely embodies in blast furnace production process radial each region stockline in continuous time
Changing rule in range, and by more radar signal timing spectrograms training same model, the general of model can be effectively improved
Change performance, so that promoting blast-furnace line extracts accuracy, generalization and real-time.
Detailed description of the invention
Fig. 1 is the flow diagram of blast-furnace line extracting method provided in an embodiment of the present invention;
Fig. 2 is the detailed process schematic diagram of blast-furnace line extracting method provided in an embodiment of the present invention;
Fig. 3 is 8 distributive arrays radar scheme of installation provided in an embodiment of the present invention;
Fig. 4 is collected blast furnace radar signal time-domain diagram provided in an embodiment of the present invention;
Fig. 5 is the signal sequence spectrogram schematic diagram of generation provided in an embodiment of the present invention;
Fig. 6 is stockline local detail schematic diagram provided in an embodiment of the present invention;
Fig. 7 is stockline annotation results schematic diagram provided in an embodiment of the present invention;
Fig. 8 is the full convolutional network model structure schematic diagram of modularization provided in an embodiment of the present invention;
Fig. 9 is stockline segmentation result schematic diagram provided in an embodiment of the present invention;
Figure 10 is that stockline provided in an embodiment of the present invention detects contrast schematic diagram.
Specific embodiment
To keep the technical problem to be solved in the present invention, technical solution and advantage clearer, below in conjunction with attached drawing and tool
Body embodiment is described in detail.
Aiming at the problem that missing stockline and part jump easily occurs when existing detection stockline in the present invention, a kind of blast furnace is provided
Stockline extracting method.
As shown in Figure 1, blast-furnace line extracting method provided in an embodiment of the present invention, comprising:
S101 acquires blast furnace scene radar echo signal, wherein the radar wave energy covering blast furnace that radar system issues is radial
Charge level each region;
S102 generates the signal sequence spectrogram of the radar echo signal;
S103 carries out data enhancing to the signal sequence spectrogram and obtains stockline annotation results, enhanced according to data
Signal sequence spectrogram and stockline annotation results afterwards establish stockline segmentation database;
S104 constructs the full convolutional network model of modularization, according to modularization described in the stockline of foundation segmentation database training
Full convolutional network model;
S105 acquires blast furnace radar echo signal to be detected in real time and generates signal sequence spectrogram, according to training
The full convolutional network model of modularization, to the signal sequence spectrogram carry out stockline segmentation extract.
Blast-furnace line extracting method described in the embodiment of the present invention acquires blast furnace scene radar echo signal, wherein radar
The radar wave energy that system issues covers blast furnace radial direction charge level each region;Generate the signal sequence frequency spectrum of the radar echo signal
Figure, wherein the signal sequence spectrogram of different radars in different time period is different;To the signal sequence spectrogram into
Row data enhance and obtain stockline annotation results, according to the enhanced signal sequence spectrogram of data and stockline annotation results, build
Vertical stockline divides database;The full convolutional network model of modularization is constructed, according to mould described in the stockline of foundation segmentation database training
The full convolutional network model of blockization;Blast furnace radar echo signal to be detected is acquired in real time and generates signal sequence spectrogram, according to
The trained full convolutional network model of modularization carries out stockline segmentation to the signal sequence spectrogram and extracts, and realization produced
To the real-time detection of blast-furnace line variation tendency in journey.In this way, dividing database by stockline completely embodies blast furnace production process
Changing rule of the middle each region stockline of radial direction within continuous time, and it is same by the training of more radar signal timing spectrograms
Model, can effectively improve the Generalization Capability of model, to promote accuracy, generalization and the real-time of blast-furnace line extraction.
In the specific embodiment of aforementioned blast-furnace line extracting method, further, acquisition blast furnace scene radar
Echo-signal includes:
The installation principle that distributing mode and radar system are rotated without clock-type chute used according to blast furnace, in blast furnace roof cloth
If corresponding radar system, meet blast furnace radial direction charge level all standing;
The continuous acquisition radar echo signal in real time in blast furnace production process.
It is further, described to generate the radar return in the specific embodiment of aforementioned blast-furnace line extracting method
The signal sequence spectrogram of signal includes:
To collected radar echo signal progress time-frequency transformation, normalization, timing is unfolded and grey scale mapping, described in generation
The signal sequence spectrogram of radar echo signal.
It is further, described that collected radar is returned in the specific embodiment of aforementioned blast-furnace line extracting method
Wave signal carries out time-frequency transformation, normalization, timing expansion and grey scale mapping, generates the signal sequence frequency of the radar echo signal
Spectrogram includes:
Time-frequency convert is carried out to radar echo signal by Fast Fourier Transform (FFT), obtains the amplitude frequency of radar echo signal
Spectrum;
According to the distance between furnace roof radar and charge level, section is carried out to the amplitude frequency spectrum of each group radar echo signal and is cut
It takes, and amplitude is normalized, it is assumed that interception range is N;
According to continuity of the radar echo signal in time series, with the signal after normalized described in continuous N group
Frequency spectrum data is unit, and timing is arranged to make up two-dimensional matrix, then the size of two-dimensional matrix is N*M;
Grey scale mapping is carried out to the two-dimensional matrix, generates signal sequence frequency spectrum grayscale image.
It is described in finer embodiment further in the specific embodiment of aforementioned blast-furnace line extracting method
The variation details of stockline and convenient for fine mark in signal sequence spectrogram, the generation signal sequence frequency spectrum grayscale image it
Afterwards, the method also includes:
Using a cube height for interpolation stretch signal timing frequency spectrum grayscale image.
In the present embodiment, according to radar range finding principle, stockline testing distance is directly proportional to spectrum peak spectral line number, is converted into
After image, stockline distance in image is directly proportional to its height in the picture;It is therefore possible to use cube interpolation stretch signal
The height of timing frequency spectrum grayscale image, for example, picture altitude can be stretched one times, width is constant, in this way, can guarantee stockline
Clearer display stockline changes details under the premise of constant, while facilitating mark.
It is further, described to the signal sequence frequency in the specific embodiment of aforementioned blast-furnace line extracting method
Spectrogram carries out data enhancing
Data enhancing is carried out by the signal sequence spectrogram of the flip horizontal to generation.
In the specific embodiment of aforementioned blast-furnace line extracting method, further, the acquisition stockline annotation results
Include:
Obtain the annotation results pixel-by-pixel of stockline in the enhanced signal sequence spectrogram of data, wherein marked content packet
It includes: the variation tendency of stockline and the floating on stockline, double-deck stockline, continuous tomography;
By preset pixel threshold, it will will mark the pixel value for being more than or equal to preset pixel threshold in picture and be converted into
First identifier, the pixel value less than preset pixel threshold are converted into second identifier, generate a pair of with signal sequence spectrogram one
The first identifier answered-second identifier label picture.
In the present embodiment, by preset pixel threshold, it can will be more than or equal to preset pixel threshold in mark picture
The pixel value of (that is: stockline is corresponding) is converted into first identifier (for example, 0), is less than the picture of preset pixel threshold (background is corresponding)
Plain value is converted into second identifier (for example, 1), generation and the one-to-one 0-1 label picture of signal sequence spectrogram, so as in net
In network training, result is exported by label picture and network and calculates network losses function, updates network parameter.
In the specific embodiment of aforementioned blast-furnace line extracting method, further, the full convolution of the building modularization
Network model includes:
The characteristic extracting module is intersected with pond layer by several convolution units and is constituted, for passing through different levels
Convolution unit learns the stockline feature in input signal timing spectrogram, obtains different levels, various sizes of feature
Figure;Where it is assumed that convolution unit is 5, the convolution unit of shallow-layer is (for example, the first convolution unit, the second convolution unit, third
Convolution unit, Volume Four product unit and) study edge, the local features such as texture and shape, the convolution unit of deep layer is (for example, the
Five convolution units) shallow-layer feature is blended on the basis of this again, learn more abstract, more essential feature;
The Fusion Features module, different levels, various sizes of characteristic pattern for being extracted to characteristic extracting module
It is merged, learns different shallow-layer features and further feature in pair of spatial position and interchannel will pass through the convolutional layer of addition
It should be related to, fusion of the characteristic pattern in decoding process through shallow-layer feature is made to make up deep layer abstract characteristics lacking on target detail
It loses;The difference of shallow-layer feature and further feature on spatial position and channel makes the difference of characteristic pattern amalgamation mode to network model
Performance be affected;
The feature decoder module, for carrying out the net that up-sampling operation obtains size identical as label picture to characteristic pattern
Network output, and changing output channel number not be consistent it Shuo with as several classes of, obtain stockline segmentation result.
In the present embodiment, the full convolutional network model structure of the modularization of building includes: characteristic extracting module, Fusion Features mould
Block and feature decoder module;Wherein,
Characteristic extracting module: basic model of the convolutional neural networks as characteristic extracting module is selected;Remove convolutional Neural
Full articulamentum and output layer in network;Convolution unit there are five containing in convolutional neural networks, each convolution unit is by several volumes
Batch normalization unit of lamination, 1 amendment linear unit and 1 collectively constitutes, and specific: the convolution unit of front two is by two
Convolutional layer is composed in series, behind three convolution units be composed in series by three convolutional layers, modified line is set after each convolutional layer
Property unit and batch normalization unit;Determine convolution kernel size in convolutional neural networks;Five convolution lists in convolutional neural networks
Member replaces connection with pond layer, and using maximum value pond, pond window size is 2*2, sliding step 2*2;
In the present embodiment, since linear model ability to express is insufficient, the use of linear unit is corrected, is introduced for network non-thread
Property characteristic, allow the full convolutional network model of modularization indicate nonlinear complexity between input and output arbitrary function mapping,
Have that calculating speed is fast, derivation is simple, unilateral inhibitions, sparse activity, broader excitement boundary and not will cause gradient disperse
The advantages that;Batch normalization unit can accelerate network training convergence speed by being adjusted to the variation that inter-layer data is distributed
Degree prevents over-fitting from improving the generalization ability of network.
Fusion Features module: the amalgamation mode between setting characteristic pattern is splicing;Before characteristic pattern fusion, pass through convolution kernel
Convolutional layer having a size of 1*1 changes characteristic pattern port number;Convolution after characteristic pattern fusion, by convolution kernel having a size of 1*1
Layer changes characteristic pattern port number;
Feature decoder module: after the 5th pond layer pool5, three convolutional layers is successively added, first convolution is set
The convolution kernel of layer and the output characteristic pattern comparable size of pool5, the convolution kernel of subsequent two convolutional layers are successively dropped having a size of 1*1
The port number of low characteristic pattern makes final characteristic pattern port number classification number, obtains characteristic pattern score5;Pass through deconvolution pair
Score5 carries out two times of up-samplings, obtains pool5 layers of feature decoding result score_up5;By the spy of the 4th pond layer pool4
Result after sign figure progress 1*1 convolution algorithm is merged with score_up5, passes through the feature after the change fusion of 1*1 convolution algorithm
Figure port number and pool4, which are consistent and pass through warp lamination, carries out two times of up-samplings, obtains pool4 layers of feature decoding result
score_up4;Result after the characteristic pattern of third pond layer pool3 to be carried out to 1*1 convolution algorithm is merged with score_up4,
Characteristic pattern port number and pool3 after being merged by the convolutional layer change that convolution kernel is 1*1 are consistent and pass through deconvolution
Layer carries out octuple up-sampling, obtains final stockline segmentation result, port number is consistent with classification number.
In the present embodiment, also set up the full convolutional network model loss function of modularization be cross entropy loss function, so as to
Parameter update is carried out during the full convolutional network model training of modularization, and adds L2 parameter regular terms;
The working principle of the full convolutional network model of S42, training moduleization includes: that characteristic extracting module parameter initialization is chosen
The convolutional neural networks model parameter of pre-training on ImageNet, the initialization of remaining module parameter of model are all made of truncation normal state
Distribution;Using batch gradient descent method Optimized model parameter;Initializing learning rate is 10-4, more new strategy is every iteration 5000 times
Reduce 10 times;During the network optimization, its hyper parameter is adjusted according to performance indicator of the network model on verifying collection,
Wherein, the hyper parameter includes: the quantity (batch_size) in each batch comprising sample, learning rate (learning_
Rate), network the number of iterations etc..
In the specific embodiment of aforementioned blast-furnace line extracting method, further, the stockline divides database packet
It includes: training set, verifying collection and test set;
It is described according to the stockline of foundation divide database training described in the full convolutional network model of modularization include:
Utilize the training set training full convolutional network model of modularization;
Optimize, adjust the hyper parameter of the full convolutional network model of the modularization using the verifying collection;
The performance of the full convolutional network model of the modularization is tested, evaluated using the test set.
The present embodiment is red, and the training set refers to the data set for the full convolutional network model of training moduleization, verifying collection
Refer to that the data set for being used to optimize, adjust the full convolutional network model hyper parameter of modularization in the training process, test set refer to use
Carry out the full convolutional network model performance of test moduleization and carries out the data set of performance evaluation.
In the specific embodiment of aforementioned blast-furnace line extracting method, further, the radar system includes: distribution
Formula array radar, scanning radar and/or phased-array radar.
Blast-furnace line extracting method described in the present embodiment is suitable for blast furnace production process to the company of blast-furnace line variation
Continuous prolonged accurate monitoring, helps to optimize and revise cloth strategy, improves production efficiency and realize energy-saving and emission-reduction;Applicable thunder
It is not limited to distributive array radar, scanning radar and/or phased-array radar up to system, it is all to utilize radar within continuous time
Stockline tracking occasion be applicable in.
Blast-furnace line extracting method described in the present embodiment is not limited to blast-furnace line and extracts task, is being applied to target range
There is versatility under the various industrial radar scenes of detection, essential characteristics are the full convolutional network model of binding modulesization and letter
Classical signal processing is completed in the form of image segmentation, and meets accuracy, generalization and real-time by number timing spectrogram.
To sum up, the advantageous effects of the above technical solutions of the present invention are as follows:
1) it establishes the stockline that one is extracted for blast-furnace line and divides database, which divides database by blast furnace scene
Radar signal timing spectrogram and its annotation results are constituted, and can completely be embodied radial each region stockline in blast furnace production process and be existed
Changing rule within the scope of continuous time, and the signal sequence spectrogram of different radars in different time period is different, it should
The blast-furnace line that is established as of stockline segmentation database provides data support;
2) target accurately extracted for blast-furnace line, building include characteristic extracting module, Fusion Features module and feature
The full convolutional network model of the modularization of decoder module, using the full convolutional network model parameter of training set training moduleization, using testing
Card collection optimizes and revises the parameter of the full convolutional network model of modularization, and more radar signal timing spectrograms are used to train same model,
The Generalization Capability of model, the missing and local jump of the stockline that the existing algorithm of solution is easy to appear under blast furnace environment can be effectively improved
The problems such as change.
For a better understanding of the present invention, 8 distributive array radar systems are based on, described in the embodiment of the present invention
Blast-furnace line extracting method is illustrated:
The embodiment of the present invention is made of two parts:
First part is the foundation that stockline divides database, including data acquisition, image generate, data enhance and data mark
Note provides data for subsequent algorithm design and supports;
Second part is the building full convolutional network model of modularization, passes through the full convolutional network model of training set training moduleization
Parameter optimizes and revises the hyper parameter of the full convolutional network model of modularization by verifying collection, and test moduleization is complete on test set
The performance of convolutional network model.
The embodiment of the present invention is under Ubuntu16.04 operating system, in the GPU hardware based on Nvidia GTX 1080Ti
TensorFlow deep learning frame is built on platform, completes the training and test of the full convolutional network model of modularization.
Blast-furnace line extracting method described in the present embodiment, as shown in Fig. 2, specific steps may include:
S1, the live 8 distributive array radar system radar echo signals of acquisition blast furnace;
S2 carries out time-frequency transformation, normalization, timing expansion and grey scale mapping to collected radar echo signal, generates
Signal sequence spectrogram;
S3 carries out data enhancing to the signal sequence spectrogram of generation and obtains stockline annotation results, enhanced according to data
Signal sequence spectrogram and stockline annotation results afterwards establish stockline segmentation database;
S4, building include the full convolutional network of modularization of characteristic extracting module, Fusion Features module and feature decoder module
Model, and the full convolutional network model of database training modularization is divided according to the stockline;
S5 acquires blast furnace radar echo signal to be detected and generates signal sequence spectrogram, in real time according to trained
The full convolutional network model of modularization carries out stockline segmentation to the signal sequence spectrogram and extracts, realizes the change to blast-furnace line
Change trend is measured in real time.
In the present embodiment, the live 8 distributive array radar system radar echo signals of the acquisition blast furnace include:
The installation principle that distributing mode and distributed radar system are rotated without clock-type chute used according to blast furnace, in blast furnace
Furnace roof lays 8 distributive array radar systems, meets radar charge level radial direction ring type all standing, scheme of installation such as Fig. 3 institute
Show, in Fig. 3 1. -8. indicate 8 distributive array radar systems in eight radars;Continuous four in blast furnace production process
The echo-signal of its eight radar of acquisition, respectively corresponds the different zones of blast furnace charge level radially, as shown in Figure 4;
It is described that time-frequency transformation, normalization, timing expansion and ash are carried out to collected radar echo signal in the present embodiment
Degree mapping, generating signal sequence spectrogram includes:
It include as shown in Figure 4 1024 data points in single radar echo signal, by Fast Fourier Transform (FFT) to radar
Echo-signal carries out time-frequency convert, obtains the amplitude frequency spectrum of radar echo signal, availability vector is expressed as [x1,x2,…x512]T;
According to the limited distance between furnace roof radar and charge level, section interception is carried out to the amplitude frequency spectrum of each group radar echo signal, is cut
The vector after range is taken to be expressed as [x1,x2,…x128]T, and amplitude is normalized;Existed according to radar echo signal
Continuity and time interval in time series, as unit of the signal spectrum data after normalized described in continuous 256 groups,
Timing is arranged to make up two-dimensional matrix, having a size of 128*256;Grey scale mapping is carried out to the two-dimensional matrix, generates signal sequence
Frequency spectrum grayscale image;It is marked for the variation details of stockline in the finer embodiment signal sequence spectrogram and convenient for fine,
A cube interpolation is used to be sized as 256*256, wherein the signal sequence spectrogram of 1#, 6#, 7# radar is as shown in Figure 5;
In the present embodiment, the signal sequence spectrogram of described pair of generation carries out data enhancing and obtains stockline annotation results,
According to the enhanced signal sequence spectrogram of data and stockline annotation results, establishing stockline segmentation database includes:
Data enhancing is carried out by the signal sequence spectrogram of the flip horizontal to generation;It is enhanced to data all
Stockline in signal sequence spectrogram carries out pixel-by-pixel mark by hand, and marked content includes: the overall variation trend and material of stockline
The localized variations details such as small floating, the double-deck stockline, continuous faint tomography on line will be marked by preset pixel threshold
Picture transformation generates 0-1 label, and stockline localized variation details is as shown in Figure 6, wherein (a) is small floating, is (b) the double-deck material
Line is (c) continuous faint tomography, and 6# radar stockline mark figure is as shown in Figure 7;1#, 6#, 7# radar return are selected in the present embodiment
Signal establishes stockline segmentation database, altogether includes 5692 width signal sequence spectrograms, upsets that choose after sequence wherein 75% be instruction
White silk collection, 10% are that verifying integrates, 15% is test set, and three data, which are concentrated, does not include identical picture;
In the present embodiment, the structure of the full convolutional network model of the modularization constructed in S4 includes: characteristic extracting module, feature
Fusion Module and feature decoder module;Wherein,
Characteristic extracting module: basic model of the VGG16 as characteristic extracting module is selected;Remove the full connection in VGG16
Layer and output layer;Containing there are five convolution unit in VGG16, each convolution unit by several convolutional layers, 1 amendment linear unit and
1 batch normalization unit collectively constitutes, and specific: the convolution unit of front two is composed in series by two convolutional layers, behind three
A convolution unit is composed in series by three convolutional layers, and setting amendment linear unit and batch normalize single after each convolutional layer
Member;Determine that convolution kernel is having a size of (3,6) in convolutional neural networks;Five pond layers replace with convolution unit in convolutional neural networks
Connection, using maximum value pond, pond window size is 2*2, sliding step 2*2;
Fusion Features module: before characteristic pattern fusion, adding convolution kernel having a size of the convolutional layer of 1*1 makes two characteristic patterns
Port number is consistent;After characteristic pattern fusion, reduce characteristic pattern port number having a size of the convolutional layer of 1*1 by convolution kernel
Half;Amalgamation mode between characteristic pattern is set for splicing, if two characteristic patterns to be fused are respectively f1=f (x1,x2,…,xn)
And f2=f (y1,y2,…,ym), wherein xi, yiIndicate the single channel in characteristic pattern, n, m are port number, and splicing amalgamation mode can
To indicate are as follows:
fconcat=f (x1,x2,…,xn,y1,y2,…ym)
Feature decoder module: after the 5th pond layer maxpool5, three convolutional layers is successively added, first volume is set
The convolution kernel of the output characteristic pattern comparable size of lamination and maxpool5, the convolution kernels of subsequent two convolutional layers having a size of 1*1, according to
The secondary port number for reducing characteristic pattern, makes final characteristic pattern port number as several classes of other numbers, obtains characteristic pattern score5;Pass through
Deconvolution carries out two times of up-samplings to score5, obtains maxpool5 layers of feature decoding result score_up5;By the 4th pond
Result after the characteristic pattern progress 1*1 convolution algorithm of layer maxpool4 is merged with score_up5, is changed by 1*1 convolution algorithm
Characteristic pattern port number and maxpool4 after fusion, which are consistent and pass through warp lamination, carries out two times of up-samplings, obtains
Maxpool4 layers of feature decoding result score_up4;The characteristic pattern of third pond layer maxpool3 is subjected to 1*1 convolution algorithm
Result afterwards is merged with score_up4, by convolution kernel be 1*1 convolutional layer change fusion after characteristic pattern port number with
Maxpool3, which is consistent and passes through warp lamination, carries out octuple up-sampling, obtains final stockline segmentation result, port number is
2, stockline and background are respectively corresponded, complete network model structure is as shown in Figure 8, wherein a is characterized extraction module, and b is characterized
Fusion Module, c are characterized decoder module;In Fig. 8, conv1, conv2, conv3, conv4, conv5 respectively indicate first volume
Product unit, second convolution unit, third convolution unit, the 4th convolution unit, the 5th convolution unit, each convolution list
Member is collectively constituted by several convolutional layers, amendment linear unit, batch normalization unit;maxpool1,maxpool2,maxpool3,
Maxpool4, maxpool5 respectively indicate the first pond layer, the second pond layer, third pond layer, the 4th pond layer, the 5th pond
Layer, the pond mode of each pond layer are maximum value pond;Fuse is characterized fused layer, herein splices character pair figure;
Deconv is warp lamination, carries out up-sampling operation to characteristic pattern.
In this implementation, the full convolutional network model loss function of setup moduleization is cross entropy loss function, and adds L2 ginseng
Number regular terms, loss function may be expressed as:
Wherein, m is training set sample size, yiThe true distribution of pixel class, yi' be pixel class prediction distribution, ω
For model convolution nuclear parameter.
It include L2 parameter regular terms in loss function loss, loss is when seeking local derviation to design parameter in the present embodiment
Part comprising L2 regular terms, pixel class refer to: two class of stockline and background.
In the present embodiment, at the beginning of the working principle of the full convolutional network model of training moduleization includes: characteristic extracting module parameter
Beginningization is chosen at the VGG16 model parameter (parameter before pool5) of pre-training on ImageNet, at the beginning of remaining module parameter of model
Beginningization is all made of cutting gearbox;Network model parameter, after loss function has been calculated, root are optimized using batch gradient descent method
It is lost according to error and carries out backpropagation, parameter ω is adjusted according to following formula:
Wherein, ωnewAnd ωoldTo update front and back weighting parameter;ε is learning rate, and initialization learning rate is 10-4, update plan
Slightly every iteration 5000 times 10 times of diminutions;It is every to pass through an iteration at network repetitive exercise 15000 times, calculate primary loss letter
Number, and carry out parameter update;In an iterative process if since verifying collection error become larger being gradually reduced, then it is assumed that network
There is over-fitting, terminate training and adjust hyper parameter, until network is optimal;After training, test set is sent into network
Model carries out stockline extraction, and extraction process only carries out a propagated forward, carries out pixel conversion to network output result, makes to extract
Stockline visualization, test set segmentation result is as shown in Figure 9;
It is described to acquire blast furnace radar echo signal to be detected in real time and generate signal sequence spectrogram in the present embodiment,
According to the full convolutional network model of trained modularization, stockline segmentation is carried out to the signal sequence spectrogram and is extracted, realization pair
The variation tendency of blast-furnace line, which is measured in real time, includes:
Radar echo signal is acquired in real time and is generated by 8 distributive array radar systems in blast furnace production process
Signal sequence spectrogram is sent into the network model that training is completed, divides to stockline therein and extract, realizes in production process to height
The real-time detection of furnace charge line variation tendency, it is as shown in Figure 10 with the comparing result of existing algorithm, wherein grey is represented by limited
Impulse response (Finite Impulse Response, FIR) bandpass filter combination Fast Fourier Transform (FFT) (Fast
Fourier Transformation, FFT) and a kind of ChirpZ (ChirpZ is frequency spectrum refinement algorithm) Spectrum Correction algorithm extraction
Stockline, black represents the stockline that method of the invention is extracted.
Above-described embodiment the result shows that the method for the present invention can be accurately in real time to high furnace charge under blast furnace adverse circumstances
Line continuously monitors, and method performance is consistent in different radars, different moments, and Generalization Capability is good.
It should be noted that, in this document, relational terms such as first and second and the like are used merely to a reality
Body or operation are distinguished with another entity or operation, are deposited without necessarily requiring or implying between these entities or operation
In any actual relationship or order or sequence.
The above is a preferred embodiment of the present invention, it is noted that for those skilled in the art
For, without departing from the principles of the present invention, several improvements and modifications can also be made, these improvements and modifications
It should be regarded as protection scope of the present invention.
Claims (10)
1. a kind of blast-furnace line extracting method characterized by comprising
Acquire blast furnace scene radar echo signal, wherein the radar wave energy covering blast furnace radial direction charge level that radar system issues is each
Region;
Generate the signal sequence spectrogram of the radar echo signal;
Data enhancing is carried out to the signal sequence spectrogram and obtains stockline annotation results, when signal enhanced according to data
Sequence spectrogram and stockline annotation results establish stockline segmentation database;
The full convolutional network model of modularization is constructed, according to the full convolutional network of modularization described in the stockline of foundation segmentation database training
Model;
Blast furnace radar echo signal to be detected is acquired in real time and generates signal sequence spectrogram, it is complete according to trained modularization
Convolutional network model carries out stockline segmentation to the signal sequence spectrogram and extracts.
2. blast-furnace line extracting method according to claim 1, which is characterized in that acquisition blast furnace scene radar return
Signal includes:
The installation principle that distributing mode and radar system are rotated without clock-type chute used according to blast furnace lays phase in blast furnace roof
Radar system is answered, blast furnace radial direction charge level all standing is met;
The continuous acquisition radar echo signal in real time in blast furnace production process.
3. blast-furnace line extracting method according to claim 1, which is characterized in that described to generate the radar echo signal
Signal sequence spectrogram include:
Time-frequency transformation, normalization, timing expansion and grey scale mapping are carried out to collected radar echo signal, generate the radar
The signal sequence spectrogram of echo-signal.
4. blast-furnace line extracting method according to claim 3, which is characterized in that described to believe collected radar return
Number carry out time-frequency transformation, normalization, timing expansion and grey scale mapping, generate the signal sequence spectrogram of the radar echo signal
Include:
Time-frequency convert is carried out to radar echo signal by Fast Fourier Transform (FFT), obtains the amplitude frequency spectrum of radar echo signal;
According to the distance between furnace roof radar and charge level, section interception is carried out to the amplitude frequency spectrum of each group radar echo signal, and
Amplitude is normalized;
According to continuity of the radar echo signal in time series, with the signal spectrum after normalized described in continuous N group
Data are unit, and timing is arranged to make up two-dimensional matrix;
Grey scale mapping is carried out to the two-dimensional matrix, generates signal sequence frequency spectrum grayscale image.
5. blast-furnace line extracting method according to claim 4, which is characterized in that in the generation signal sequence frequency spectrum ash
It spends after figure, the method also includes:
Using a cube height for interpolation stretch signal timing frequency spectrum grayscale image.
6. blast-furnace line extracting method according to claim 1, which is characterized in that described to the signal sequence spectrogram
Carrying out data enhancing includes:
Data enhancing is carried out by the signal sequence spectrogram of the flip horizontal to generation.
7. blast-furnace line extracting method according to claim 1, which is characterized in that the acquisition stockline annotation results packet
It includes:
Obtain the annotation results pixel-by-pixel of stockline in the enhanced signal sequence spectrogram of data, wherein marked content includes: material
The variation tendency of line and the floating on stockline, double-deck stockline, continuous tomography;
By preset pixel threshold, it will will mark the pixel value for being more than or equal to preset pixel threshold in picture and be converted into first
Mark, the pixel value less than preset pixel threshold are converted into second identifier, generate one-to-one with signal sequence spectrogram
First identifier-second identifier label picture.
8. blast-furnace line extracting method according to claim 7, which is characterized in that the full convolutional network of the building modularization
Model includes:
Building includes the full convolutional network model of modularization of characteristic extracting module, Fusion Features module and feature decoder module;Its
In,
The characteristic extracting module is intersected with pond layer by several convolution units and is constituted, for the convolution by different levels
Unit learns the stockline feature in input signal timing spectrogram, obtains different levels, various sizes of characteristic pattern;
The Fusion Features module, different levels, various sizes of characteristic pattern for extracting to characteristic extracting module carry out
Fusion;
The feature decoder module, the network for carrying out up-sampling operation acquisition size identical as label picture to characteristic pattern are defeated
Out, and changing output channel number not be consistent it Shuo with as several classes of, obtain stockline segmentation result.
9. blast-furnace line extracting method according to claim 1, which is characterized in that the stockline divides database and includes:
Training set, verifying collection and test set;
It is described according to the stockline of foundation divide database training described in the full convolutional network model of modularization include:
Utilize the training set training full convolutional network model of modularization;
Optimize, adjust the hyper parameter of the full convolutional network model of the modularization using the verifying collection;
The performance of the full convolutional network model of the modularization is tested, evaluated using the test set.
10. blast-furnace line extracting method according to claim 1, which is characterized in that the radar system includes: distribution
Array radar, scanning radar and/or phased-array radar.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811268646.2A CN109490861B (en) | 2018-10-29 | 2018-10-29 | Blast furnace burden line extraction method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811268646.2A CN109490861B (en) | 2018-10-29 | 2018-10-29 | Blast furnace burden line extraction method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109490861A true CN109490861A (en) | 2019-03-19 |
CN109490861B CN109490861B (en) | 2020-06-02 |
Family
ID=65693310
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811268646.2A Active CN109490861B (en) | 2018-10-29 | 2018-10-29 | Blast furnace burden line extraction method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109490861B (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110647891A (en) * | 2019-09-17 | 2020-01-03 | 上海仪电(集团)有限公司中央研究院 | CNN (convolutional neural network) -based automatic extraction method and system for time sequence data characteristics of self-encoder |
CN110819751A (en) * | 2020-01-09 | 2020-02-21 | 江苏金恒信息科技股份有限公司 | Blast furnace burden line obtaining method and device based on radar data processing |
CN111449644A (en) * | 2020-03-19 | 2020-07-28 | 复旦大学 | Bioelectricity signal classification method based on time-frequency transformation and data enhancement technology |
CN111784642A (en) * | 2020-06-10 | 2020-10-16 | 中铁四局集团有限公司 | Image processing method, target recognition model training method and target recognition method |
CN113627283A (en) * | 2021-07-23 | 2021-11-09 | 中冶南方工程技术有限公司 | Material level measuring method based on radar echo signals, terminal equipment and storage medium |
CN114136194A (en) * | 2021-10-12 | 2022-03-04 | 江苏丰尚智能科技有限公司 | Method and device for monitoring volume of material in bin, monitoring equipment and storage medium |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20160341825A1 (en) * | 2014-01-20 | 2016-11-24 | Tmt Tapping-Measuring-Technology Sàrl | Device for determining the topography of the burden surface in a shaft furnace |
CN107993215A (en) * | 2017-11-27 | 2018-05-04 | 象辑知源(武汉)科技有限公司 | A kind of weather radar image processing method and system |
CN108009629A (en) * | 2017-11-20 | 2018-05-08 | 天津大学 | A kind of station symbol dividing method based on full convolution station symbol segmentation network |
CN108062753A (en) * | 2017-12-29 | 2018-05-22 | 重庆理工大学 | The adaptive brain tumor semantic segmentation method in unsupervised domain based on depth confrontation study |
CN108564587A (en) * | 2018-03-07 | 2018-09-21 | 浙江大学 | A kind of a wide range of remote sensing image semantic segmentation method based on full convolutional neural networks |
CN108647568A (en) * | 2018-03-30 | 2018-10-12 | 电子科技大学 | Grassland degeneration extraction method based on full convolutional neural networks |
-
2018
- 2018-10-29 CN CN201811268646.2A patent/CN109490861B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20160341825A1 (en) * | 2014-01-20 | 2016-11-24 | Tmt Tapping-Measuring-Technology Sàrl | Device for determining the topography of the burden surface in a shaft furnace |
CN108009629A (en) * | 2017-11-20 | 2018-05-08 | 天津大学 | A kind of station symbol dividing method based on full convolution station symbol segmentation network |
CN107993215A (en) * | 2017-11-27 | 2018-05-04 | 象辑知源(武汉)科技有限公司 | A kind of weather radar image processing method and system |
CN108062753A (en) * | 2017-12-29 | 2018-05-22 | 重庆理工大学 | The adaptive brain tumor semantic segmentation method in unsupervised domain based on depth confrontation study |
CN108564587A (en) * | 2018-03-07 | 2018-09-21 | 浙江大学 | A kind of a wide range of remote sensing image semantic segmentation method based on full convolutional neural networks |
CN108647568A (en) * | 2018-03-30 | 2018-10-12 | 电子科技大学 | Grassland degeneration extraction method based on full convolutional neural networks |
Non-Patent Citations (2)
Title |
---|
HOU QINGWEN 等: "Radar Data Processing of Blast Furnace Stock-Line Based on Spatio-Temporal Data Association", 《2015 34TH CHINESE CONTROL CONFERENCE (CCC)》 * |
陈先中 等: "高炉雷达料面测量信号处理***改进", 《北京科技大学学报》 * |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110647891A (en) * | 2019-09-17 | 2020-01-03 | 上海仪电(集团)有限公司中央研究院 | CNN (convolutional neural network) -based automatic extraction method and system for time sequence data characteristics of self-encoder |
CN110647891B (en) * | 2019-09-17 | 2023-01-24 | 上海仪电(集团)有限公司中央研究院 | CNN (convolutional neural network) -based automatic extraction method and system for time sequence data characteristics of self-encoder |
CN110819751A (en) * | 2020-01-09 | 2020-02-21 | 江苏金恒信息科技股份有限公司 | Blast furnace burden line obtaining method and device based on radar data processing |
WO2021139399A1 (en) * | 2020-01-09 | 2021-07-15 | 江苏金恒信息科技股份有限公司 | Radar data processing based blast furnace stock line obtaining method and apparatus |
CN111449644A (en) * | 2020-03-19 | 2020-07-28 | 复旦大学 | Bioelectricity signal classification method based on time-frequency transformation and data enhancement technology |
CN111784642A (en) * | 2020-06-10 | 2020-10-16 | 中铁四局集团有限公司 | Image processing method, target recognition model training method and target recognition method |
CN111784642B (en) * | 2020-06-10 | 2021-12-28 | 中铁四局集团有限公司 | Image processing method, target recognition model training method and target recognition method |
CN113627283A (en) * | 2021-07-23 | 2021-11-09 | 中冶南方工程技术有限公司 | Material level measuring method based on radar echo signals, terminal equipment and storage medium |
CN113627283B (en) * | 2021-07-23 | 2024-05-24 | 中冶南方工程技术有限公司 | Material level measuring method based on radar echo signals, terminal equipment and storage medium |
CN114136194A (en) * | 2021-10-12 | 2022-03-04 | 江苏丰尚智能科技有限公司 | Method and device for monitoring volume of material in bin, monitoring equipment and storage medium |
Also Published As
Publication number | Publication date |
---|---|
CN109490861B (en) | 2020-06-02 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109490861A (en) | A kind of blast-furnace line extracting method | |
CN109191476A (en) | The automatic segmentation of Biomedical Image based on U-net network structure | |
CN109859190A (en) | A kind of target area detection method based on deep learning | |
CN110119780A (en) | Based on the hyperspectral image super-resolution reconstruction method for generating confrontation network | |
CN107194872A (en) | Remote sensed image super-resolution reconstruction method based on perception of content deep learning network | |
CN109934154B (en) | Remote sensing image change detection method and detection device | |
CN107766794A (en) | The image, semantic dividing method that a kind of Fusion Features coefficient can learn | |
CN106683048A (en) | Image super-resolution method and image super-resolution equipment | |
CN106355151A (en) | Recognition method, based on deep belief network, of three-dimensional SAR images | |
CN107862668A (en) | A kind of cultural relic images restored method based on GNN | |
CN106570893A (en) | Rapid stable visual tracking method based on correlation filtering | |
CN108764298B (en) | Electric power image environment influence identification method based on single classifier | |
CN109492596B (en) | Pedestrian detection method and system based on K-means clustering and regional recommendation network | |
CN109685716A (en) | A kind of image super-resolution rebuilding method of the generation confrontation network based on Gauss encoder feedback | |
CN109001736A (en) | Radar echo extrapolation method based on deep space-time prediction neural network | |
CN104376529A (en) | Gray level image colorization system and method based on GLCM | |
CN104992403B (en) | Mixed operation operator image redirection method based on visual similarity measurement | |
CN103325120A (en) | Rapid self-adaption binocular vision stereo matching method capable of supporting weight | |
CN110349185A (en) | A kind of training method and device of RGBT target following model | |
CN107967474A (en) | A kind of sea-surface target conspicuousness detection method based on convolutional neural networks | |
CN108399430B (en) | A kind of SAR image Ship Target Detection method based on super-pixel and random forest | |
CN109712183A (en) | Electronic speckle interference intelligent information retrieval method based on deep learning | |
CN108664994A (en) | A kind of remote sensing image processing model construction system and method | |
CN110322403A (en) | A kind of more supervision Image Super-resolution Reconstruction methods based on generation confrontation network | |
CN107607942A (en) | Based on the large scale electromagnetic scattering of deep learning model and the Forecasting Methodology of back scattering |
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 |