CN107122796B - A kind of remote sensing image classification method based on multiple-limb network integration model - Google Patents
A kind of remote sensing image classification method based on multiple-limb network integration model Download PDFInfo
- Publication number
- CN107122796B CN107122796B CN201710213577.4A CN201710213577A CN107122796B CN 107122796 B CN107122796 B CN 107122796B CN 201710213577 A CN201710213577 A CN 201710213577A CN 107122796 B CN107122796 B CN 107122796B
- Authority
- CN
- China
- Prior art keywords
- layer
- image
- branch
- input
- output
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/254—Fusion techniques of classification results, e.g. of results related to same input data
-
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/285—Selection of pattern recognition techniques, e.g. of classifiers in a multi-classifier system
-
- 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)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Biophysics (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The present invention relates to a kind of remote sensing image classification methods based on multiple-limb network integration model, and this method comprises the following steps: S1, construct multiple-limb network integration model, include a main split S, at least one other branch { Tk, k=1 ..., K, K >=1 }；S2 is acquired remote sensing image as training image, and carry out classification mark to training image, is learnt multiple-limb network integration model parameter out from the training image marked using back-propagation algorithm；S3, image are input in the multiple-limb network integration model for having learnt parameter by the input layer I in main split S, layer-by-layer processing and all other branch { T of the image by main split SkLayer-by-layer processing, finally by the cutting layer W output category score image in main split S.The present invention greatly improves remote sensing image nicety of grading using the multiple-limb network integration shallow-layer and deep layer characteristics of image, and can once generate the intensive semantic segmentation of size identical as input picture as a result, improving image classification efficiency.
Description
Technical field
This hair belongs to the technical field of remote sensing image classification, more particularly to a kind of based on multiple-limb network integration model
Remote sensing image classification method.
Background technique
With the rapid development and extensive use of image sensor technologies, image data explosive growth, image data amount
Have become magnanimity grade.Conventional artificial image's classification method is no longer feasible, therefore currently based on the figure of computer vision technique
As automatic classification has become a very powerful and exceedingly arrogant research topic.Recent study persons have proposed many image classification sides
Method, but remote sensing image classification at this stage still faces huge challenge: and content complexity, object category are numerous in remote sensing image
More, difference is very big in image scene feature class.Previous shallow-layer method does not utilize the layering of multilayered structure extraction image
Feature realizes classification merely with shallow-layer feature and shallow-layer classifier, and thus the expression ability of shallow-layer method, classification capacity have very much
Limit, image classification accuracy be not high.In recent years, with the development of deep learning, many image classification methods based on deep learning
Also it is suggested, such as SAE (stacked autoencoder), DBN (deep belief network), CNN
(convolutional neural network) even depth model, although the method based on depth model compares shallow-layer method pole
Image classification accuracy is improved greatly, but for complicated remote sensing image, nicety of grading is still to be improved.In addition most of
Depth model does not merge layered characteristic for classifying, and only receives fixed-size input picture, can not once calculate
The intensive semantic segmentation result of (needing through multiple network query function) size identical as input picture.
Summary of the invention
The technical problems to be solved by the present invention are: current depth model most of in the prior art does not merge layering
Feature only receives fixed-size input picture for classifying, and can not once calculate and (need through multiple network meter
Calculate) the intensive semantic segmentation result of identical as input picture size.
To solve technical problem above, the present invention provides a kind of optical remote sensings based on multiple-limb network integration model
Image classification method, the algorithm include the following steps:
S1 constructs multiple-limb network integration model, includes a main split S, at least one other branch { Tk, k=1 ...,
K, K >=1 }, wherein K indicates the item number of other branch；
S2 acquires remote sensing image as training image, and carry out classification mark to training image, utilizes backpropagation
Algorithm learns multiple-limb network integration model parameter out from the training image marked；
S3, image are input to the multiple-limb network integration model for having learnt parameter by the input layer I in main split S
In, layer-by-layer processing and all other branch { T of the image by main split SkLayer-by-layer processing, finally by the sanction in main split S
Cut a layer W output category score image.
Beneficial effects of the present invention: multiple-branching construction can in the multiple-limb network integration model used in through the invention
It merges shallow-layer and further feature, and by fusion feature classification remote sensing image, greatly improves nicety of grading, more points
Layer is up-sampled in branch network integration model uses the image for enabling input layer in the method for the present invention to receive arbitrary dimension defeated
Enter, and can once generate the intensive semantic segmentation of the size identical as input picture (without multiple network query function) as a result, improving
Image classification efficiency.
Further, for the step S2 specifically, the number of species of remote sensing image are arranged, classification number is C, acquires light
Remote sensing images are learned as training image, and classification mark is carried out to training image according to the remote sensing image type, are utilized
Back-propagation algorithm learns multiple-limb network integration model parameter out from the training image marked.
Further, in the step S1, the main split S successively includes an input layer I, at least one convolutional layer
{ni, i=1 ..., N, N >=1 }, at least one excitation layer { lj, j=1 ..., L, L >=1 }, at least one dimensionality reduction layer { mp, p=
1 ..., M, M >=1 }, at least one up-sampling layer { uq, q=1 ..., U, U >=1 }, a cutting layer W, as N > 1 or M > 1
Convolutional layer and dimensionality reduction layer cross arrangement, wherein N indicates the number of plies of convolutional layer, and M indicates the number of plies of dimensionality reduction layer, and L indicates excitation layer
The number of plies, U indicate the number of plies of up-sampling layer.
Further, in main split S, the excitation layer ljIn, each excitation layer ljOne layer of front is all a convolution
Layer；The last one convolutional layer n in main split SNConvolution kernel number be the classification number C.
Further, every Tiao Pang branch TkIt successively include a convolutional layer ak, one cut layer bk, a superposition
Layer ck, a up-sampling layer dk；Other branch { Tk, k=1 ..., K, K >=1 } in all convolutional layer { ak, k=1 ..., K, K >=1 }
Convolution kernel number be all the classification number C.
Further, include: in the step S3
Input layer I in S31, the main split S receives image input；
Convolutional layer n in S32, main split SiConvolution operation is executed to adjacent upper one layer of output image, extracts characteristics of image,
And characteristic image is exported；
Excitation layer l in S33, main split SjUsing correction linear unit to the output characteristic image of adjacent upper one layer of convolutional layer
It executes nonlinear function to calculate, and export the image generated is calculated；
Dimensionality reduction layer m in S34, main split SpTo adjacent upper one layer of excitation layer ljOutput image execute dimensionality reduction, and by dimensionality reduction figure
As output；
Layer u is up-sampled in S35, main split SqDeconvolution operation is executed to adjacent upper one layer of output image, is realized to figure
The liter of picture is tieed up, and exports a liter dimension image；
S36 selects K dimensionality reduction layer from main split S, and the output image of the K dimensionality reduction layer is separately input to other point
Branch { Tk, k=1 ..., K, K >=1 } in, other branch's classification is arranged to make up multilevel structure, by the layer-by-layer meter of each other branch
It calculates, by branch T by afterbodyKOutput image be input in the cutting layer W in main split S, the cutting layer W in main split S
The input picture is cut into size identical with the image of input layer I input in main split S, while the image after cutting is made
For the output of entire multiple-limb network integration model, classification score image is obtained.
It is above-mentioned further the utility model has the advantages that above-mentioned main split S by the processing of each level and it is each by branch it is layer-by-layer
Processing, eventually by output category score image in cutting layer W last in main split.
Further, the step S32, specifically:
The convolutional layer niConvolution operation is executed to adjacent upper one layer of output image, if convolutional layer niAdjacent upper one layer
For excitation layer lj, then to excitation layer ljThe image of output executes convolution operation；If convolutional layer niAdjacent upper one layer is convolutional layer
ni-1, then to convolutional layer ni-1The image of output executes convolution operation；If convolutional layer niAdjacent upper one layer is dimensionality reduction layer mp, then right
Dimensionality reduction layer mpThe image of output executes convolution operation.
Further, it is specifically included in the step S36:
S361 selects K dimensionality reduction layer from main split S, and the output image of the K dimensionality reduction layer is separately input to other point
Branch { Tk, k=1 ..., K, K >=1 } in, i.e. a dimensionality reduction layer m in main split SpOutput image be input to a Ge Pang branch TkIn
Convolutional layer akIn, convolutional layer akConvolution operation is executed to the input picture of this layer；
S362, according to S361 Zhong Pang branch { Tk, k=1 ..., K, K >=1 } convolutional layer { ak, k=1 ..., K }
The size of input picture, by ascending sequence by other branch { Tk, k=1 ..., K } and it is arranged as multilevel structure, simultaneously
Every Tiao Pang branch TkOutput image by side branch TkIn up-sampling layer dkOutput, and side branch TkOutput image
Input picture as branch by adjacent next stage；
S363, every Tiao Pang branch TkMiddle cutting layer bkTwo inputs are received, if side branch TkIt is branch by the first order, then
Side branch TkMiddle cutting layer bkFirst input be current other branch TkMiddle convolutional layer akOutput image, second input
It is up-sampling layer u in main split SUOutput；If side branch TkIt is not branch by the first order, then side branch TkMiddle cutting layer
bkFirst input be current other branch TkMiddle convolutional layer akOutput image, second input is branch T by upper levelk-1's
Up-sample layer dk-1Output image；Side branch TkMiddle cutting layer bkFirst input is cut into and second input size phase
Same image, and the image after cutting is exported；
S364, every Tiao Pang branch TkMiddle superimposed layer ckTwo inputs are received, if side branch TkIt is branch by the first order, then
Side branch TkMiddle superimposed layer ckFirst input be current other branch TkMiddle cutting layer bkOutput image, second input
It is up-sampling layer u in main split SUOutput；If side branch TkIt is not branch by the first order, then side branch TkMiddle superimposed layer
ckFirst input be current other branch TkMiddle cutting layer bkOutput image, second input is branch T by upper levelk-1's
Up-sample layer dk-1Output image；Side branch TkMiddle superimposed layer ckTwo input pictures are added up, and will it is cumulative after image
Output；
S365, every Tiao Pang branch TkMiddle up-sampling layer dkTo the superimposed layer c in current other branchkOutput image execute it is anti-
Convolution operation, and the image output after deconvolution is operated；
S366, by branch T by afterbodykIn up-sampling layer dkThe image of output is input to the cutting layer in main split S
In W, which is cut into identical with the image of input layer I input in main split S by the cutting layer W in main split S
Size, while output of the image after cutting as entire multiple-limb network integration model, obtain classification score image.
Detailed description of the invention
Fig. 1 is a kind of remote sensing image classification method flow chart based on multiple-limb network integration model of the invention.
Fig. 2 is the schematic diagram of multiple-limb of embodiment of the present invention network integration model structure.
Specific embodiment
The principle and features of the present invention will be described below with reference to the accompanying drawings, and the given examples are served only to explain the present invention, and
It is non-to be used to limit the scope of the invention.
As shown in Figure 1, a kind of remote sensing image classification method based on multiple-limb network integration model of the present invention should
Algorithm includes the following steps:
The first step first constructs multiple-limb network integration model, and wherein multiple-limb network integration model specifically includes: a master
Branch S, at least one other branch { Tk, k=1 ..., K, K >=1 }；The main split S successively includes input layer I, extremely
A few convolutional layer { ni, i=1 ..., N, N >=1 }, at least one excitation layer { lj, j=1 ..., L, L >=1 }, at least one drop
Tie up layer { mp, p=1 ..., M, M >=1 }, at least one up-sampling layer { uq, q=1 ..., U, U >=1 }, a cutting layer W, work as N
Convolutional layer and dimensionality reduction layer cross arrangement when > 1 or M > 1, wherein N indicates the number of plies of convolutional layer, and M indicates the number of plies of dimensionality reduction layer, L table
Show the number of plies of excitation layer, U indicates the number of plies of up-sampling layer；In main split S, the excitation layer ljIn, each excitation layer ljFront
One layer is all a convolutional layer；The last one convolutional layer n in main split SNConvolution kernel number be the classification number C；It is described
Every Tiao Pang branch TkIt successively include a convolutional layer ak, one cut layer bk, a superimposed layer ck, a up-sampling layer dk；Often
The image for receiving two different layers output in main split S as two inputs and is merged the two inputs by Tiao Pang branch, is realized
The fusion of the characteristics of image of different levels；Other branch { Tk, k=1 ..., K, K >=1 } in all convolutional layer { ak, k=1 ...,
K, K >=1 } convolution kernel number be all the classification number C；Convolutional layer in the present invention is mentioned from input picture using convolution operation
Characteristics of image is taken, convolution operation mainly carries out convolution algorithm using convolution collecting image.Convolution kernel is generally the square of h × h,
H indicates the size of convolution kernel, and the value of each position needs to learn from training image to come out in square；In the present invention
Excitation layer using correct linear unit to image execute nonlinear function calculate, enhance the non-linear property of network；The present invention
In dimensionality reduction layer to image execute dimensionality reduction operation, realize the dimensionality reduction of image；Up-sampling layer in the present invention executes warp to image
The liter dimension of image is realized in product operation；Two image croppings at identical size, are realized picture size by the cutting layer in the present invention
Normalization；Two image superpositions are realized the fusion of characteristics of image by the superimposed layer in the present invention
The number of species of remote sensing image are arranged in second step, and classification number is C, such as remote sensing image type is to build
Build object, road, meadow, trees etc.；Remote sensing image is acquired as training image, and according to the remote sensing image type
Classification mark is carried out to training image, learns multiple-limb network out from the training image marked using back-propagation algorithm and melts
Mold shape parameter；
Third step, image are input to the multiple-limb network integration mould for having learnt parameter by the input layer I in main split S
In type, layer-by-layer processing and all other branch { T of the image by main split SkLayer-by-layer processing, finally by main split S
Cut layer W output category score image；Wherein start, the input layer I in the main split S receives image input；Then, main point
Convolutional layer n in branch SiConvolution operation is executed to adjacent upper one layer of output image, extracts characteristics of image, and characteristic image is defeated
Out；Followed by excitation layer lj holds the output characteristic image of adjacent upper one layer of convolutional layer using linear unit is corrected in main split S
Row nonlinear function calculates, and exports the image generated is calculated；Then, dimensionality reduction layer m in main split SpTo adjacent upper one
Layer excitation layer ljOutput image execute dimensionality reduction, and dimensionality reduction image is exported；Subsequently, layer u is up-sampled in main split SqTo adjacent
Upper one layer of output image executes deconvolution operation, realizes and ties up to the liter of image, and exports a liter dimension image；Finally, from main split S
The output image of the K dimensionality reduction layer is separately input to other branch { T by K dimensionality reduction layer of middle selectionk, k=1 ..., K, K >=1 }
In, other branch's classification is arranged to make up multilevel structure, by the layer-by-layer calculating of each other branch, by branch T by afterbodyK's
Output image is input in the cutting layer W in main split S, which is cut into and main point by the cutting layer W in main split S
The identical size of image of input layer I input in branch S, while the image after cutting is as entire multiple-limb network integration model
Output, obtain classification score image.
Above-mentioned described excitation layer l in the present inventionjFront all immediately convolutional layer ni, such as convolutional layer n in Fig. 21With swash
Encourage a layer l1, convolutional layer n2With excitation layer l2, convolutional layer n3With excitation layer l3Deng.In addition, dimensionality reduction layer is to intersect with convolutional layer, excitation layer
Arrangement, such as in Fig. 2, all dimensionality reduction layer m1、m2、m3、m4、m5It is interspersed in convolutional layer n1-n15And excitation layer l1-l15In.
Preferably, the convolutional layer niConvolution operation is executed to adjacent upper one layer of output image, if convolutional layer niIt is adjacent
Upper one layer is excitation layer lj, then to excitation layer ljThe image of output executes convolution operation；If convolutional layer niAdjacent upper one layer is volume
Lamination ni-1, then to convolutional layer ni-1The image of output executes convolution operation；If convolutional layer niAdjacent upper one layer is dimensionality reduction layer mp,
Then to dimensionality reduction layer mpThe image of output executes convolution operation.
In the present invention, K dimensionality reduction layer is selected from main split S, the output image of the K dimensionality reduction layer is separately input to
Other branch { Tk, k=1 ..., K, K >=1 } in, i.e. a dimensionality reduction layer m in main split SpOutput image be input to a Ge Pang branch
TkMiddle convolutional layer akIn, convolutional layer akConvolution operation is executed to the input picture of this layer；
Then, according to S361 Zhong Pang branch { Tk, k=1 ..., K, K >=1 } convolutional layer { ak, k=1 ..., K }
The size of input picture, by ascending sequence by other branch { Tk, k=1 ..., K } and it is arranged as multilevel structure, simultaneously
Every Tiao Pang branch TkOutput image by side branch TkIn up-sampling layer dkOutput, and side branch TkOutput image
Input picture as branch by adjacent next stage；
Followed by every Tiao Pang branch TkMiddle cutting layer bkTwo inputs are received, if side branch TkIt is branch by the first order,
Then side branch TkMiddle superimposed layer bkFirst input be current other branch TkMiddle convolutional layer akOutput image, second is defeated
Entering is up-sampling layer u in main split SUOutput；If side branch TkIt is not branch by the first order, then side branch TkMiddle cutting
Layer bkFirst input be current other branch TkMiddle convolutional layer akOutput image, second input is branch T by upper levelk-1
Up-sampling layer dk-1Output image；Side branch TkMiddle cutting layer bkFirst input is cut into and second input size
Identical image, and the image after cutting is exported；
Followed by every Tiao Pang branch TkMiddle superimposed layer ckTwo inputs are received, if side branch TkIt is branch by the first order,
Then side branch TkMiddle superimposed layer ckFirst input be current other branch TkMiddle cutting layer bkOutput image, second is defeated
Entering is up-sampling layer u in main split SUOutput；If side branch TkIt is not branch by the first order, then side branch TkMiddle superposition
Layer ckFirst input be current other branch TkMiddle cutting layer bkOutput image, second input is branch T by upper levelk-1
Up-sampling layer dk-1Output image；Side branch TkMiddle superimposed layer ckTwo input pictures are added up, and will it is cumulative after figure
As output；
Followed by every Tiao Pang branch TkMiddle up-sampling layer dkTo the superimposed layer c in current other branchkOutput image execute
Deconvolution operation, and the image output after deconvolution is operated；
Finally, by branch T by afterbodykIn up-sampling layer dkThe image of output is input to the cutting layer in main split S
In W, which is cut into identical with the image of input layer I input in main split S by the cutting layer W in main split S
Size, while output of the image after cutting as entire multiple-limb network integration model, obtain classification score image.
As shown in Fig. 2, for the embodiment in the present invention, the layered structure of main split S in embodiment are as follows: input layer-convolution
3 convolutional layer 4- excitation layer 4- of layer 1- excitation layer 1- convolutional layer 2- excitation layer 2- dimensionality reduction layer 1- convolutional layer 3- excitation layer
Dimensionality reduction layer 2- convolutional layer 5- excitation layer 5- convolutional layer 6- excitation layer 6- convolutional layer 7- excitation layer 7- dimensionality reduction layer 3- convolution
Layer 8- excitation layer 8- convolutional layer 9- excitation layer 9- convolutional layer 10- excitation layer 10- dimensionality reduction layer 4- convolutional layer 11- excitation layer
11- convolutional layer 12- excitation layer 12- convolutional layer 13- excitation layer 13- dimensionality reduction layer 5- convolutional layer 14- excitation layer 14- convolution
Layer 15- excitation layer 15- convolutional layer 16 up-samples layer 2 followed by up-sampling layer 1-, is finally to cut layer 3.Wherein, in Fig. 2
First layer is the input layer of main split S, and image is inputted by the input layer, is then passed through convolutional layer, convolutional layer starts to input
The image of layer carries out convolution, then executes nonlinear function by excitation layer and calculates, and a volume is connected to before each excitation layer
Lamination, dimensionality reduction layer 1-5 are interspersed among convolutional layer 1-15 and excitation layer 1-15, and convolutional layer 16 is using C convolution kernel to excitation layer
15 input picture executes convolution operation, and the input picture of up-sampling 1 pair of convolutional layer 16 of layer executes deconvolution operation, up-samples layer
The output image of 2 pairs of up-sampling layers 1 finally cuts 3 output category score image of layer after dimension of continuing rising.
In the embodiment of the present invention as shown in Figure 2, other branch receives image from the dimensionality reduction layer of main split, if the dimensionality reduction layer
Output picture size be it is the smallest, then the side branch is the first order, and the input of convolutional layer 17 is to drop in branch by the first order in Fig. 2
The output image of layer 4 is tieed up, the output image for cutting the up-sampling layer 2 that layer 1 and superimposed layer 1 receive main split is inputted as one,
Image is exported finally by the up-sampling layer 3 of the side branch；The input of convolutional layer 18 is dimensionality reduction layer 3 in the other branch of the second level
Output image cuts the output for up-sampling layer 3 that layer 2 and superimposed layer 2 receive by the first order in branch and is used as input, finally by
The up-sampling layer 4 of the side branch exports image；It is same reason for third, the other other branch of the fourth estate, until most
Branch process finishes by rear stage, exports image by the up-sampling layer of branch by afterbody, while the output image enters
In the cutting layer 3 of main split, the classification score image that layer 3 exports the application is finally cut.
In the present specification, the schematic representation of the above terms does not necessarily have to refer to the same embodiment or example.
Moreover, particular features, structures, materials, or characteristics described can be in any one or more of the embodiments or examples with suitable
Mode combines.In addition, without conflicting with each other, those skilled in the art can be by difference described in this specification
The feature of embodiment or example and different embodiments or examples is combined.
The foregoing is merely presently preferred embodiments of the present invention, is not intended to limit the invention, it is all in spirit of the invention and
Within principle, any modification, equivalent replacement, improvement and so on be should all be included in the protection scope of the present invention.
Claims (4)
1. a kind of remote sensing image classification method based on multiple-limb network integration model, which is characterized in that this method includes
Following steps:
S1 constructs multiple-limb network integration model, includes a main split S, at least one other branch { Tk, k=1 ..., K, K >=
1 }, wherein K indicates the item number of other branch；
S2 acquires remote sensing image as training image, and carry out classification mark to training image, utilizes back-propagation algorithm
Learn multiple-limb network integration model parameter out from the training image marked；
S3, image are input in the multiple-limb network integration model for having learnt parameter by the input layer I in main split S, figure
Layer-by-layer processing and all other branch { T of the picture by main split SkLayer-by-layer processing, finally by the cutting layer W in main split S
Output category score image；
The step S2 is specifically included:
The number of species of remote sensing image are set, classification number is C, remote sensing image is acquired as training image, and according to
The remote sensing image type carries out classification mark to training image, using back-propagation algorithm from the training image marked
In learn multiple-limb network integration model parameter out；
It is specifically included in the step S3:
Input layer I in S31, the main split S receives image input；
Convolutional layer n in S32, main split SiConvolution operation is executed to adjacent upper one layer of output image, extracts characteristics of image, and will
Characteristic image output；
Excitation layer l in S33, main split SjThe output characteristic image of adjacent upper one layer of convolutional layer is executed using linear unit is corrected
Nonlinear function calculates, and exports the image generated is calculated；
Dimensionality reduction layer m in S34, main split SpTo adjacent upper one layer of excitation layer ljOutput image execute dimensionality reduction, and it is dimensionality reduction image is defeated
Out；
Layer u is up-sampled in S35, main split SqDeconvolution operation is executed to adjacent upper one layer of output image, realizes the liter to image
Dimension, and export a liter dimension image；
S36 selects K dimensionality reduction layer from main split S, the output image of the K dimensionality reduction layer is separately input to other branch { Tk,
K=1 ..., K, K >=1 } in, other branch's classification is arranged to make up multilevel structure, will most by the layer-by-layer calculating of each other branch
Branch T by rear stageKOutput image be input in the cutting layer W in main split S, the cutting layer W in main split S is by the input
Image cropping is at size identical with the image of input layer I input in main split S, while the image after cutting is as entire more
The output of branching networks Fusion Model obtains classification score image；
It is specifically included in the step S36:
S361 selects K dimensionality reduction layer from main split S, the output image of the K dimensionality reduction layer is separately input to other branch
{Tk, k=1 ..., K, K >=1 } in, i.e. a dimensionality reduction layer m in main split SpOutput image be input to a Ge Pang branch TkMiddle volume
Lamination akIn, convolutional layer akConvolution operation is executed to the input picture of this layer；
S362, according to S361 Zhong Pang branch { Tk, k=1 ..., K, K >=1 } convolutional layer { ak, k=1 ..., K } input
The size of image, by ascending sequence by other branch { Tk, k=1 ..., K } and it is arranged as multilevel structure, while every
Other branch TkOutput image by side branch TkIn up-sampling layer dkOutput, and side branch TkOutput image conduct
The input picture of branch by adjacent next stage；
S363, every Tiao Pang branch TkMiddle cutting layer bkTwo inputs are received, if side branch TkIt is branch by the first order, then by this
Branch TkMiddle superimposed layer bkFirst input be current other branch TkMiddle convolutional layer akOutput image, second input is main
Layer u is up-sampled in branch SUOutput；If side branch TkIt is not branch by the first order, then side branch TkMiddle cutting layer bk's
First input is current other branch TkMiddle convolutional layer akOutput image, second input is branch Tk by upper level-1It is upper
Sample level dk-1Output image；Side branch TkMiddle cutting layer bkFirst input is cut into identical as second input size
Image, and by after cutting image export；
S364, every Tiao Pang branch TkMiddle superimposed layer ckTwo inputs are received, if side branch TkIt is branch by the first order, then by this
Branch TkMiddle superimposed layer ckFirst input be current other branch TkMiddle cutting layer bkOutput image, second input is main
Layer u is up-sampled in branch SUOutput；If side branch TkIt is not branch by the first order, then side branch TkMiddle superimposed layer ck's
First input is current other branch TkMiddle cutting layer bkOutput image, second input is branch T by upper levelk-1On adopt
Sample layer dk-1Output image；Side branch TkMiddle superimposed layer ckTwo input pictures are added up, and will it is cumulative after image output；
S365, every Tiao Pang branch TkMiddle up-sampling layer dkTo the superimposed layer c in current other branchkOutput image execute deconvolution
Operation, and the image output after deconvolution is operated；
S366, by branch T by afterbodykIn up-sampling layer dkThe image of output is input in the cutting layer W in main split S,
The input picture is cut into size identical with the image of input layer I input in main split S by the cutting layer W in main split S,
Output of the image as entire multiple-limb network integration model after cutting simultaneously, obtains classification score image.
2. the method for remote sensing image classification according to claim 1, which is characterized in that the main split S is successively
Include an input layer I, at least one convolutional layer { ni, i=1 ..., N, N >=1 }, at least one excitation layer { lj, j=1 ...,
L, L >=1 }, at least one dimensionality reduction layer { mp, p=1 ..., M, M >=1 }, at least one up-sampling layer { uq, q=1 ..., U, U >=
1 }, a cutting layer W, convolutional layer and dimensionality reduction layer cross arrangement as N > 1 or M > 1, wherein N indicates the number of plies of convolutional layer, M table
Show the number of plies of dimensionality reduction layer, L indicates the number of plies of excitation layer, and U indicates the number of plies of up-sampling layer.
3. the method for remote sensing image classification according to claim 2, which is characterized in that described to swash in main split S
Encourage a layer ljIn, each excitation layer ljOne layer of front is all a convolutional layer；The last one convolutional layer n in main split SNConvolution
Core number is the classification number C.
4. the method for remote sensing image classification according to claim 1, which is characterized in that every Tiao Pang branch Tk
It successively include a convolutional layer ak, one cut layer bk, a superimposed layer ck, a up-sampling layer dk；Other branch { Tk, k=
1 ..., K, K >=1 } in all convolutional layer { ak, k=1 ..., K, K >=1 } convolution kernel number be all the classification number C.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710213577.4A CN107122796B (en) | 2017-04-01 | 2017-04-01 | A kind of remote sensing image classification method based on multiple-limb network integration model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710213577.4A CN107122796B (en) | 2017-04-01 | 2017-04-01 | A kind of remote sensing image classification method based on multiple-limb network integration model |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107122796A CN107122796A (en) | 2017-09-01 |
| CN107122796B true CN107122796B (en) | 2019-07-12 |
Family
ID=59724599
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710213577.4A Active CN107122796B (en) | 2017-04-01 | 2017-04-01 | A kind of remote sensing image classification method based on multiple-limb network integration model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107122796B (en) |
Families Citing this family (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108875751B (en) * | 2017-11-02 | 2021-03-02 | 北京旷视科技有限公司 | Image processing method and device, neural network training method and storage medium |
| CN107993217B (en) * | 2017-12-22 | 2021-04-09 | 北京奇虎科技有限公司 | Video data real-time processing method and device and computing equipment |
| CN108664994A (en) * | 2018-04-17 | 2018-10-16 | 哈尔滨工业大学深圳研究生院 | A kind of remote sensing image processing model construction system and method |
| CN109146885B (en) * | 2018-08-17 | 2021-08-17 | 深圳蓝胖子机器智能有限公司 | Image segmentation method, apparatus, and computer-readable storage medium |
| CN109523569B (en) * | 2018-10-18 | 2020-01-31 | 中国科学院空间应用工程与技术中心 | optical remote sensing image segmentation method and device based on multi-granularity network fusion |
| CN109544554B (en) * | 2018-10-18 | 2020-01-31 | 中国科学院空间应用工程与技术中心 | plant image segmentation and leaf skeleton extraction method and system |
| CN109558806B (en) * | 2018-11-07 | 2021-09-14 | 北京科技大学 | Method for detecting high-resolution remote sensing image change |
| CN109902584B (en) * | 2019-01-28 | 2022-02-22 | 深圳大学 | Mask defect identification method, device, equipment and storage medium |
| CN111667045A (en) * | 2019-03-08 | 2020-09-15 | 北京奇虎科技有限公司 | Multi-channel neural network model training method and device and computer storage medium |
| CN110298396B (en) * | 2019-06-25 | 2022-02-08 | 北京工业大学 | Hyperspectral image classification method based on deep learning multi-feature fusion |
| CN110472668B (en) * | 2019-07-22 | 2021-02-19 | 华北电力大学（保定） | Image classification method |
| CN110659726B (en) * | 2019-09-24 | 2022-05-06 | 北京达佳互联信息技术有限公司 | Image processing method and device, electronic equipment and storage medium |
| CN111833273B (en) * | 2020-07-17 | 2021-08-13 | 华东师范大学 | Semantic boundary enhancement method based on long-distance dependence |
| CN112560706B (en) * | 2020-12-18 | 2022-03-29 | 广东电网有限责任公司电力科学研究院 | Method and device for identifying water body target of multi-source satellite image |
| CN113554655B (en) * | 2021-07-13 | 2021-12-31 | 中国科学院空间应用工程与技术中心 | Optical remote sensing image segmentation method and device based on multi-feature enhancement |
| CN114419059B (en) * | 2021-11-24 | 2024-08-13 | 哈尔滨工程大学 | Vision-based embryo-retaining rice embryo retaining degree detection method |
| CN115169561A (en) * | 2022-05-16 | 2022-10-11 | 清华大学 | Multi-branch network collaborative reasoning method and system for Internet of things |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104077613A (en) * | 2014-07-16 | 2014-10-01 | 电子科技大学 | Crowd density estimation method based on cascaded multilevel convolution neural network |
| CN105701508A (en) * | 2016-01-12 | 2016-06-22 | 西安交通大学 | Global-local optimization model based on multistage convolution neural network and significant detection algorithm |
| CN105912990A (en) * | 2016-04-05 | 2016-08-31 | 深圳先进技术研究院 | Face detection method and face detection device |
-
2017
- 2017-04-01 CN CN201710213577.4A patent/CN107122796B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104077613A (en) * | 2014-07-16 | 2014-10-01 | 电子科技大学 | Crowd density estimation method based on cascaded multilevel convolution neural network |
| CN105701508A (en) * | 2016-01-12 | 2016-06-22 | 西安交通大学 | Global-local optimization model based on multistage convolution neural network and significant detection algorithm |
| CN105912990A (en) * | 2016-04-05 | 2016-08-31 | 深圳先进技术研究院 | Face detection method and face detection device |
Non-Patent Citations (3)
| Title |
|---|
| Multi-Scale Convolutional Neural Networks for Time Series Classification;Zhicheng Cui 等;《arXiv:1603.06995v4 [cs.CV]》;20160511;1-10 * |
| 基于多分支BP网络模型的车型分类研究;林培群 等;《可持续发展的中国交通——2005全国博士生学术论坛（交通运输工程学科）论文集（下册）》;20050701;1643-1648 * |
| 基于多分支卷积神经网络的车辆图像比对方法;蔡晓东 等;《电视技术》;20161117;第40卷(第11期);116－120 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107122796A (en) | 2017-09-01 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107122796B (en) | A kind of remote sensing image classification method based on multiple-limb network integration model | |
| CN108154194B (en) | Method for extracting high-dimensional features by using tensor-based convolutional network | |
| CN111079795B (en) | Image classification method based on CNN (content-centric networking) fragment multi-scale feature fusion | |
| CN107220657A (en) | A kind of method of high-resolution remote sensing image scene classification towards small data set | |
| CN105005774B (en) | A kind of recognition methods of face kinship and device based on convolutional neural networks | |
| CN106022300B (en) | Traffic sign recognition method and system based on cascade deep study | |
| CN107943967A (en) | Algorithm of documents categorization based on multi-angle convolutional neural networks and Recognition with Recurrent Neural Network | |
| CN110298361A (en) | A kind of semantic segmentation method and system of RGB-D image | |
| CN109543502A (en) | A kind of semantic segmentation method based on the multiple dimensioned neural network of depth | |
| CN109255334A (en) | Remote sensing image terrain classification method based on deep learning semantic segmentation network | |
| CN112132197B (en) | Model training, image processing method, device, computer equipment and storage medium | |
| CN107862261A (en) | Image people counting method based on multiple dimensioned convolutional neural networks | |
| CN106845529A (en) | Image feature recognition methods based on many visual field convolutional neural networks | |
| CN105678216A (en) | Spatio-temporal data stream video behavior recognition method based on deep learning | |
| CN109359559A (en) | A kind of recognition methods again of the pedestrian based on dynamic barriers sample | |
| CN108986058A (en) | The image interfusion method of lightness Consistency Learning | |
| CN110046550A (en) | Pedestrian's Attribute Recognition system and method based on multilayer feature study | |
| CN108009637B (en) | Station caption segmentation method of pixel-level station caption identification network based on cross-layer feature extraction | |
| CN107958067A (en) | It is a kind of based on without mark Automatic Feature Extraction extensive electric business picture retrieval system | |
| CN113034506B (en) | Remote sensing image semantic segmentation method and device, computer equipment and storage medium | |
| CN111368734B (en) | Micro expression recognition method based on normal expression assistance | |
| CN110414483A (en) | A kind of face identification method and system based on deep neural network and random forest | |
| CN109544554A (en) | A kind of segmentation of plant image and blade framework extracting method and system | |
| CN108388904B (en) | Dimensionality reduction method based on convolutional neural network and covariance tensor matrix | |
| CN104156464A (en) | Micro-video retrieval method and device based on micro-video feature database |
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 |