The content of the invention
The invention provides a kind of image based on deep learning to remove grid method and device, to solve the removal grid having
Method in the presence of grid poor removal effect technical problem.
The technical solution adopted by the present invention is as follows:
According to an aspect of the present invention, there is provided a kind of image based on deep learning goes grid method, including step:
Grid of different sizes in multiple original mesh images is extracted, constructs a variety of nets corresponding with the type of grid
Grid template;
Grid image, the generation multi-class grid number corresponding with net template are constructed online using a variety of net templates
According to;
Sorter network and full convolutional network is respectively trained using the multi-class grid data of generation as training data;
Classification processing is carried out to grid image to be removed using the sorter network trained;And according to the knot of classification processing
Fruit, grid is carried out to the grid image to be removed classified using the full convolutional network trained and handled.
Further, grid of different sizes in multiple original mesh images is extracted, is constructed relative with the type of grid
The step of a variety of net templates answered, includes:
The different grids in multiple the original mesh images collected are extracted, obtain the pixel size of the grid of extraction;
According to the pixel size of the grid of acquisition, grid is classified, by mesh generation into polytype, and build with
The corresponding net template of the type of grid.
Further, grid image, the generation multiclass corresponding with net template are constructed online using a variety of net templates
The step of other grid data, includes:
A variety of net templates are superimposed in non-screening table images, grid image corresponding to online construction;
Non- screening table images and grid image are integrated to form training image mapping pair, and the training image of formation is mapped
To as multi-class grid data.
Further, sorter network includes two sorter networks, and multi-class grid data includes refined net data and coarse grid
Data, wrapped using the multi-class grid data of generation as training data the step of sorter network and full convolutional network is respectively trained
Include:
Two sorter networks are trained using refined net data and coarse grid data as training data, two classification that will be trained
Network is as two sorter network models;
Full convolutional network is trained using refined net data and coarse grid data as training data, the full convolution that will be trained
Full convolutional network model corresponding to network conduct.
Further, being carried out the step of classification processing to grid image to be removed using the sorter network trained is included:
Classification processing is carried out to grid image to be removed using sorter network, judges the network mould of grid image to be removed
Formula;
According to the network mode judged, grid image to be removed is transported in corresponding full convolutional network model, with
Obtain removing grid image after grid.
According to another aspect of the present invention, a kind of image based on deep learning is also provided and goes mesh device, including:
Module is built, for extracting grid of different sizes in multiple original mesh images, constructs the type with grid
Corresponding a variety of net templates;
Generation module, for constructing grid image online using a variety of net templates, generation is corresponding with net template
Multi-class grid data;
Train template, for using the multi-class grid data of generation as training data come be respectively trained sorter network and entirely
Convolutional network;
Grid processing module, for carrying out classification processing to grid image to be removed using the sorter network trained;And
According to the result of classification processing, the grid image to be removed classified is carried out at grid using the full convolutional network trained
Reason.
Further, structure module includes extraction unit and construction unit,
Extraction unit, for extracting the different grids in multiple the original mesh images collected, obtain the grid of extraction
Pixel size;
Construction unit, for the pixel size of the grid according to acquisition, grid is classified, by mesh generation into a variety of
Type, and build the net template corresponding with the type of grid.
Further, generation module includes structural unit and integral unit,
Structural unit, for a variety of net templates to be superimposed in non-screening table images, grid chart corresponding to online construction
Picture;
Integral unit, for integrating non-screening table images and grid image to form training image mapping pair, and it will be formed
Training image map to as multi-class grid data.
Further, sorter network includes two sorter networks, and multi-class grid data includes refined net data and coarse grid
Data, training template include the first training unit and the second training unit,
First training unit, for training two classification nets using refined net data and coarse grid data as training data
Network, using two sorter networks trained as two sorter network models;
Second training unit, for training full convolution net using refined net data and coarse grid data as training data
Network, using the full convolutional network trained as corresponding to full convolutional network model.
Further, grid processing module includes judging unit and supply unit,
Judging unit, for carrying out classification processing to grid image to be removed using sorter network, judge grid to be removed
The network mode of image;
Supply unit, for according to the network mode judged, grid image to be removed to be transported into corresponding full convolution
In network model, to obtain removing grid image after grid.
The invention has the advantages that:
Image provided by the invention based on deep learning removes grid method and device, and net is constructed online by net template
Table images, the generation multi-class grid data corresponding with net template, and using multi-class grid data as training data point
Xun Lian not sorter network and full convolutional network;Classification processing is carried out to grid image to be removed using the sorter network trained,
And according to the result of classification processing, grid is carried out to the grid image to be removed classified using the full convolutional network trained
Processing.Image provided by the invention based on deep learning removes grid method and device, using National Federation of Trade Unions's convolutional network as training net
Network, grid is carried out to grid image to be removed by the total convolutional network trained and handled, with the grid to be removed with grid
Image is exported as being originally inputted without the removal grid image of grid as target, learns relevant parameter;And pass through instruction
The sorter network perfected carries out classification processing to grid image to be removed in advance, then by the corresponding total convolutional network pair trained
The grid image to be removed classified carries out accordingly going grid to handle, and goes the effect of grid good.
In addition to objects, features and advantages described above, the present invention also has other objects, features and advantages.
Below with reference to figure, the present invention is further detailed explanation.
Embodiment
It should be noted that in the case where not conflicting, the feature in embodiment and embodiment in the application can phase
Mutually combination.Describe the present invention in detail below with reference to the accompanying drawings and in conjunction with the embodiments.
Reference picture 1, the preferred embodiments of the present invention provide a kind of image based on deep learning and remove grid method, including
Step:
Step S100, grid of different sizes in multiple original mesh images is extracted, is constructed relative with the type of grid
The a variety of net templates answered.
As shown in Figure 3 and Figure 4, the different original mesh image of multiple grids is collected, extracts size in original mesh image
Different grids, construct a variety of net templates corresponding with the type of grid.
Step S200, grid image, the generation multiclass corresponding with net template are constructed online using a variety of net templates
Other grid data.
Grid image is constructed online using a variety of net templates, based on not band grid image, grid corresponding to generation
Image, band grid image and corresponding grid image it will not be combined to form training image mapping pair, and by the training figure of formation
As mapping is to as the multi-class grid data corresponding with net template.And the grid image of corresponding generation is subjected to data and added
By force, a variety of noises are added, to ensure the robustness of data, and the grid image of generation is more nearly real grid image.
Step S300, sorter network and full convolution is respectively trained using the multi-class grid data of generation as training data
Network.
According to the thickness of grid, the multi-class grid data of generation is divided into two classes, one kind is coarse grid data, another kind of
For refined net data.A sorter network is trained using the coarse grid data of generation and refined net data as training data,
Simultaneously to FCN (Fully Convolutional Networks, full convolutional network) corresponding to every a kind of grid training.
Step S400, classification processing is carried out to grid image to be removed using the sorter network trained;And according to classification
The result of processing, grid is carried out to the grid image to be removed classified using the full convolutional network trained and handled.
In test phase, classification processing is carried out to grid image to be removed using the sorter network trained;And according to point
The result of class processing, carries out grid to the grid image to be removed classified using the FCN trained and handles, obtain grid
Afterwards remove grid image.
The image based on deep learning that the present embodiment provides removes grid method, and grid chart is constructed online by net template
Picture, the generation multi-class grid data corresponding with net template, and instructed respectively using multi-class grid data as training data
Practice sorter network and full convolutional network;Classification processing, and root are carried out to grid image to be removed using the sorter network trained
According to the result of classification processing, the grid image to be removed classified is carried out at grid using the full convolutional network trained
Reason.The image based on deep learning that the present embodiment provides removes grid method, using National Federation of Trade Unions's convolutional network as training network, passes through
The total convolutional network trained carries out grid to grid image to be removed and handled, using the grid image to be removed with grid as
It is originally inputted, is exported without the removal grid image of grid as target, learns relevant parameter;And point by training
Class network carries out classification processing to grid image to be removed in advance, then by the corresponding total convolutional network trained to having classified
Grid image to be removed carries out accordingly going grid to handle, and goes the effect of grid good.
Preferably, as shown in Fig. 2 the image based on deep learning that the present embodiment provides removes grid method, step S100
Including:
Step S110, the different grids in multiple original mesh images that extraction has been collected, the picture of the grid of extraction is obtained
Plain size.
See Fig. 3 and Fig. 4, the different original mesh image of multiple grids is collected, extracts multiple original meshes collected
Different grids in image, obtain the pixel size of the grid of extraction.
Step S120, according to the pixel size of the grid of acquisition, grid is classified, by mesh generation into multiple types
Type, and build the net template corresponding with the type of grid.
The pixel size of grid is obtained, grid is classified according to the pixel size of the grid of acquisition, by mesh generation
Into polytype, and build the net template corresponding with the type of grid.Specifically, in the present embodiment, by mesh generation
Into refined net and the class of coarse grid two, classification foundation be in net template net lattice control width be less than 5 pixels for refined net, greatly
In 5 pixels be coarse grid (referring to Fig. 4, the first row secondary series and the second row first are classified as refined net, and remaining is coarse grid).
The image based on deep learning that the present embodiment provides removes grid method, extracts multiple the original mesh figures collected
Different grids as in, obtain the pixel size of the grid of extraction;According to the pixel size of the grid of acquisition, grid is divided
Class, by mesh generation into polytype, and build the net template corresponding with the type of grid.The present embodiment provide based on
The image of deep learning removes grid method, and according to the pixel size of the grid of acquisition, grid is classified, by mesh generation into
Polytype, and the net template corresponding with the type of grid is built, ensure the diversity of sample, so as to make obtained classification
Network and full convolutional network are preferably trained and learnt, and are made the better of grid, are demonstrate,proved and gone after network processes
Grid image is as far as possible close to original image.
Preferably, as shown in figure 5, the image based on deep learning that the present embodiment provides removes grid method, step S200
Including:
Step S210, a variety of net templates are superimposed in non-screening table images, grid image corresponding to online construction.
As shown in Figure 6 and Figure 7, according to a variety of net templates constructed, based on not band grid image, by a variety of nets
Grid template is superimposed in non-screening table images, grid image corresponding to online construction.
Step S220, non-screening table images and grid image are integrated to form training image mapping pair, and by the instruction of formation
Practice image mapping to as multi-class grid data.
Non- screening table images and corresponding grid image are integrated, form training image mapping pair, and by formation
Training image is mapped to being trained respectively to sorter network and full convolutional network as multi-class grid data.
The image based on deep learning that the present embodiment provides removes grid method, by the way that a variety of net templates are superimposed on not
In screening table images, grid image corresponding to online construction;Non- screening table images and grid image are integrated to form training image
Mapping pair, and the training image of formation is mapped to as multi-class grid data.The present embodiment provide based on deep learning
Image remove grid method, using training image mapping to training sorter network and full convolutional network respectively, make to train
Sorter network and full convolutional network grid is carried out to grid image to be removed and handled, go the effect of grid good;Online generation
Multi-class grid data in add a variety of net templates, ensure the diversity of sample.
Preferably, as shown in figure 8, the image based on deep learning that the present embodiment provides removes grid method, sorter network
Including two sorter networks, multi-class grid data includes refined net data and coarse grid data, and step S300 includes:
Step S310, two sorter networks are trained using refined net data and coarse grid data as training data, will trained
Two good sorter networks are as two sorter network models.
Two sorter networks are trained using the refined net data constructed and coarse grid data as training data, will be instructed
Two sorter networks perfected as two sorter network models to be detected to images to be recognized, classification network such as Fig. 9 institutes
Show, sorter network includes the first convolutional layer, the second convolutional layer, the 3rd convolutional layer, the first full articulamentum and the second full articulamentum.When
So, the first excitation layer and the first pond layer can also be connect behind the first convolutional layer, the second excitation can be connect behind the second convolutional layer
Layer and the second pond layer, the 3rd excitation layer and the 3rd pond layer can be connect behind the 3rd convolutional layer.The input of first convolutional layer is 112
× 112 × 3 sized images, followed by the convolution kernel that 20 yardsticks are 5 × 5;The core of first pond layer is 3 × 3, step-length 2.
Second convolutional layer has the convolution kernel of 40 3 × 3;3rd convolutional layer has the convolution kernel of 60 3 × 3.First excitation layer, the second excitation
Layer is identical with the structure of the 3rd excitation layer, and the structure of the first pond layer, the second pond layer and the second pond layer is identical.First connects entirely
The characteristic length for connecing layer is 128.Second full articulamentum is divided to refined net and the class of coarse grid two to be exported.In the present embodiment, net
Network input size be 112*112*3, after three convolutional layers, carries out the first full articulamentum processing, formed length for 128 the
One feature vector chart (featureMap), reconnect second and connect the class label output for corresponding to that length is 2 entirely.
In the present embodiment, classification task is one two classification, as follows using cross entropy loss function, its formula:
Wherein, piIt is the probability of refined net to represent the sample,Value be 0 or 1, respectively to should sample reality
Label is refined net or coarse grid.
Step S320, full convolutional network is trained using refined net data and coarse grid data as training data, will trained
Full convolutional network model corresponding to good full convolutional network conduct.
FCN networks are trained using the refined net data constructed and coarse grid data as training data, by what is trained
FCN networks be used as corresponding to full convolutional network model to be detected to images to be recognized.FCN network structures are as indicated by 10:FCN
Network includes Volume Four lamination, the 5th convolutional layer and the 6th convolutional layer, and the input of FCN networks is 56 × 56 × 3 sized images, the
The size of one convolution kernel is 9 × 9, and the number of core is 64, and the size for mending side (pad) is 6, second after the processing of Volume Four lamination
Feature vector chart (featureMap) size is 60 × 60 × 64.Second convolution kernel size is 5 × 5, and the number of core is 32,
Pad is 1, and the third feature vectogram featureMap sizes handled well by the 5th convolutional layer are 32 × 60 × 60.3rd
Convolution kernel size be 5 × 5, number 3, pad is 0, by the 6th convolutional layer processing after fourth feature vectogram size be 56 ×
56 × 3, and input it is identical, this method is also called the learning method of end-to-end (end to end).In addition, in Volume Four lamination
Behind connect the first excitation layer (Relu), connect behind the 5th convolutional layer the second excitation layer (Relu) and carry out non-linear place
Reason.The error of output is finally input to as loss function, reaction using Euclidean distance.In the present embodiment, in full convolution net
In network, without full articulamentum, only convolutional layer and excitation layer.Three-layer coil is set to accumulate herein, by setting pad and convolution kernel size
Etc. parameter, input is equal with output size after ensureing three-layer coil product processing.Using Euclidean distance as loss function, ensure warp
The image crossed after network processes is as far as possible close to original image (non-screening table images).During training, using less learning rate, very much
One of to ten a ten thousandths, ensure preferably convergence effect.
The image based on deep learning that the present embodiment provides removes grid method, by by refined net data and coarse grid number
According to two sorter networks are trained as training data, using two sorter networks trained as two sorter network models;By fine-structure mesh
Lattice data and coarse grid data train full convolutional network as training data, using the full convolutional network trained as corresponding to
Full convolutional network model.The image based on deep learning that the present embodiment provides removes grid method, by end to end FCN nets
Network structure is applied to grid and handled, and goes the effect of grid good.
Preferably, as shown in figure 11, the image based on deep learning that the present embodiment provides removes grid method, step S400
Including:
Step S410, classification processing is carried out to grid image to be removed using sorter network, judges grid image to be removed
Network mode.
Test phase, the grid image to be removed for inputting sorter network is subjected to classification processing, judges grid chart to be removed
Seem to belong to coarse grid pattern or refined net pattern.
Step S420, according to the network mode judged, grid image to be removed is transported to corresponding full convolutional network
In model, to obtain removing grid image after grid.
The network mode finally judged according to sorter network, for different models, grid image to be removed is conveyed
Into corresponding full convolutional network model, to obtain removing grid image after grid.
The image based on deep learning that the present embodiment provides removes grid method, by using sorter network to net to be removed
Table images carry out classification processing, judge the network mode of grid image to be removed;, will be to be removed according to the network mode judged
Grid image is transported in corresponding full convolutional network model, to obtain removing grid image after grid.The present embodiment provides
The image based on deep learning remove grid method, grid image to be removed is divided in advance by the sorter network trained
Class processing, then the grid image to be removed classified is carried out accordingly going grid to handle by corresponding full convolutional network model, go
The effect of grid is good, and ensures to go grid image as far as possible close to original image after network processes.
Preferably, as shown in figure 12, the present embodiment also provides a kind of image based on deep learning and goes mesh device, bag
Include:Module 10 is built, for extracting grid of different sizes in multiple original mesh images, is constructed relative with the type of grid
The a variety of net templates answered;Generation module 20, for constructing grid image, generation and grid mould online using a variety of net templates
The corresponding multi-class grid data of plate;Template 30 is trained, for using the multi-class grid data of generation as training data
Sorter network and full convolutional network is respectively trained;Grid processing module 40, for using the sorter network trained to be removed
Grid image carries out classification processing;And according to the result of classification processing, treated using the full convolutional network trained to what is classified
Removal grid image carries out grid and handled.
As shown in Figure 3 and Figure 4, build module 10 and collect the different original mesh image of multiple grids, extract original mesh
Grid of different sizes in image, construct a variety of net templates corresponding with the type of grid.
Generation module 20 constructs grid image online using a variety of net templates, based on not band grid image, generation
Corresponding grid image, band grid image and corresponding grid image it will not be combined to form training image mapping pair, and by shape
Into training image map to as the multi-class grid data corresponding with net template.And the grid image that generation will be corresponded to
Data reinforcement is carried out, a variety of noises is added, to ensure the robustness of data, and the grid image of generation is more nearly really
Grid image.
According to the thickness of grid, the multi-class grid data of generation is divided into two classes, one kind is coarse grid data, another kind of
For refined net data.Template 30 is trained using the coarse grid data of generation and refined net data as training data to train one
Sorter network, while to FCN (Fully Convolutional Networks, full convolution net corresponding to every a kind of grid training
Network).
In test phase, grid processing module 40 is classified using the sorter network trained to grid image to be removed
Processing;And according to the result of classification processing, the grid image to be removed classified is carried out at grid using the FCN trained
Reason, obtains removing grid image after grid.
The image based on deep learning that the present embodiment provides goes mesh device, and grid chart is constructed online by net template
Picture, the generation multi-class grid data corresponding with net template, and instructed respectively using multi-class grid data as training data
Practice sorter network and full convolutional network;Classification processing, and root are carried out to grid image to be removed using the sorter network trained
According to the result of classification processing, the grid image to be removed classified is carried out at grid using the full convolutional network trained
Reason.The image based on deep learning that the present embodiment provides goes mesh device, using National Federation of Trade Unions's convolutional network as training network, passes through
The total convolutional network trained carries out grid to grid image to be removed and handled, using the grid image to be removed with grid as
It is originally inputted, is exported without the removal grid image of grid as target, learns relevant parameter;And point by training
Class network carries out classification processing to grid image to be removed in advance, then by the corresponding total convolutional network trained to having classified
Grid image to be removed carries out accordingly going grid to handle, and goes the effect of grid good.
Preferably, as shown in figure 13, the present embodiment also provides a kind of image based on deep learning and goes mesh device, structure
Module 10 includes extraction unit 11 and construction unit 12, extraction unit 11, for extracting multiple the original mesh images collected
In different grids, obtain the pixel size of the grid of extraction;Construction unit 12, the pixel for the grid according to acquisition are big
It is small, grid is classified, by mesh generation into polytype, and builds the net template corresponding with the type of grid.
See Fig. 3 and Fig. 4, the different original mesh image of multiple grids is collected, extracts multiple original meshes collected
Different grids in image, obtain the pixel size of the grid of extraction.
The pixel size of grid is obtained, grid is classified according to the pixel size of the grid of acquisition, by mesh generation
Into polytype, and build the net template corresponding with the type of grid.Specifically, in the present embodiment, by mesh generation
Into refined net and the class of coarse grid two, classification foundation be in net template net lattice control width be less than 5 pixels for refined net, greatly
In 5 pixels be coarse grid (referring to Fig. 4, the first row secondary series and the second row first are classified as refined net, and remaining is coarse grid).
The image based on deep learning that the present embodiment provides goes mesh device, extracts multiple the original mesh figures collected
Different grids as in, obtain the pixel size of the grid of extraction;According to the pixel size of the grid of acquisition, grid is divided
Class, by mesh generation into polytype, and build the net template corresponding with the type of grid.The present embodiment provide based on
The image of deep learning goes mesh device, and according to the pixel size of the grid of acquisition, grid is classified, by mesh generation into
Polytype, and the net template corresponding with the type of grid is built, ensure the diversity of sample, so as to make obtained classification
Network and full convolutional network are preferably trained and learnt, and are made the better of grid, are demonstrate,proved and gone after network processes
Grid image is as far as possible close to original image.
Preferably, as shown in figure 14, the present embodiment also provides a kind of image based on deep learning and goes mesh device, generation
Module 20 includes structural unit 21 and integral unit 22, structural unit 21, for a variety of net templates to be superimposed on into non-screening lattice
In image, grid image corresponding to online construction;Integral unit 22, for integrating non-screening table images and grid image to be formed
Training image mapping pair, and the training image of formation is mapped to as multi-class grid data.
As shown in Figure 6 and Figure 7, structural unit 21 is according to a variety of net templates constructed, using not band grid image as base
Plinth, a variety of net templates are superimposed in non-screening table images, grid image corresponding to online construction.
Integral unit 22 is integrated non-screening table images and corresponding grid image, forms training image mapping pair,
And the training image of formation is mapped to being instructed respectively to sorter network and full convolutional network as multi-class grid data
Practice.
The image based on deep learning that the present embodiment provides goes mesh device, by the way that a variety of net templates are superimposed on not
In screening table images, grid image corresponding to online construction;Non- screening table images and grid image are integrated to form training image
Mapping pair, and the training image of formation is mapped to as multi-class grid data.The present embodiment provide based on deep learning
Image go mesh device, using training image mapping to training sorter network and full convolutional network respectively, make to train
Sorter network and full convolutional network grid is carried out to grid image to be removed and handled, go the effect of grid good;Online generation
Multi-class grid data in add a variety of net templates, ensure the diversity of sample.
Preferably, as shown in figure 15, the present embodiment also provides a kind of image based on deep learning and goes mesh device, training
Template 30 includes the first training unit 31 and the second training unit 32, the first training unit 31, for by refined net data and slightly
Grid data trains two sorter networks as training data, using two sorter networks trained as two sorter network models;
Second training unit 32, for training full convolutional network using refined net data and coarse grid data as training data, it will instruct
Full convolutional network model corresponding to the full convolutional network conduct perfected.
First training unit 31 trains one using the refined net data constructed and coarse grid data as training data
Two sorter networks, using two sorter networks trained as two sorter network models to be detected to images to be recognized, classify
Network as shown in figure 9, sorter network include the first convolutional layer, the second convolutional layer, the 3rd convolutional layer, the first full articulamentum and
Second full articulamentum.Certainly, the first excitation layer and the first pond layer can also be connect behind the first convolutional layer, behind the second convolutional layer
The second excitation layer and the second pond layer can be connect, the 3rd excitation layer and the 3rd pond layer can be connect behind the 3rd convolutional layer.First
Convolutional layer input is 112 × 112 × 3 sized images, followed by the convolution kernel that 20 yardsticks are 5 × 5;The core of first pond layer
For 3 × 3, step-length 2.Second convolutional layer has the convolution kernel of 40 3 × 3;3rd convolutional layer has the convolution kernel of 60 3 × 3.First
The structure of excitation layer, the second excitation layer and the 3rd excitation layer is identical, the knot of the first pond layer, the second pond layer and the second pond layer
Structure is identical.The characteristic length of first full articulamentum is 128.Second full articulamentum is divided to refined net and the class of coarse grid two to be exported.
In the present embodiment, network inputs size is 112*112*3, after three convolutional layers, carries out the first full articulamentum processing, shape
Into the first eigenvector figure (featureMap) that length is 128, reconnect second and connect the class label for corresponding to that length is 2 entirely
Output.
In the present embodiment, classification task is one two classification, as follows using cross entropy loss function, its formula:
Wherein, piIt is the probability of refined net to represent the sample,Value be 0 or 1, respectively to should sample actual mark
Sign is refined net or coarse grid.
Second training unit 32 trains FCN nets using the refined net data constructed and coarse grid data as training data
Network, using the FCN networks trained as corresponding to full convolutional network model to be detected to images to be recognized.FCN network knots
Structure is as indicated by 10:FCN networks include Volume Four lamination, the 5th convolutional layer and the 6th convolutional layer, and the input of FCN networks is 56 × 56
× 3 sized images, the size of first convolution kernel is 9 × 9, and the number of core is 64, and the size for mending side (pad) is 6, Volume Four product
Second feature vectogram (featureMap) size after layer processing is 60 × 60 × 64.Second convolution kernel size is 5 × 5,
The number of core is 32, pad 1, the third feature vectogram featureMap sizes handled well by the 5th convolutional layer for 32 ×
60×60.3rd convolution kernel size is 5 × 5, and number 3, pad is 0, the fourth feature vector after the processing of the 6th convolutional layer
Figure size is 56 × 56 × 3, and input is identical, and this method is also called the learning method of end-to-end (end to end).In addition,
The first excitation layer (Relu) is connected behind Volume Four lamination, the second excitation layer (Relu) is connected behind the 5th convolutional layer and enters
Row Nonlinear Processing.The error of output is finally input to as loss function, reaction using Euclidean distance.In the present embodiment,
In full convolutional network, without full articulamentum, only convolutional layer and excitation layer.Herein set three-layer coil accumulate, by set pad and
The parameters such as convolution kernel size, input is equal with output size after ensureing three-layer coil product processing.Using Euclidean distance as loss
Function, ensure the image after network processes as far as possible close to original image (non-screening table images).During training, use is less
Learning rate, a ten thousandth to ten a ten thousandths, ensure preferably convergence effect.
The image based on deep learning that the present embodiment provides goes mesh device, by by refined net data and coarse grid number
According to two sorter networks are trained as training data, using two sorter networks trained as two sorter network models;By fine-structure mesh
Lattice data and coarse grid data train full convolutional network as training data, using the full convolutional network trained as corresponding to
Full convolutional network model.The image based on deep learning that the present embodiment provides goes mesh device, by end to end FCN nets
Network structure is applied to grid and handled, and goes the effect of grid good.
Preferably, as shown in figure 16, the present embodiment also provides a kind of image based on deep learning and removes mesh device, grid
Processing module 40 includes judging unit 41 and supply unit 42, judging unit 41, for using sorter network to grid to be removed
Image carries out classification processing, judges the network mode of grid image to be removed;Supply unit 42, for according to the network judged
Pattern, grid image to be removed is transported in corresponding full convolutional network model, to obtain removing grid image after grid.
The grid image to be removed for inputting sorter network is carried out classification processing, judges to treat by test phase, judging unit 41
Remove grid image and belong to coarse grid pattern or refined net pattern.
The network mode that supply unit 42 is finally judged according to sorter network, for different models, by net to be removed
Table images are transported in corresponding full convolutional network model, to obtain removing grid image after grid.
The image based on deep learning that the present embodiment provides goes mesh device, by using sorter network to net to be removed
Table images carry out classification processing, judge the network mode of grid image to be removed;, will be to be removed according to the network mode judged
Grid image is transported in corresponding full convolutional network model, to obtain removing grid image after grid.The present embodiment provides
The image based on deep learning go mesh device, grid image to be removed is divided in advance by the sorter network trained
Class processing, then the grid image to be removed classified is carried out accordingly going grid to handle by corresponding full convolutional network model, go
The effect of grid is good, and ensures to go grid image as far as possible close to original image after network processes.
The preferred embodiments of the present invention are the foregoing is only, are not intended to limit the invention, for the skill of this area
For art personnel, the present invention can have various modifications and variations.Within the spirit and principles of the invention, that is made any repaiies
Change, equivalent substitution, improvement etc., should be included in the scope of the protection.