CN109242859A - Remote Sensing Image Segmentation based on multilayer perceptron - Google Patents
Remote Sensing Image Segmentation based on multilayer perceptron Download PDFInfo
- Publication number
- CN109242859A CN109242859A CN201810889311.6A CN201810889311A CN109242859A CN 109242859 A CN109242859 A CN 109242859A CN 201810889311 A CN201810889311 A CN 201810889311A CN 109242859 A CN109242859 A CN 109242859A
- Authority
- CN
- China
- Prior art keywords
- remote sensing
- sensing image
- image
- multilayer perceptron
- neural network
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000003709 image segmentation Methods 0.000 title claims abstract description 21
- 238000013528 artificial neural network Methods 0.000 claims abstract description 33
- 230000011218 segmentation Effects 0.000 claims abstract description 30
- 238000012549 training Methods 0.000 claims abstract description 15
- 239000011159 matrix material Substances 0.000 claims abstract description 6
- 238000012360 testing method Methods 0.000 claims abstract description 4
- 230000004913 activation Effects 0.000 claims description 20
- 238000000513 principal component analysis Methods 0.000 claims description 3
- 230000000694 effects Effects 0.000 abstract description 9
- 230000006870 function Effects 0.000 description 19
- 238000010276 construction Methods 0.000 description 8
- 238000000034 method Methods 0.000 description 5
- 230000008901 benefit Effects 0.000 description 3
- 238000002474 experimental method Methods 0.000 description 3
- 230000009467 reduction Effects 0.000 description 3
- 238000001914 filtration Methods 0.000 description 2
- 230000002068 genetic effect Effects 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 230000000977 initiatory effect Effects 0.000 description 2
- 210000004218 nerve net Anatomy 0.000 description 2
- 241001269238 Data Species 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 239000004744 fabric Substances 0.000 description 1
- 210000002364 input neuron Anatomy 0.000 description 1
- 230000017074 necrotic cell death Effects 0.000 description 1
- 210000005036 nerve Anatomy 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 210000004205 output neuron Anatomy 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
- G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/20—Image enhancement or restoration using local operators
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10032—Satellite or aerial image; Remote sensing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20024—Filtering details
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Image Processing (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses the Remote Sensing Image Segmentations based on multilayer perceptron, comprising the following steps: S1: being filtered to original remote sensing image；S2: dimension-reduction treatment is carried out to filtered remote sensing image, obtains new characteristic image and corresponding eigenmatrix；S3: new characteristic image is equably divided into the identical small image of several pixel sizes, therefrom chooses several small images respectively as training sample and test sample；S4: initialization neural network, input training sample start to train, and stopping when network error reaches setting value or frequency of training reaches maximum value saves corresponding weight and threshold value；S5: selecting remote sensing image to be split, is trained using trained neural network to input vector, segmentation result of the last output vector as remote sensing image；S6: segmentation result is converted to the gray matrix of image by vector form.Model learning effect of the present invention is good, positive axis direction is surfed the Internet network fast convergence rate.
Description
Technical field
The present invention relates to target in hyperspectral remotely sensed image dividing methods, more particularly to the remote sensing image based on multilayer perceptron point
Segmentation method.
Background technique
In the late three decades, constantly extensive and in-depth research is unfolded in Remote Sensing Image Segmentation and its application aspect in scholars
And propose numerous image division methods.
Multilayer perceptron neural network uses back-propagation algorithm (BP algorithm) usually to carry out network training, this use
The multilayer perceptron neural network of BP algorithm is the neural network the most universal used so far.In multilayer perceptron nerve
In network, often using ReLu function or Softsign function as activation primitive.ReLu function compares Softsign function,
The advantage unsaturated with gradient in positive axis direction, calculating speed is fast, but in negative axis directions, the rough pressure of ReLu function
Sparse processing shields many useful features, and model learning effect is caused to be deteriorated.Softsign function compares ReLu function,
Do not have good sparsity in negative axis directions, avoids the missing of negative axis useful feature information, but the net in positive axis direction
Network convergence rate ratio ReLu function is many slowly.Therefore, activation primitive in the prior art have there are model learning effect is poor
The problem of, it is some to there is a problem of that positive axis direction online network convergence rate is slow.
Summary of the invention
Goal of the invention: the purpose of the present invention is aiming at the problems existing in the prior art, provide a kind of model learning effect
The Remote Sensing Image Segmentation based on multilayer perceptron of good, positive axis direction online network fast convergence rate.
Technical solution: the Remote Sensing Image Segmentation of the present invention based on multilayer perceptron, comprising the following steps:
S1: original remote sensing image is filtered；
S2: dimension-reduction treatment is carried out to filtered remote sensing image, obtains new characteristic image and corresponding eigenmatrix；
S3: new characteristic image is equably divided into the identical small image of several pixel sizes, is therefrom chosen respectively
Several small images are as training sample and test sample；
S4: initialization neural network, input training sample starts to train, until network error reaches setting value or training
Stopping when number reaches maximum value saves corresponding weight and threshold value；
S5: selecting remote sensing image to be split, is trained using trained neural network to input vector, most
Segmentation result of the output vector afterwards as remote sensing image；
S6: segmentation result is converted to the gray matrix of image by vector form.
Further, the dimension-reduction treatment in the step S2 is realized by Principal Component Analysis.
Further, it is filtered in the step S1 using Lee filter.
Further, in the step S6, after being converted to gray matrix, segmentation result is shown.
Further, in the step S4, neural network is trained using activation primitive f (x):
In formula (1), x is input data.
The utility model has the advantages that constructing one the invention discloses a kind of Remote Sensing Image Segmentation based on multilayer perceptron
Linear activation primitive is corrected in the unsaturation of new segmentation, is effectively increased network convergence rate and segmentation precision, is also effectively improved
Model learning effect.
Detailed description of the invention
Fig. 1 is the flow chart of neural network in the specific embodiment of the invention；
Fig. 2 is the flow chart of Image Segmentation in the specific embodiment of the invention；
Fig. 3 is the schematic diagram of activation primitive in the specific embodiment of the invention；
Fig. 4 be in the specific embodiment of the invention accuracy that is split with the network of different activation primitives with training
The change curve of number.
Specific embodiment
Present embodiment discloses a kind of Remote Sensing Image Segmentation based on multilayer perceptron, as shown in Fig. 2, packet
Include following steps:
S1: original remote sensing image is filtered using Lee filter；
S2: carrying out dimension-reduction treatment to filtered remote sensing image by Principal Component Analysis, obtain new characteristic image and
Corresponding eigenmatrix；
S3: new characteristic image is equably divided into the identical small image of several pixel sizes, is therefrom chosen respectively
Several small images are as training sample and test sample；
S4: as shown in Figure 1, initialization neural network, input training sample starts to train, until network error reaches setting
Stopping when value or frequency of training reach maximum value saves corresponding weight and threshold value；
S5: selecting remote sensing image to be split, is trained using trained neural network to input vector, most
Segmentation result of the output vector afterwards as remote sensing image, shows segmentation result；
S6: segmentation result is converted to the gray matrix of image by vector form.
In step S4, neural network is trained using activation primitive f (x):
In formula (1), x is input data.Activation primitive f (x) is denoted as Relu_Softsign function, if Fig. 3 is that the function shows
It is intended to.
Experiment simulation process and result are introduced below:
1, experimental image
It is in 2011 by Belgian Royal Military College in Belgian damp cloth that experiment remote sensing image data used, which collects,
A part that the harbour Lu and urban area are collected by airborne platform.Remote sensing image is by filtering and obtaining after PCA dimensionality reduction
The characteristic image of 10000*10000, emulation experiment use MultiSpecWin64, MATLABR2013b, UltraEdit
V24.10.0.32 (x64) and the programming of JetBrains PyCharm 2017.2.3x64 software systems.
2, experimentation
Remote sensing image is by filtering and obtaining the characteristic image of 10000*10000 after PCA dimensionality reduction in this algorithm, by characteristic pattern
Small image as being cut into 500*500 20*20 pixel.Present embodiment chooses totally 100000 small image and is used as training
Sample, that is to say, that the input sample number of neural network is 100000, has 20*20 to be equal to 400 inputs in each sample special
Data are levied, i.e. input vector is 400 dimensions, and input neuron number is 400.100000 groups of sample datas are divided into target, back
Scape two categories, i.e. output vector are 2 dimensions, and output neuron number is 2.
3, arithmetic result compares
Based on four kinds of traditional activation primitives and improved activation primitive, five kinds of different multilayer perceptron nerve nets are designed
Network is used for the segmentation of remote sensing image, and compares point with GA (Genetic Algorithm, genetic algorithm)-BP neural network
Analysis.The network structure of neural network is as shown in the table:
The structure of 1 multilayer perceptron neural network of table
When constructing multilayer perceptron neural network using traditional S type activation primitive, the relevant parameter initialization of network
It is as shown in the table:
2 multilayer perceptron neural network initiation parameter table (S type activation primitive) of table
When using linear activation primitive construction multilayer perceptron neural network is corrected, since the performance of function is by network science
Habit rate is affected, and is easier to lead to the generation of neuron " necrosis " phenomenon when learning rate is big, so by the study of network
Rate is arranged to a suitable smaller value.The relevant parameter initialization of network is as shown in the table:
3 multilayer perceptron neural network initiation parameter table (correcting linear activation primitive) of table
The Image Segmentation precision of the different neural networks of table 4
The BP network of Sigmoid construction of function is used it can be seen from Fig. 4 and table 4, Image Segmentation precision is minimum
, only 72.23%, and the BP network of Sigmoid construction of function is difficult to restrain, need to carry out network parameter it is a large amount of repeatedly
Adjustment；Compared to the BP network for using Sigmoid construction of function, obviously mentioned using the convergence rate of the BP network of Tanh construction of function
It is high much, but its Image Segmentation precision also only has 77.83%；Use the BP network of Softsign construction of function, shadow
It compares the first two network as segmentation precision to improve a lot, highest segmentation precision reaches 82.32%, but its shortcoming exists
It is relatively slow in the convergence rate of network；Using the BP network of ReLu construction of function, not only its Image Segmentation precision is higher, but also
The convergence rate of network is also very fast, and highest segmentation precision is 91.46%；And use improved activation primitive Relu_Softsign
The BP network of construction, Image Segmentation precision be it is highest, highest segmentation precision reaches 93.60%, and remains ReLu letter
The fast advantage of number network convergence rate.It therefore, not only can be with using improved activation primitive compared to common activation primitive
The convergence rate of network is improved, and the segmentation precision of network can be improved.
The BP neural network for improving front and back and GA-BP neural network are compared again, it can be seen that the BP nerve net before improvement
Network, noise proof feature is poor, and being also not for edge segmentation of image body target will be apparent that；Using GA-BP neural network
The segmentation effect of acquisition is slightly better than the BP neural network effect before improving, and noiseproof feature also increases；And it is improved
BP neural network further eliminates a large amount of noise and isolated point, compares first two algorithm, and segmentation effect is ideal.
The Image Segmentation precision of the different neural networks of table 5
As shown in Table 5, when carrying out Image Segmentation using the BP neural network before improvement, using ReLu function as activation primitive
Network segmentation precision be significantly improved than the network segmentation precision using Softsign function as activation primitive, but it is anti-
Performance of making an uproar is all poor, and there are the also very fuzzy of the edge of a large amount of noise speckle and subject goal segmentation；Use GA-BP mind
Through network obtain segmentation precision than improve before BP neural network obtain segmentation precision it is more slightly higher, noiseproof feature and
Marginal definition also all increases, but still there are many noise speckles；And it is obtained using improved BP neural network
Segmentation precision be it is highest, noise speckle has apparent reduction relative to first two network, and segmentation effect is ideal.
Claims (5)
1. the Remote Sensing Image Segmentation based on multilayer perceptron, it is characterised in that: the following steps are included:
S1: original remote sensing image is filtered；
S2: dimension-reduction treatment is carried out to filtered remote sensing image, obtains new characteristic image and corresponding eigenmatrix；
S3: new characteristic image is equably divided into the identical small image of several pixel sizes, is therefrom chosen respectively several
A small image is as training sample and test sample；
S4: initialization neural network, input training sample starts to train, until network error reaches setting value or frequency of training
Stopping when reaching maximum value saves corresponding weight and threshold value；
S5: selecting remote sensing image to be split, is trained using trained neural network to input vector, last
Segmentation result of the output vector as remote sensing image；
S6: segmentation result is converted to the gray matrix of image by vector form.
2. the Remote Sensing Image Segmentation according to claim 1 based on multilayer perceptron, it is characterised in that: the step
Dimension-reduction treatment in S2 is realized by Principal Component Analysis.
3. the Remote Sensing Image Segmentation according to claim 1 based on multilayer perceptron, it is characterised in that: the step
It is filtered in S1 using Lee filter.
4. the Remote Sensing Image Segmentation according to claim 1 based on multilayer perceptron, it is characterised in that: the step
In S6, after being converted to gray matrix, segmentation result is shown.
5. the Remote Sensing Image Segmentation according to claim 1 based on multilayer perceptron, it is characterised in that: the step
In S4, neural network is trained using activation primitive f (x):
In formula (1), x is input data.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810889311.6A CN109242859A (en) | 2018-08-07 | 2018-08-07 | Remote Sensing Image Segmentation based on multilayer perceptron |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810889311.6A CN109242859A (en) | 2018-08-07 | 2018-08-07 | Remote Sensing Image Segmentation based on multilayer perceptron |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109242859A true CN109242859A (en) | 2019-01-18 |
Family
ID=65070582
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810889311.6A Pending CN109242859A (en) | 2018-08-07 | 2018-08-07 | Remote Sensing Image Segmentation based on multilayer perceptron |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109242859A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112819077A (en) * | 2021-02-02 | 2021-05-18 | 河南大学 | Hyperspectral image classification method based on novel activation function |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101814183A (en) * | 2010-01-08 | 2010-08-25 | 清华大学 | Image segmentation method and system |
| CN102915445A (en) * | 2012-09-17 | 2013-02-06 | 杭州电子科技大学 | Method for classifying hyperspectral remote sensing images of improved neural network |
| CN106096652A (en) * | 2016-06-12 | 2016-11-09 | 西安电子科技大学 | Based on sparse coding and the Classification of Polarimetric SAR Image method of small echo own coding device |
-
2018
- 2018-08-07 CN CN201810889311.6A patent/CN109242859A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101814183A (en) * | 2010-01-08 | 2010-08-25 | 清华大学 | Image segmentation method and system |
| CN102915445A (en) * | 2012-09-17 | 2013-02-06 | 杭州电子科技大学 | Method for classifying hyperspectral remote sensing images of improved neural network |
| CN106096652A (en) * | 2016-06-12 | 2016-11-09 | 西安电子科技大学 | Based on sparse coding and the Classification of Polarimetric SAR Image method of small echo own coding device |
Non-Patent Citations (1)
| Title |
|---|
| GUIFANG LIN等: "Research on convolutional neural network based on improved Relu piecewise activation function", 《PROCEDIA COMPUTER SCIENCE》 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112819077A (en) * | 2021-02-02 | 2021-05-18 | 河南大学 | Hyperspectral image classification method based on novel activation function |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Divakar et al. | Image denoising via CNNs: An adversarial approach | |
| Zhu et al. | Data Augmentation using Conditional Generative Adversarial Networks for Leaf Counting in Arabidopsis Plants. | |
| CN107609638B (en) | method for optimizing convolutional neural network based on linear encoder and interpolation sampling | |
| GB2596886A (en) | Method and device for performing inversion of crop leaf area index | |
| CN107330405A (en) | Remote sensing images Aircraft Target Recognition based on convolutional neural networks | |
| CN112285712B (en) | Method for improving detection precision of coasting ship in SAR image | |
| CN112862792B (en) | Wheat powdery mildew spore segmentation method for small sample image dataset | |
| CN106991440B (en) | Image classification method of convolutional neural network based on spatial pyramid | |
| CN108304826A (en) | Facial expression recognizing method based on convolutional neural networks | |
| CN105320965A (en) | Hyperspectral image classification method based on spectral-spatial cooperation of deep convolutional neural network | |
| CN106682569A (en) | Fast traffic signboard recognition method based on convolution neural network | |
| CN110503610A (en) | A kind of image sleet trace minimizing technology based on GAN network | |
| CN107103338A (en) | Merge the SAR target identification methods of convolution feature and the integrated learning machine that transfinites | |
| CN107256423A (en) | A kind of neural planar network architecture of augmentation and its training method, computer-readable recording medium | |
| CN106600595A (en) | Human body characteristic dimension automatic measuring method based on artificial intelligence algorithm | |
| CN104282003B (en) | Digital blurred image blind restoration method based on gradient screening | |
| CN107967474A (en) | A kind of sea-surface target conspicuousness detection method based on convolutional neural networks | |
| CN104268524A (en) | Convolutional neural network image recognition method based on dynamic adjustment of training targets | |
| CN109711401A (en) | A kind of Method for text detection in natural scene image based on Faster Rcnn | |
| CN110503613A (en) | Based on the empty convolutional neural networks of cascade towards removing rain based on single image method | |
| CN111553462A (en) | Class activation mapping method | |
| CN106529458A (en) | Deep neural network space spectrum classification method for high-spectral image | |
| CN110096976A (en) | Human behavior micro-Doppler classification method based on sparse migration network | |
| CN109872326A (en) | Profile testing method based on the connection of deeply network hop | |
| CN111062310B (en) | Few-sample unmanned aerial vehicle image identification method based on virtual sample generation |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190118 |