CN109785344A - The remote sensing image segmentation method of binary channel residual error network based on feature recalibration - Google Patents
The remote sensing image segmentation method of binary channel residual error network based on feature recalibration Download PDFInfo
- Publication number
- CN109785344A CN109785344A CN201910055677.8A CN201910055677A CN109785344A CN 109785344 A CN109785344 A CN 109785344A CN 201910055677 A CN201910055677 A CN 201910055677A CN 109785344 A CN109785344 A CN 109785344A
- Authority
- CN
- China
- Prior art keywords
- remote sensing
- residual error
- binary channel
- data
- error 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
Abstract
The remote sensing image segmentation method of the invention discloses a kind of binary channel residual error network based on feature recalibration comprising: remote sensing images to be detected S1, are obtained, and it is normalized to obtain normalized image;S2, random window sampling is carried out to normalized image, obtains several sample datas being sized;And S3, using the binary channel residual error network remote sensing parted pattern pre-established several sample datas are split, the remote sensing images after being divided;The training process of binary channel residual error network remote sensing parted pattern are as follows: binary channel residual error network is trained according to known remote sensing images to obtain binary channel residual error network remote sensing parted pattern.
Description
Technical field
This programme is related to remote sensing images recognition methods, and in particular to a kind of binary channel residual error network based on feature recalibration
Remote sensing image segmentation method.
Background technique
Currently, the image partition method based on neural network model is quickly grown, trained neural network point can be used
New image data is cut, it will be more excellent than traditional dividing method on segmentation precision and efficiency.With the hair of deep learning
Exhibition, a large amount of neural network models continuously emerge, and simple convolutional layer cannot be guaranteed the convergence rate and image segmentation of network
Precision.Existing full convolutional neural networks use end-to-end, pixel to pixel structure, avoid due to using block of pixels and band
The repetition storage that comes and the problem of calculate convolution, but the network is not sensitive enough to the details of image, do not account for pixel it
Between relationship, the result because obtained from is fine not enough.
The output that depth residual error network has broken n-1 layers of traditional neural network can only be to n-layer thought as input, solution
Determined increase depth bring degenerate problem, network can become deeper, but residual error structure do not consider feature channel it
Between correlation, characteristic information cannot maximize the use, in Remote Sensing Image Segmentation, have enough network depths,
It is easy optimization, but has ignored the correlation of feature space, network performance also needs further to improve.
Summary of the invention
For above-mentioned deficiency in the prior art, the binary channel residual error network provided by the invention based on feature recalibration
Remote sensing image segmentation method solves the problems, such as that existing Remote Sensing Image Segmentation precision is low.
In order to achieve the above object of the invention, the technical solution adopted by the present invention are as follows:
A kind of remote sensing image segmentation method of binary channel residual error network based on feature recalibration is provided comprising:
S1, remote sensing images to be detected are obtained, and it is normalized to obtain normalized image;
S2, random window sampling is carried out to normalized image, obtains several sample datas being sized;And
S3, several sample datas are split using the binary channel residual error network remote sensing parted pattern pre-established, are obtained
Remote sensing images after to segmentation;
The training process of binary channel residual error network remote sensing parted pattern are as follows:
Binary channel residual error network is trained according to known remote sensing images to obtain binary channel residual error network remote sensing segmentation
Model.
Further, binary channel residual error network includes average pond layer and at least one image processing module, the last one
Image processing module treated image inputs the remote sensing images after average pond layer is divided;
Image processing module includes the binary channel SE module that feature recalibration is carried out to the sample data or characteristic pattern of input,
The convolutional layer of convolution operation is carried out to the characteristic pattern after binary channel SE resume module and to input sample data or characteristic pattern and volume
Lamination treated convolution characteristic pattern carries out the superimposed layer of characteristic value superposition;
When there are at least two image processing modules, the upper superimposed layer in two adjacent images processing module is under
One binary channel SE module connection.
Further, binary channel SE module carries out the method packet of feature recalibration to the sample data or characteristic pattern of input
It includes:
A1, BN normalization operation, the operation of ReLU activation primitive, 3*3 are successively carried out to the sample data or characteristic pattern of input
Convolution operation and the operation of random drop neural network unit, obtain fisrt feature figure;
A2, C1 data and C2 data are divided into according to the channel of fisrt feature figure, while successively by fisrt feature figure
BN normalization operation, the operation of ReLU activation primitive, 3*3 convolution operation and the operation of random drop neural network unit are carried out, is obtained
Second feature figure;
A3, D1 data and D2 data are divided into according to the channel of second feature figure, using binary channel module to C1 number
According to being added to obtain third feature figure with the characteristic value of D1 data and stack to the characteristic value of C2 data and D2 data
To fourth feature figure;
A4, third feature figure and fourth feature figure are stacked to obtain fifth feature figure, and successively by fifth feature figure
The global average pond layer of input, full articulamentum, ReLU activation primitive layer, full articulamentum and sigmoid activation primitive layer, obtain every
The weighted value in a channel;And
A5, the feature by each channel of fifth feature figure multiplied by the weighted value of corresponding channel, after obtaining feature recalibration
Figure.
Further, binary channel residual error network is trained according to known remote sensing images to obtain binary channel residual error network
The method of remote sensing parted pattern includes:
Remote sensing image known to B1, acquisition, and it is divided into training set and test set, it is pre-processed later;
B2, training set data is divided into several batches, and initializes convolution kernel weight and bias to cross entropy loss function
Derivative is 0:
ΔW(l)=0, Δ b(l)=0
Wherein, Δ W(l)It is the convolution kernel weight in l layers of convolutional layer to the derivative of cross entropy loss function;Δb(l)It is
The derivative of convolution kernel bias in l layers of convolutional layer to cross entropy loss function;
B3, wherein a collection of training set data binary channel residual error network will be inputted, and by training set data and binary channel residual error
All layers of each node parameter is calculated in network, realizes the propagated forward of network training, exports reality output data;
B4, the error between reality output data and training set label is calculated using cross entropy loss function, found best
Convolution kernel weight and bias make the value of cross entropy loss function minimum;
B5, judge whether to be trained all training set datas input binary channel residual error network, if then completing one
Secondary iteration enters step B6, otherwise return step B2;
B6, current iteration is judged by convolution kernel weight and bias toward the update of negative gradient direction using gradient descent method later
Whether number reaches default iteration threshold, if so, entering step B7, otherwise enters step B3;
B7, output binary channel residual error network remote sensing parted pattern.
Further, the expression formula of the propagated forward of network training is realized are as follows:
z(l+1)=w(l+1)a(t)+b(t+1), a(l+1)=f (z(l+1))
Wherein, l=1,2 ... l-1, l are the number of plies of binary channel residual error network;z(l+1)For l+1 layers of output;F is ReLU
Activation primitive;a(l+1)、a(l)Respectively l+1, l layers of reality output data;b(l+1)For the bias of l+1 layers of convolutional layer.
Further, the expression formula of cross entropy loss function are as follows:
Wherein, L is the error in reality output data and training set data between label;M is the sample of training set data
Number;X is the network unit quantity in binary channel residual error network, and y is label data;A is that the reality of binary channel residual error network is defeated
Data out.
Further, using gradient descent method, the expression formula that convolution kernel weight and bias are updated toward negative gradient direction are as follows:
Wherein, α is learning rate;w(l)' be l layers of updated convolution kernel weighted value;It is weighed for l layers of convolution kernel
Again to the partial derivative of loss function;b(l)' it is l layers of updated convolution kernel bias;b(l)For l layers of convolution kernel bias;It is l layers of convolution kernel bias to the partial derivative of loss function.
Further, obtain known remote sensing image, and be divided into training set and test set, pre-processed later into
One step includes:
Remote sensing images known to B11, acquisition, and it is divided into training set and test set;
B12, known remote sensing images in training set and test set are normalized, to map that [0,1] area
Between;
B13, random window sampling is carried out to data normalized in training set, obtained several having a size of the small of 256*256
Sample data;And
B14, data enhancement operations are carried out to Small Sample Database.
Further, carrying out data enhancement operations to Small Sample Database includes rotation transformation, turning-over changed, gamma transformation
And/or change of scale;
Wherein the rotation angle of rotation transformation is 90 °, 180 ° or 270 °;Turning-over changed is to turn over along x-axis or y-axis direction
Turn;Gamma transformation are as follows: as gamma > 1, brighter area grayscale is stretched, and darker area grayscale is compressed, so that image
It is whole dimmed;As gamma < 1, brighter area grayscale is compressed, and darker area grayscale is stretched, so that image integrally becomes
It is bright;
Change of scale are as follows: image is zoomed in or out according to the scale factor of setting, or is mentioned referring to SIFT feature
Thought is taken, using specified scale factor to image filtering tectonic scale space, changes the size or fog-level of picture material.
The invention has the benefit that this programme used binary channel residual error network when carrying out remote sensing images identification exists
It helps to capture spatial coherence, the relation of interdependence between analog channel by being explicitly embedded in study mechanism in training process
The characteristic response in channel is readjusted, gets the significance level in each feature channel automatically, then according to this important journey
Degree goes to promote useful feature and inhibits the feature little to current task use;In combination with binary channel mechanism, more effectively
Using the characteristic information of upper level, it is made to have higher precision and faster convergence rate on Remote Sensing Image Segmentation.
The binary channel residual error network remote sensing parted pattern that training is formed first passes around binary channel and is reusing the same of primitive character
When generate new feature, then using the dependence between channel, weight assessment is carried out to channel, has redescribed entire spy
The information for levying figure, is finally mapped on an equal basis using the residual error network architecture, under the premise of not increasing computation complexity, is not only mentioned
The high convergence rate of network training, also improves the training precision of network.
Detailed description of the invention
Fig. 1 is the flow chart of the remote sensing image segmentation method of the binary channel residual error network based on feature recalibration.
Fig. 2 is the network architecture of binary channel residual error network.
Fig. 3 is the architecture diagram of the binary channel SE module in binary channel residual error network.
Fig. 4 is the number of iterations and training precision and the curve graph of training loss in binary channel residual error network training process,
In (a) be the number of iterations and training precision curved line relation, (b) for the number of iterations and training loss curved line relation.
Fig. 5 is the segmentation result that three width remote sensing images are input to trained binary channel residual error network remote sensing parted pattern;
Wherein first it is classified as initial data, secondary series is the segmentation result of U-Net model, and third column are the segmentation results of this programme.
Specific embodiment
A specific embodiment of the invention is described below, in order to facilitate understanding by those skilled in the art this hair
It is bright, it should be apparent that the present invention is not limited to the ranges of specific embodiment, for those skilled in the art,
As long as various change is in the spirit and scope of the present invention that the attached claims limit and determine, these variations are aobvious and easy
See, all are using the innovation and creation of present inventive concept in the column of protection.
The stream of the remote sensing image segmentation method of the binary channel residual error network based on feature recalibration is shown with reference to Fig. 1, Fig. 1
Cheng Tu;As shown in Figure 1, remote sensing image segmentation method S includes S1 to S3.
In step sl, remote sensing images to be detected are obtained, and it is normalized to obtain normalized image, this
The main purpose for locating normalized is that its image data is mapped to [0,1] section.
In step s 2, random window sampling is carried out to normalized image, obtains several sample datas being sized, sample
The size of notebook data is the Small Sample Database of 256*256.And
In step s3, several sample datas are carried out using the binary channel residual error network remote sensing parted pattern pre-established
Segmentation, the remote sensing images after being divided;
It is illustrated below with reference to effect of the Fig. 5 to the remote sensing image segmentation method of this programme:
The remote sensing images (with reference to first row in Fig. 5) of the secondary diverse geographic locations of input three are to trained U-Net model
Image segmentation is carried out with the binary channel residual error network remote sensing parted pattern of this programme, the segmentation result of U-Net model is with reference in Fig. 5
Secondary series, the segmentation result of binary channel residual error network remote sensing parted pattern compare secondary series and third column with reference to third column in Fig. 5
Segmentation result it can be concluded that, the segmentation result of this method is substantially better than the U-Net model of the prior art.
Wherein, the training process of binary channel residual error network remote sensing parted pattern are as follows:
Binary channel residual error network is trained according to known remote sensing images to obtain binary channel residual error network remote sensing segmentation
Model.
As shown in Fig. 2, the preferred binary channel residual error network of this programme includes average pond layer and at least one image when implementing
Processing module, the last one image processing module treated image input the remote sensing images after average pond layer is divided;
Image processing module includes the binary channel SE module that feature recalibration is carried out to the sample data or characteristic pattern of input,
The convolutional layer of convolution operation is carried out to the characteristic pattern after binary channel SE resume module and to input sample data or characteristic pattern and volume
Lamination treated convolution characteristic pattern carries out the superimposed layer of characteristic value superposition;
When there are at least two image processing modules, the upper superimposed layer in two adjacent images processing module is under
One binary channel SE module connection.
The number of plies of network is deeper, and the characteristic information that can be extracted is abundanter, still, simply increases network depth,
It will lead to gradient disperse or gradient explosion, while the performance of network can also degenerate.This programme uses binary channel residual error network structure
The same mapping carried out between upper and lower level solves degenerate problem while increasing network depth, and the performance of network also obtains
It is promoted.
If image processing module is located at first, the binary channel SE module for inputting the image processing module is sample number
According to, positioned at first image processing module binary channel SE module below it is received be all the output of a upper superimposed layer feature
Figure.
As shown in figure 3, in one embodiment of the invention, sample data or characteristic pattern of the binary channel SE module to input
The method A for carrying out feature recalibration includes step A1 to step A5:
In step A1, sample data or characteristic pattern to input successively carry out BN normalization operation, ReLU activation primitive
Operation, 3*3 convolution operation and the operation of random drop neural network unit, obtain fisrt feature figure.
Wherein, random drop neural network unit (dropout) is indicated according to certain probability at random by binary channel residual error net
Neural network unit in network is temporarily abandoned from network, prevents over-fitting.
In step A2, C1 data and C2 data are divided into according to the channel of fisrt feature figure, while special by first
Sign figure successively carries out BN normalization operation, the operation of ReLU activation primitive, 3*3 convolution operation and random drop neural network unit behaviour
Make, obtains second feature figure.
In step A3, D1 data and D2 data are divided into according to the channel of second feature figure, using binary channel mould
Block C1 data are added obtain with the characteristic value of D1 data third feature figure and to the characteristic values of C2 data and D2 data into
Row stacks and obtains fourth feature figure.
Binary channel module includes summation access and concatenate access, and summation access is for realizing C1
Data are added with the characteristic value of D1 data, concatenate access for realizing C2 data and D2 data stacking (size phase
Add).
In step A4, third feature figure and fourth feature figure are stacked to obtain fifth feature figure, and special by the 5th
Sign figure sequentially inputs global average pond layer, full articulamentum, ReLU activation primitive layer, full articulamentum and sigmoid activation primitive
Layer, obtains the weighted value in each channel.
Global draw pond layer shows the numeric distribution situation of C characteristic pattern, and the effect of two full articulamentums is fusion
The profile information in each channel, and reduce channel number to reduce calculation amount.
In step A5, by each channel of fifth feature figure multiplied by the weighted value of corresponding channel, feature recalibration is obtained
Characteristic pattern afterwards.
The relation of interdependence of binary channel SE module simulation feature interchannel readjusts the characteristic response in channel, from
Then the dynamic significance level for getting each feature channel is gone to be promoted useful feature according to this significance level and is inhibited to working as
The little feature of preceding task use;In combination with binary channel mechanism, new feature is generated while reusing primitive character, is more had
Effect ground utilizes the characteristic information of upper level, it is made to have higher precision and faster convergence rate on Remote Sensing Image Segmentation.
In one embodiment of the invention, binary channel residual error network is trained to obtain according to known remote sensing images
The method B of binary channel residual error network remote sensing parted pattern includes step B1 to B7:
In step bl is determined, known remote sensing image is obtained, and is divided into training set and test set, is located in advance later
Reason;
Then the concrete methods of realizing of B1 step is described in detail below:
Remote sensing images known to B11, acquisition, and it is divided into training set and test set, wherein in training set and test set
Remote sensing images ratio be 8:1.
B12, known remote sensing images in training set and test set are normalized, to map that [0,1] area
Between;After this programme can also be normalized remote sensing images during realization, it is being trained collection and test set
Division, after can also first dividing training set and test set, then be normalized, no matter first carry out which step operation,
Final image processing effect is not influenced.
B13, random window sampling is carried out to data normalized in training set, obtained several having a size of the small of 256*256
Sample data, it is to be adapted to binary channel residual error network remote sensing that the main purpose of random window sampling, which is by the picture segmentation of training set,
Picture format and the size requirement of parted pattern.
B14, data enhancement operations are carried out to Small Sample Database, carrying out data enhancement operations to Small Sample Database includes rotation
Transformation, turning-over changed, gamma transformation and/or change of scale;
Wherein the rotation angle of rotation transformation is 90 °, 180 ° or 270 °;Turning-over changed is to turn over along x-axis or y-axis direction
Turn;Gamma transformation are as follows: as gamma > 1, brighter area grayscale is stretched, and darker area grayscale is compressed, so that image
It is whole dimmed;As gamma < 1, brighter area grayscale is compressed, and darker area grayscale is stretched, so that image integrally becomes
It is bright;
Change of scale are as follows: image is zoomed in or out according to the scale factor of setting, or is mentioned referring to SIFT feature
Thought is taken, using specified scale factor to image filtering tectonic scale space, changes the size or fog-level of picture material.
In step B2, training set data is divided into several batches, and initializes convolution kernel weight and convolution kernel bias to friendship
The derivative for pitching entropy loss function is 0:
ΔW(l)=0, Δ b(l)=0
Wherein, Δ W(l)It is the convolution kernel weight in l layers of convolutional layer to the derivative of cross entropy loss function;Δb(l)It is
The derivative of convolution kernel bias in l layers of convolutional layer to cross entropy loss function;
Wherein, the expression formula of cross entropy loss function are as follows:
Wherein, L is the error in reality output data and training set data between label;M is the sample of training set data
Number;X is the network unit quantity in binary channel residual error network, and y is label data;A is that the reality of binary channel residual error network is defeated
Data out.
In step B3, wherein a collection of training set data binary channel residual error network will be inputted, and by training set data and double
Each node parameter of all layers (all layers of each layer referred in whole network structure) is calculated in access residual error network,
It realizes the propagated forward of network training, exports reality output data.
In the calculating of binary channel residual error network, need to calculate the spy of the forward-propagating output according to binary channel residual error network
The score P1 that sign figure calculates, and according to the gap between the score P2 of true tag calculating, cross entropy loss function Loss is calculated,
Backpropagation could be applied.
Wherein, the expression formula of the propagated forward of network training is realized are as follows:
z(l+1)=w(l+1)a(l)+b(l+1), a(l+1)=f (z(l+1))
Wherein, l=1,2 ... l-1, l are the number of plies of binary channel residual error network;z(l+1)For l+1 layers of output;F is ReLU
Activation primitive;a(l+1)、a(l)Respectively l+1, l layers of reality output data;b(l+1)For the bias of l+1 layers of convolutional layer.
In step B4, the error between reality output data and training set label is calculated using cross entropy loss function,
It finds optimal convolution kernel weight and bias makes the value of cross entropy loss function minimum;
In step B5, judge whether to be trained all training set datas input binary channel residual error network, if
An iteration is then completed, enters step B6, otherwise return step B2;
In step B6, using gradient descent method, convolution kernel weight and bias is updated toward negative gradient direction, judged later
Whether current iteration number reaches default iteration threshold, if so, entering step B7, otherwise enters step B3;
Wherein, using gradient descent method, the expression formula that convolution kernel weight and bias are updated toward negative gradient direction are as follows:
Wherein, α is learning rate;w(l)' be l layers of updated convolution kernel weighted value;It is weighed for l layers of convolution kernel
Again to the partial derivative of loss function;b(l)' it is l layers of updated convolution kernel bias;b(l)For l layers of convolution kernel bias;It is l layers of convolution kernel bias to the partial derivative of loss function.
In step B7, binary channel residual error network remote sensing parted pattern is exported.
During obtaining binary channel residual error network remote sensing parted pattern, the number of iterations (abscissa) will affect training precision
It is lost with training, can specifically refer to Fig. 4, from (a) in Fig. 4 as can be seen that the number of iterations is more, training precision is higher,
Otherwise training loss just it is smaller, referring to fig. 4 in (b).
Claims (9)
1. the remote sensing image segmentation method of the binary channel residual error network based on feature recalibration characterized by comprising
S1, remote sensing images to be detected are obtained, and it is normalized to obtain normalized image;
S2, random window sampling is carried out to normalized image, obtains several sample datas being sized;And
S3, several sample datas are split using the binary channel residual error network remote sensing parted pattern pre-established, are divided
Remote sensing images after cutting;
The training process of the binary channel residual error network remote sensing parted pattern are as follows:
Binary channel residual error network is trained according to known remote sensing images to obtain binary channel residual error network remote sensing parted pattern.
2. the remote sensing image segmentation method of the binary channel residual error network according to claim 1 based on feature recalibration,
It is characterized in that, the binary channel residual error network includes average pond layer and at least one image processing module, the last one image
Processing module treated image inputs the remote sensing images after average pond layer is divided;
Described image processing module includes that the binary channel SE module of feature recalibration is carried out to the sample data or characteristic pattern of input,
The convolutional layer of convolution operation is carried out to the characteristic pattern after binary channel SE resume module and to the sample data or characteristic pattern of input and
Convolutional layer treated convolution characteristic pattern carries out the superimposed layer of characteristic value superposition;
When there are at least two image processing modules, a upper superimposed layer in two adjacent images processing module with it is next
The connection of binary channel SE module.
3. the remote sensing image segmentation method of the binary channel residual error network according to claim 2 based on feature recalibration,
It is characterized in that, the method that the binary channel SE module carries out feature recalibration to the sample data or characteristic pattern of input includes:
A1, BN normalization operation, the operation of ReLU activation primitive, 3*3 convolution are successively carried out to the sample data or characteristic pattern of input
Operation and the operation of random drop neural network unit, obtain fisrt feature figure;
A2, it is divided into C1 data and C2 data according to the channel of fisrt feature figure, while fisrt feature figure is successively carried out
BN normalization operation, the operation of ReLU activation primitive, 3*3 convolution operation and the operation of random drop neural network unit, obtain second
Characteristic pattern;
A3, D1 data and D2 data are divided into according to the channel of second feature figure, using binary channel module to C1 data with
The characteristic value of D1 data is added to obtain third feature figure and is stacked to obtain the to the characteristic value of C2 data and D2 data
Four characteristic patterns;
A4, third feature figure and fourth feature figure are stacked to obtain fifth feature figure, and fifth feature figure is sequentially input
The average pond layer of the overall situation, full articulamentum, ReLU activation primitive layer, full articulamentum and sigmoid activation primitive layer obtain each logical
The weighted value in road;And
A5, the characteristic pattern by each channel of fifth feature figure multiplied by the weighted value of corresponding channel, after obtaining feature recalibration.
4. the remote sensing image segmentation method of the binary channel residual error network according to claim 3 based on feature recalibration,
It is characterized in that, the remote sensing images according to known to are trained binary channel residual error network to obtain binary channel residual error network remote sensing
The method of parted pattern includes:
Remote sensing image known to B1, acquisition, and it is divided into training set and test set, it is pre-processed later;
B2, training set data is divided into several batches, and initializes convolution kernel weight and bias to the derivative of cross entropy loss function
It is 0:
ΔW(l)=0, Δ b(l)=0
Wherein, Δ W(l)It is the convolution kernel weight in l layers of convolutional layer to the derivative of cross entropy loss function;Δb(l)It is l layers
The derivative of convolution kernel bias in convolutional layer to cross entropy loss function;
B3, wherein a collection of training set data binary channel residual error network will be inputted, and by training set data and binary channel residual error network
In all layers of each node parameter calculated, realize the propagated forward of network training, export reality output data;
B4, the error between reality output data and training set label is calculated using cross entropy loss function, finds optimal volume
Product core weight and bias make the value of cross entropy loss function minimum;
B5, judge whether to be trained all training set datas input binary channel residual error network, if then completing primary change
In generation, enters step B6, otherwise return step B2;
B6, current iteration number is judged by convolution kernel weight and bias toward the update of negative gradient direction using gradient descent method later
Whether reach default iteration threshold, if so, entering step B7, otherwise enters step B3;
B7, output binary channel residual error network remote sensing parted pattern.
5. the remote sensing image segmentation method of the binary channel residual error network according to claim 4 based on feature recalibration,
It is characterized in that, the expression formula of the propagated forward for realizing network training are as follows:
z(l+1)=w(l+1)a(l)+b(l+1), a(l+1)=f (z(l+1))
Wherein, l=1,2 ... l-1, l are the number of plies of binary channel residual error network;z(l+1)For l+1 layers of output;F is ReLU activation
Function;a(l+1)、a(l)Respectively l+1, l layers of reality output data;b(l+1)For the bias of l+1 layers of convolutional layer.
6. the remote sensing image segmentation method of the binary channel residual error network according to claim 4 based on feature recalibration,
It is characterized in that, the expression formula of the cross entropy loss function are as follows:
Wherein, L is the error in reality output data and training set data between label;M is the number of samples of training set data;
X is the network unit quantity in binary channel residual error network, and y is label data;A is the reality output number of binary channel residual error network
According to.
7. the remote sensing image segmentation method of the binary channel residual error network according to claim 6 based on feature recalibration,
It is characterized in that, it is described to use gradient descent method, the expression formula that convolution kernel weight and bias are updated toward negative gradient direction are as follows:
Wherein, α is learning rate;w(l)′For the weighted value of l layers of updated convolution kernel;It is l layers of convolution kernel weight to damage
Lose the partial derivative of function;b(l)′For l layers of updated convolution kernel bias;b(l)For l layers of convolution kernel bias;For l
Partial derivative of the convolution kernel bias of layer to loss function.
8. the remote sensing image segmentation method of the binary channel residual error network according to claim 4 based on feature recalibration,
It is characterized in that, remote sensing image known to the acquisition, and is divided into training set and test set, carry out pre-processing further later
Include:
Remote sensing images known to B11, acquisition, and it is divided into training set and test set;
B12, known remote sensing images in training set and test set are normalized, to map that [0,1] section;
B13, random window sampling is carried out to data normalized in training set, obtains several small samples having a size of 256*256
Data;And
B14, data enhancement operations are carried out to Small Sample Database.
9. the remote sensing image segmentation method of the binary channel residual error network according to claim 8 based on feature recalibration,
It is characterized in that, carrying out data enhancement operations to Small Sample Database includes rotation transformation, turning-over changed, gamma transformation and/or scale
Transformation;
Wherein the rotation angle of rotation transformation is 90 °, 180 ° or 270 °;Turning-over changed is to overturn along x-axis or y-axis direction;Gal
Horse transformation are as follows: as gamma > 1, brighter area grayscale is stretched, and darker area grayscale is compressed, so that image integrally becomes
Secretly;As gamma < 1, brighter area grayscale is compressed, and darker area grayscale is stretched, so that image integrally brightens;
Change of scale are as follows: image is zoomed in or out according to the scale factor of setting, or extracts and thinks referring to SIFT feature
Think, using specified scale factor to image filtering tectonic scale space, changes the size or fog-level of picture material.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910055677.8A CN109785344A (en) | 2019-01-22 | 2019-01-22 | The remote sensing image segmentation method of binary channel residual error network based on feature recalibration |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910055677.8A CN109785344A (en) | 2019-01-22 | 2019-01-22 | The remote sensing image segmentation method of binary channel residual error network based on feature recalibration |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109785344A true CN109785344A (en) | 2019-05-21 |
Family
ID=66501803
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910055677.8A Pending CN109785344A (en) | 2019-01-22 | 2019-01-22 | The remote sensing image segmentation method of binary channel residual error network based on feature recalibration |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109785344A (en) |
Cited By (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110211137A (en) * | 2019-06-08 | 2019-09-06 | 西安电子科技大学 | Satellite Image Segmentation method based on residual error network and U-Net segmentation network |
CN110309879A (en) * | 2019-07-01 | 2019-10-08 | 天津工业大学 | A kind of jujube exterior quality classification method, device, equipment and storage medium |
CN110472668A (en) * | 2019-07-22 | 2019-11-19 | 华北电力大学(保定) | A kind of image classification method |
CN110781756A (en) * | 2019-09-29 | 2020-02-11 | 北京化工大学 | Urban road extraction method and device based on remote sensing image |
CN110796115A (en) * | 2019-11-08 | 2020-02-14 | 厦门美图之家科技有限公司 | Image detection method and device, electronic equipment and readable storage medium |
CN110807376A (en) * | 2019-10-17 | 2020-02-18 | 北京化工大学 | Method and device for extracting urban road based on remote sensing image |
CN110991568A (en) * | 2020-03-02 | 2020-04-10 | 佳都新太科技股份有限公司 | Target identification method, device, equipment and storage medium |
CN111220786A (en) * | 2020-03-09 | 2020-06-02 | 生态环境部华南环境科学研究所 | Method for rapidly monitoring organic pollution of deep water sediments |
CN111242028A (en) * | 2020-01-13 | 2020-06-05 | 北京工业大学 | Remote sensing image ground object segmentation method based on U-Net |
CN112991354A (en) * | 2021-03-11 | 2021-06-18 | 东北大学 | High-resolution remote sensing image semantic segmentation method based on deep learning |
CN113269778A (en) * | 2021-06-21 | 2021-08-17 | 复旦大学 | Image weak supervision segmentation method based on iteration |
CN114795258A (en) * | 2022-04-18 | 2022-07-29 | 浙江大学 | Child hip joint dysplasia diagnosis system |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108492286A (en) * | 2018-03-13 | 2018-09-04 | 成都大学 | A kind of medical image cutting method based on the U-shaped convolutional neural networks of binary channel |
CN108875587A (en) * | 2018-05-24 | 2018-11-23 | 北京飞搜科技有限公司 | Target distribution detection method and equipment |
-
2019
- 2019-01-22 CN CN201910055677.8A patent/CN109785344A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108492286A (en) * | 2018-03-13 | 2018-09-04 | 成都大学 | A kind of medical image cutting method based on the U-shaped convolutional neural networks of binary channel |
CN108875587A (en) * | 2018-05-24 | 2018-11-23 | 北京飞搜科技有限公司 | Target distribution detection method and equipment |
Non-Patent Citations (6)
Title |
---|
JIE HU 等: "Squeeze-and-Excitation Networks", 《ARXIV》 * |
KAIMING HE 等: "Deep Residual Learning for Image Recognition", 《ARXIV》 * |
SAINING XIE 等: "Aggregated Residual Transformations for Deep Neural Networks", 《ARXIV》 * |
YUNPENG CHEN 等: "Dual Path Networks", 《ARXIV》 * |
司马海峰 等著: "《遥感图像分类中的智能计算方法》", 31 January 2018, 吉林大学出版社 * |
韦鹏程 等著: "《大数据巨量分析与机器学习的整合与开发》", 31 May 2017, 电子科技大学出版社 * |
Cited By (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110211137A (en) * | 2019-06-08 | 2019-09-06 | 西安电子科技大学 | Satellite Image Segmentation method based on residual error network and U-Net segmentation network |
CN110211137B (en) * | 2019-06-08 | 2022-12-06 | 西安电子科技大学 | Satellite image segmentation method based on residual error network and U-Net segmentation network |
CN110309879A (en) * | 2019-07-01 | 2019-10-08 | 天津工业大学 | A kind of jujube exterior quality classification method, device, equipment and storage medium |
CN110472668A (en) * | 2019-07-22 | 2019-11-19 | 华北电力大学(保定) | A kind of image classification method |
CN110781756A (en) * | 2019-09-29 | 2020-02-11 | 北京化工大学 | Urban road extraction method and device based on remote sensing image |
CN110807376A (en) * | 2019-10-17 | 2020-02-18 | 北京化工大学 | Method and device for extracting urban road based on remote sensing image |
CN110796115A (en) * | 2019-11-08 | 2020-02-14 | 厦门美图之家科技有限公司 | Image detection method and device, electronic equipment and readable storage medium |
CN110796115B (en) * | 2019-11-08 | 2022-12-23 | 厦门美图宜肤科技有限公司 | Image detection method and device, electronic equipment and readable storage medium |
CN111242028A (en) * | 2020-01-13 | 2020-06-05 | 北京工业大学 | Remote sensing image ground object segmentation method based on U-Net |
CN110991568A (en) * | 2020-03-02 | 2020-04-10 | 佳都新太科技股份有限公司 | Target identification method, device, equipment and storage medium |
CN111220786A (en) * | 2020-03-09 | 2020-06-02 | 生态环境部华南环境科学研究所 | Method for rapidly monitoring organic pollution of deep water sediments |
CN112991354A (en) * | 2021-03-11 | 2021-06-18 | 东北大学 | High-resolution remote sensing image semantic segmentation method based on deep learning |
CN112991354B (en) * | 2021-03-11 | 2024-02-13 | 东北大学 | High-resolution remote sensing image semantic segmentation method based on deep learning |
CN113269778A (en) * | 2021-06-21 | 2021-08-17 | 复旦大学 | Image weak supervision segmentation method based on iteration |
CN113269778B (en) * | 2021-06-21 | 2022-11-29 | 复旦大学 | Image weak supervision segmentation method based on iteration |
CN114795258A (en) * | 2022-04-18 | 2022-07-29 | 浙江大学 | Child hip joint dysplasia diagnosis system |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109785344A (en) | The remote sensing image segmentation method of binary channel residual error network based on feature recalibration | |
CN103955702B (en) | SAR image terrain classification method based on depth RBF network | |
CN110135267B (en) | Large-scene SAR image fine target detection method | |
CN107169956B (en) | Color woven fabric defect detection method based on convolutional neural network | |
CN109670528B (en) | Data expansion method facing pedestrian re-identification task and based on paired sample random occlusion strategy | |
CN109829399A (en) | A kind of vehicle mounted road scene point cloud automatic classification method based on deep learning | |
CN111199214B (en) | Residual network multispectral image ground object classification method | |
CN109344821A (en) | Small target detecting method based on Fusion Features and deep learning | |
CN110378381A (en) | Object detecting method, device and computer storage medium | |
CN109035260A (en) | A kind of sky areas dividing method, device and convolutional neural networks | |
CN106778835A (en) | The airport target by using remote sensing image recognition methods of fusion scene information and depth characteristic | |
CN108491849A (en) | Hyperspectral image classification method based on three-dimensional dense connection convolutional neural networks | |
CN107016357A (en) | A kind of video pedestrian detection method based on time-domain convolutional neural networks | |
CN105678278A (en) | Scene recognition method based on single-hidden-layer neural network | |
CN107239733A (en) | Continuous hand-written character recognizing method and system | |
CN109215034A (en) | A kind of Weakly supervised image, semantic dividing method for covering pond based on spatial pyramid | |
CN113627472B (en) | Intelligent garden leaf feeding pest identification method based on layered deep learning model | |
CN108121931A (en) | two-dimensional code data processing method, device and mobile terminal | |
CN110490265B (en) | Image steganalysis method based on double-path convolution and feature fusion | |
CN108122003A (en) | A kind of Weak target recognition methods based on deep neural network | |
CN107909084A (en) | A kind of haze concentration prediction method based on convolution linear regression network | |
CN110222760A (en) | A kind of fast image processing method based on winograd algorithm | |
CN107516128A (en) | A kind of flowers recognition methods of the convolutional neural networks based on ReLU activation primitives | |
CN114842208A (en) | Power grid harmful bird species target detection method based on deep learning | |
CN108596818A (en) | A kind of image latent writing analysis method based on multi-task learning convolutional neural networks |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190521 |
|
RJ01 | Rejection of invention patent application after publication |