CN107833183A - A kind of satellite image based on multitask deep neural network while super-resolution and the method for coloring - Google Patents
A kind of satellite image based on multitask deep neural network while super-resolution and the method for coloring Download PDFInfo
- Publication number
- CN107833183A CN107833183A CN201711224807.3A CN201711224807A CN107833183A CN 107833183 A CN107833183 A CN 107833183A CN 201711224807 A CN201711224807 A CN 201711224807A CN 107833183 A CN107833183 A CN 107833183A
- Authority
- CN
- China
- Prior art keywords
- resolution
- mrow
- network
- image
- msub
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformation in the plane of the image
- G06T3/40—Scaling the whole image or part thereof
- G06T3/4053—Super resolution, i.e. output image resolution higher than sensor resolution
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T11/00—2D [Two Dimensional] image generation
- G06T11/001—Texturing; Colouring; Generation of texture or colour
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformation in the plane of the image
- G06T3/40—Scaling the whole image or part thereof
- G06T3/4046—Scaling the whole image or part thereof using neural networks
Abstract
The invention discloses a kind of satellite image based on multitask deep neural network while super-resolution and the method for coloring, belong to technical field of image processing.The invention mainly includes steps：1st, the gray level image block training set of high-resolution and low resolution is made；2nd, the deep neural network for building a multitask is used for model training；3rd, the depth network based on structure and the training set of making are trained to network model；4th, the model parameter according to study, the gray level image of a width low resolution is inputted, obtained output is the high-resolution coloured image of reconstruction.The present invention possesses the depth super-resolution network and coloured networks of premium properties by combining, not only increase the detail section of satellite image, but also can carry out coloring to gray level image simultaneously makes it automatically generate the color remote sensing image for meeting the sense of reality, the step of also reducing execution and time, had a wide range of applications in the fields such as gray level image coloring, satellite remote sensing remote measurement.
Description
Technical field
The invention belongs to technical field of image processing, and multitask deep neural network is based on more specifically to one kind
Satellite image simultaneously super-resolution and coloring method.
Background technology
As the network performance with simple function is more and more stronger, people are to the network that can handle complicated multitask
Demand also increasingly increase.Traditional method is using the input exported as another network of a network, then could be obtained
Result to the end.Because this method not only needs man-machine interactively, and execution can waste many times one by one, it is also contemplated that
With the presence or absence of problems such as compatibility issues between two networks so that people have to seek other method.
Existing network need to have two very important functions, i.e. super-resolution and coloring.In terms of super-resolution, rebuild
Technology can be divided into different types, can mainly be divided into 3 classes：Method based on interpolation, the method based on reconstruction, it is based on
The method of study.Wherein, the method based on study is generally from an external data focusing study high-definition picture and low resolution
Mapping relations between rate image, high-definition picture then is rebuild using the mapping relations of study, be most popular at present
Method.For example convolutional neural networks are applied in image super-resolution rebuilding task by Dong et al. first, they pass through structure
One three-layer coil accumulates neutral net to generate super-resolution image；Also He et al. rebuilds high resolution graphics by residual unit
Picture.In terms of coloring, method can from it is earliest based on interaction, such as：Luan et al. proposes the phase using neighbor pixel
Likelihood metric colours, and then has scholar to propose an automanual method, and it is reference picture or multiple images to the defeated of grey
The Color Statistical entered.Later, full-automatic method was also suggested, for example Zhang et al. makes e-learning combination rudimentary and advanced
Clue dyes.
In recent years, the training method due to the powerful learning ability of convolutional neural networks and end to end so that computer regards
Feel is obtained for progress, such as image classification and recognition of face in many aspects, and many networks have good performance, because
This someone starts to consider how to allow a real-time performance multiple tasks, such as Iizuka et al. to propose a global network, study
The environment of image is semantic so that the result of coloring is more accurate.But because the network is still used as driving, institute using data
So that if the picture/mb-type tested is not included in training set class, then poor effect, and the instruction of sorter network can be produced
Practice difficulty greatly, it is necessary to be lot more time to restrain, allow it to aid in another network to colour gray level image, Ke Nengyi
Start to have negative effect.
It is September in 2016 27 through retrieval, Chinese Patent Application No. 201610856231.1, the applying date, innovation and creation
It is entitled：The face character analysis method of convolutional neural networks based on multi-task learning, this application case comprise the following steps：1、
Single task model analysis：1) original sample of each age facial image is subjected to face critical point detection, pedestrian's face of going forward side by side alignment
The new samples for generating and including facial image are cut according to pre-set dimension afterwards；2) using the new samples of step 1) generation, it is respectively trained
Three age estimation network, sex identification network, species network single task convolutional neural networks, the convergence of more each network
Speed, obtain the weights of a most slow single task convolutional neural networks of convergence rate；Pre-set dimension cuts generation and includes face
The new samples of image；2nd, multi task model is trained：1) multitask convolutional neural networks are built, it is defeated that the network shares three tasks
Go out, correspond to age estimation, sex identification and species respectively, three tasks are all using softmax loss functions as target
Function；The multitask convolutional neural networks include being used for the shared part that data sharing and information exchange in multi-task learning,
And for calculating the independent sector of above three task output；Weights using the single task convolutional neural networks of acquisition are initial
Change the shared part of multitask convolutional neural networks, the multitask convolutional neural networks formed after initialization；2) generation is utilized
New samples, train multitask convolutional neural networks, the multitask convolutional neural networks model trained；3rd, face character is sentenced
It is disconnected：1) Face datection is carried out to inputted picture, judges whether to include facial image, as carried out face to input picture comprising if
Critical point detection, pedestrian's face of going forward side by side alignment, the new picture for generating and including facial image is then cut according to pre-set dimension；2) by institute
New picture is obtained, is input to obtained multitask convolutional neural networks model, carries out age estimation, sex identification and species.
Although the method realizes a kind of multitask network, by the way that face is inputted, each attribute of face is obtained, such as：Year
Age estimation, sex identification and species, but this application case has the following disadvantages：1) although this application case can realize more
The function of business, but analyzed the attribute of face, so three networks are similar, but affirm the problem of in reality
It is complicated and changeable, it is impossible to so close situation occur；2) due to the network solve the problems, such as it is similar, so network structure
Also it is similar, two completely different tasks are if desired handled simultaneously, then need to consider combination and the feature of network on this basis
It is shared.
Analyzed based on more than, prior art needs a kind of method for the deep neural network that can obtain more preferable multitask.
The content of the invention
1. invention technical problems to be solved
In order to overcome existing for above-mentioned prior art can not be complicated and changeable in Coping with Reality the problem of and multitask between deposit
May it is incompatible the problem of；The present invention proposes a kind of satellite image based on multitask deep neural network super-resolution simultaneously
With the method for coloring；The present invention not only carries out multi-task learnings to completely different two aspects, and solve network it
Between the problem of incompatible be present, meet requirement complicated in reality.
2. technical scheme
To reach above-mentioned purpose, technical scheme provided by the invention is：
A kind of satellite image based on multitask deep neural network while super-resolution of the present invention and the method for coloring, its
Step is：
Step 1, using satellite color image data collection, make the image block training set of high-resolution and low resolution；
Step 2, the deep neural network of one multitask of structure are used for model training；
The network that step 3, the training set made according to step 1 and step 2 are built, adjusts network parameter, carries out network instruction
Practice；
Step 4, the input using the gray level image of a width low resolution as network, the parameter obtained using step 3 study
A high-resolution coloured image is rebuild as output.
Further, the process of the coloured image block training set of step 1 making high-resolution and low resolution is：
Every coloured image is concentrated for a conventional satellite image processing data, two are carried out to high-definition picture first
Secondary bicubic interpolation, obtain with high-definition picture corresponding to identical size low-resolution image；Then by every high-resolution
Image and low-resolution image are cut into multiple images block, and intersection be present between adjacent image block, thus obtain
Set for the high-definition picture block and low-resolution image block of depth network training.
Further, one 43 layers of depth network model is built in step 2, one is divided into three parts, right first
Image is pre-processed, 20 layers of composition super-resolution network afterwards, last 23 layers of composition coloured networks；In image preprocessing section
Point, coloured image is transformed into Lab color spaces from RGB, coloured image is then divided into two parts, a part is L vectors, is made
For the input of whole network；Another part be ab vector, the label as last coloured networks；In super-resolution network, altogether
9 residual error layers and two convolutional layers are contained, wherein each residual unit there are two convolutional layers；Each convolutional layer is followed by one
PReLU active coatings；Shown in residual unit such as formula (1)：
yi=0.9*h (xi)+0.1*F(xi,wi)
xi+1=f (yi) (1)
Wherein, xiIt is expressed as the feature input of i-th layer of residual unit, wiIt is expressed as setting for i-th layer of weight and bias term
Put, F represents residual error function, and it is that identity maps h (x that f, which then represents activation primitive PReLU, h,i)=xi；
The depth network by learn low resolution gray level image block and high-resolution gray level image block between mapping
Relation, as shown in formula (2)：
X=F (y, Φ) (2)
Wherein, x, y are respectively high-resolution gray level image block and the gray level image block of low resolution, and Φ is super-resolution
The model parameter that e-learning arrives, for follow-up high resolution image reconstruction；
In coloured networks, 4th layer reciprocal is warp lamination, and residue is convolutional layer；In each convolutional layer and deconvolution
Layer is followed by a Relu active coating；Network inputs be super-resolution network output high-resolution gray level image block, the network
The mapping relations between high-resolution gray level image block and ab color component images block will be learnt, as shown in formula (3)：
X=f (y, θ) (3)
Wherein, x, y are respectively ab color component images block and high-resolution gray level image block, and θ learns for coloured networks
The model parameter arrived, for predicting the ab chrominance components in high-resolution gray-scale map corresponding to the brightness L of each pixel afterwards,
And the result of prediction is combined with high-resolution gray level image L, has obtained the high-definition picture of Lab color spaces, then
It is transformed into RGB color, it is possible to obtain our desired coloured images.
Further, the loss function of network training uses difference in super-resolution network and coloured networks in step 2
Method, in super-resolution network, loss function using mean square error represent, as shown in formula (4)：
Wherein, N be step 1 gained training set in sample size, xi,yiFor i-th of high-definition picture block and corresponding low
Image in different resolution block；
In coloured networks, loss function intersects entropy loss using multinomial and represented, as shown in formula (5)：
Wherein,The probability distribution of prediction is represented, and Z then represents real probability distribution, function v is a rebalancing
The factor, it is obtained based on the statistics to training set ab color components, and h and w represent the length and width of image respectively, and q is then instruction
Practice the classification sum for concentrating ab color components.
Further, the activation primitive of ReLU active coatings represents as follows with formula (6) in step 2：
F (x)=max (0, x) (6)
Wherein, x is the input of ReLU activation primitives, and f (x) is the output of ReLU activation primitives；
The activation primitive of PReLU active coatings represents as follows with formula (7)：
F (x)=max (0, x)+a*min (0, x) (7)
Wherein, x is the input of PReLU activation primitives, and f (x) is the output of PReLU activation primitives, and a is can learning parameter.
Further, in step 2, except last layer of convolutional layer, all convolutional layers of the depth network of structure
Convolution kernel size is set to 3*3, and the convolution kernel of last layer is dimensioned to 1*1；The convolution kernel of warp lamination is dimensioned to
4*4。
Further, in super-resolution network, the quantity of the characteristic pattern of preceding 19 convolutional layers is all set to 64, finally
One layer of characteristic pattern quantity is 1；In coloured networks, the quantity of characteristic pattern corresponding to preceding 7 convolutional layers is set to 64,64,
128th, 128,256,256,256, followed by 12 convolutional layers be arranged to 512, the feature of 3 convolutional layers followed by
The quantity of figure is set as 256, and the quantity of the characteristic pattern of last layer of convolution is set to 244, and each convolutional layer and warp lamination obtains
Output such as formula (8) represent：
yi=f (Wixi+bi), i=1,2 ..., 43 (8)
Wherein, WiRepresent i-th layer of weight, biRepresent i-th layer of biasing, xiRepresent i-th layer of input, yiRepresent i-th layer
Output；
Pass through activation primitive ReLU and PReLU respectively, as a result as shown in formula (9) and (10)：
zi=max (0, yi) (9)
zj=max (0, yj)+a*min(0,yj) (10)
Wherein, yiAnd yjThe respectively output of activation primitive ReLU and PReLU last layers, ziAnd zjRespectively activation primitive
ReLU and PReLU output.
Further, step 3 is trained using Caffe deep learning platforms to network, first to being built in step 2
Multitask deep neural network weight and biasing initialized, detailed process is：
1) after initializing weight W using MSRA modes in super-resolution network, W meets following Gaussian Profile：
Wherein, n represents the layer network input block number, i.e. convolutional layer input feature vector figure quantity；
In coloured networks, weights initialisation is all set to 0, i.e. Wi=0；
2) in the entire network, biasing is all initialized as 0, i.e. bi=0.
Further, step 3 represents as follows using gradient descent method renewal network parameter with formula (12)：
Vi+1=μ Vi-α▽L(Wi),Wi+1=Wi+Vi+1 (12)
Wherein, Vi+1Represent this weight updated value, and ViLast weight updated value is represented, and μ is last ladder
The weight of angle value, α are learning rates, ▽ L (Wi) it is gradient；
In the training process, network parameter renewal is carried out by given number of iterations.
3. beneficial effect
Using technical scheme provided by the invention, compared with existing known technology, there is following remarkable result：
(1) side of a kind of satellite image based on multitask deep neural network of the invention while super-resolution and coloring
Method, it is contemplated that reality in it is complicated and changeable the problem of, have chosen completely different both direction and image handled, i.e., to image
The work of super-resolution and coloring is carried out simultaneously, not only meets the demand of multitask, and can realize to other any kind
The image of class realizes super-resolution and coloring simultaneously, and the method meets requirement complicated in reality.
(2) side of a kind of satellite image based on multitask deep neural network of the invention while super-resolution and coloring
Method, the not only each self-optimizing but also they are shared and feature interaction realizes collaboration optimization by feature of two parts of the network,
To realize more preferable result.
(3) side of a kind of satellite image based on multitask deep neural network of the invention while super-resolution and coloring
Method, the super-resolution network and coloured networks of premium properties are possessed by combining, not only increase the detail section of satellite image, and
And can also carry out coloring to gray level image simultaneously makes it automatically generate the color remote sensing image for meeting the sense of reality；Compared to original list
One prototype network method, man-machine interactively is not needed not only, and perform and greatly shortened on the time, in historical photograph and remote sensing images
Deng being had a wide range of applications in field.
Brief description of the drawings
The deep neural network based on multitask that Fig. 1 is the present invention carries out super-resolution and coloring to satellite image simultaneously
Method flow diagram；
Fig. 2 is the Making programme figure of data set in the present invention；
Fig. 3 is the network model schematic diagram that builds of the present invention, the ReLU active coatings that are not drawn into Fig. 3 after convolution and
PReLU active coatings；
Fig. 4 is the detailed annotation figure of residual unit in the present invention.
Embodiment
To further appreciate that present disclosure, the present invention is described in detail in conjunction with the accompanying drawings and embodiments.
Embodiment 1
With reference to Fig. 1, a kind of satellite image based on multitask deep neural network of the present embodiment simultaneously super-resolution and
The method of color, specifically includes following steps：
Step 1, using conventional data set, such as ImageNet and AID satellite datas collection, make high-resolution figure
As block training set and the image block training set of low resolution, specific steps are as shown in Fig. 2 i.e.：
For every coloured image in frequently-used data collection (such as AID satellite image datas collection), first to high-definition picture
Bicubic interpolation (carrying out bicubic down-sampling interpolation, second of progress bicubic up-sampling interpolation for the first time) twice is carried out, is obtained
To with high-definition picture corresponding to identical size low-resolution image；
The image block that every high-definition picture and low-resolution image are cut into multiple 93*93 (is cut into 93*93's
Image block includes the feature learning of more conducively super-resolution), cut at intervals of 27 pixels so that deposited between adjacent image block
In the part that a part overlaps, high-definition picture block and low-resolution image block for depth network training resulting in
Set.
Step 2, the deep neural network of one multitask of structure are used for model training；
2-1, one 43 layers of depth network model is built, concrete structure is as shown in figure 3, and one be divided into three portions
Point, image is pre-processed first, 20 layers of composition super-resolution network afterwards, last 23 layers of composition coloured networks；Pre-
During processing, coloured image is transformed into Lab color spaces from RGB, coloured image is then divided into two parts, a part for L to
Amount, the input as whole network；Another part be ab vector, the label as last coloured networks；In super-resolution network,
9 residual error layers and two convolutional layers are contained altogether, wherein (remaining residual unit and first by taking first residual unit as an example
Individual residual unit is consistent), the concrete structure of residual unit in each residual unit as shown in figure 4, there is two convolutional layers；Each
Convolutional layer is followed by a PReLU active coating；Shown in residual unit such as formula (1)：
yi=0.9*h (xi)+0.1*F(xi,wi)
xi+1=f (yi) (1)
Wherein, xiIt is expressed as the feature input of i-th layer of residual unit, wiIt is expressed as setting for i-th layer of weight and bias term
Put, what F was represented is residual error function, and f, which then represents activation primitive PReLU, h, to be mapped as identity：h(xi)=xi。
In this network, the network will learn between gray level image block and the high-resolution gray level image block of low resolution
Mapping relations, as shown in formula (2)：
X=F (y, Φ) (2)
Wherein, x, y represent high-definition picture block and low-resolution image block respectively, and Φ is that super-resolution e-learning arrives
Model parameter, for high resolution image reconstruction afterwards.
Finally, in coloured networks, 4th layer reciprocal is warp lamination, and remaining is entirely convolutional layer；In each convolution
A Relu active coating is connect after layer and warp lamination；Network inputs are the high-resolution gray scales of a part of network output above
Image block, the network are such as public by the mapping relations between learning high-resolution gray level image block and ab color component images block
Shown in formula (3)：
X=f (y, θ) (3)
Wherein, x, y are respectively ab color component images block and high-resolution gray level image block, and θ learns for coloured networks
The model parameter arrived, for predicting the ab chrominance components in high-resolution gray-scale map corresponding to the brightness L of each pixel afterwards,
And the result of prediction is combined with high-resolution gray level image L, has obtained the high-definition picture of Lab color spaces, then
It is transformed into RGB color, it is possible to obtain our desired coloured images.
The loss function of network training employs different methods in two parts network, in super-resolution network, damage
Lose function and employ mean square error expression, as shown in formula (4)：
Wherein, N be step 1 gained training set in sample size, xi,yiFor i-th of high-definition picture block and corresponding low
Image in different resolution block, Φ are the model parameter that super-resolution e-learning arrives.
And in coloured networks, the loss function of training intersects entropy loss using multinomial and represented, as shown in formula (5)：
Wherein,The probability distribution of prediction is represented, and Z then represents real probability distribution, function v is a rebalancing
The factor, it is obtained based on the statistics to training set ab color components, and h and w represent the length and width of image respectively, q
It is then the ab color component classifications in training set.
The activation primitive of ReLU active coatings represents as follows with formula (6)：
F (x)=max (0, x) (6)
Wherein, x is the input of ReLU activation primitives, and f (x) is the output of ReLU activation primitives.
The activation primitive of PReLU active coatings represents as follows with formula (7)：
F (x)=max (0, x)+a*min (0, x) (7)
Wherein, x is the input of PReLU activation primitives, and f (x) is the output of PReLU activation primitives, and a is can learning parameter.
2-2, the convolution kernel size of all convolutional layers of depth network of structure are set to 3*3, the convolution kernel of last layer
It is dimensioned to 1*1；The convolution kernel of warp lamination is dimensioned to 4*4.In super-resolution network, the spy of preceding 19 convolutional layers
The quantity of sign figure is all set to 64, and the characteristic pattern quantity of last layer is 1；It is special corresponding to preceding 7 convolutional layers in coloured networks
The quantity of sign figure is set to 64,64,128,128,256,256,256, followed by 12 convolutional layers characteristic pattern
Quantity is arranged to 512, and the quantity of the characteristic pattern of 3 convolutional layers followed by is set as 256, the characteristic pattern of last layer of convolution
Quantity be set to 244.Every layer of configuration is specifically as shown in table 1 in network.
The network model configuration of the present invention of table 1
The output such as formula (8) that each convolutional layer and warp lamination obtain represents：
yi=f (Wixi+bi), i=1,2 ..., 43 (8)
Wherein, WiRepresent i-th layer of weight, biRepresent i-th layer of biasing, xiRepresent i-th layer of input, yiRepresent i-th layer
Output；
Pass through activation primitive ReLU and PReLU respectively, as a result as shown in formula (9) and (10)：
zi=max (0, yi) (9)
zj=max (0, yj)+a*min(0,yj) (10)
Wherein, yiAnd yjThe respectively output of activation primitive ReLU and PReLU last layers, ziAnd zjRespectively activation primitive
ReLU and PReLU output.
The network that step 3, the training set made according to step 1 and step 2 are built, adjusts network parameter, carries out network instruction
Practice, it is specific as follows：
3-1, using Caffe deep learning platforms network is trained, to the depth god of the multitask built in step 2
Through network, super-resolution network is initialized using MSRA modes first, and coloured networks are initialized as 0, biasing is all initial
Turn to 0.Detailed process is：
1) employed in super-resolution network after MSRA modes initialize weight W, W meets following Gaussian Profile：
Wherein, n represents the layer network input block number, i.e. convolutional layer input feature vector figure quantity.
And in coloured networks, weights initialisation is all set to 0, i.e. Wi=0.
2) in the entire network, biasing is all initialized as 0, i.e. bi=0.
3-2, network parameter updated using gradient descent method, represent as follows with formula (12)：
Vi+1=μ Vi-α▽L(Wi),Wi+1=Wi+Vi+1 (12)
Wherein, Vi+1Represent this weight updated value, and ViLast weight updated value is represented, and μ is last ladder
The weight of angle value, α are learning rates, ▽ L (Wi) it is gradient；
3-3, in the training process, network parameter renewal is carried out by given number of iterations.
After step 4, training terminate, using the gray level image of a width low resolution as the input of network, learnt using step 3
Obtained parameter removes to reconstruct a high-resolution coloured image as output.
The satellite image based on multitask deep neural network while super-resolution of the present embodiment and the method for coloring, consider
The problem of complicated and changeable into reality, it have chosen completely different both direction and image handled, not only meet more
The demand of business, and can realize to other any kind of images while realize super-resolution and coloring, the method is met
Complicated requirement in reality.In addition, two parts of the network not only each self-optimizing and also they by the way that feature is shared and feature
Interaction realizes collaboration optimization, to realize more preferable result.Also it is exactly not need artificial interference not only, and performs on the time
Greatly shorten, had a wide range of applications in the field such as historical photograph and remote sensing images.
Schematically the present invention and embodiments thereof are described above, this describes no restricted, institute in accompanying drawing
What is shown is also one of embodiments of the present invention, and actual structure is not limited thereto.So if common skill of this area
Art personnel are enlightened by it, without departing from the spirit of the invention, without designing and the technical scheme for creativeness
Similar frame mode and embodiment, protection scope of the present invention all should be belonged to.
Claims (9)
1. super-resolution and the method for coloring, its step are a kind of satellite image based on multitask deep neural network simultaneously：
Step 1, using satellite color image data collection, make the image block training set of high-resolution and low resolution；
Step 2, the deep neural network of one multitask of structure are used for model training；
The network that step 3, the training set made according to step 1 and step 2 are built, adjusts network parameter, carries out network training；
Step 4, the input using the gray level image of a width low resolution as network, the Reconstruction obtained using step 3 study
One high-resolution coloured image is as output.
2. a kind of satellite image based on multitask deep neural network according to claim 1 carry out simultaneously super-resolution and
The method of coloring, it is characterised in that：The process that step 1 makes the coloured image block training set of high-resolution and low resolution is：
Every coloured image is concentrated for a conventional satellite image processing data, high-definition picture is carried out first double twice
Cubic interpolation, obtain with high-definition picture corresponding to identical size low-resolution image；Then by every high-definition picture
Multiple images block is cut into low-resolution image, and intersection be present between adjacent image block, is thus used for
The high-definition picture block of depth network training and the set of low-resolution image block.
3. a kind of satellite image based on multitask deep neural network according to claim 1 or 2 carries out oversubscription simultaneously
The method distinguished and coloured, it is characterised in that：One 43 layers of depth network model is built in step 2, one is divided into three parts,
Image is pre-processed first, 20 layers of composition super-resolution network afterwards, last 23 layers of composition coloured networks；It is pre- in image
Process part, coloured image is transformed into Lab color spaces from RGB, coloured image is then divided into two parts, a part is L
Vector, the input as whole network；Another part be ab vector, the label as last coloured networks；In super-resolution network
In, 9 residual error layers and two convolutional layers are contained altogether, wherein each residual unit there are two convolutional layers；Each convolutional layer it
It is followed by a PReLU active coating；Shown in residual unit such as formula (1)：
yi=0.9*h (xi)+0.1*F(xi,wi)
xi+1=f (yi) (1)
Wherein, xiIt is expressed as the feature input of i-th layer of residual unit, wiIt is expressed as the setting of i-th layer of weight and bias term, F generations
Table residual error function, it is that identity maps h (x that f, which then represents activation primitive PReLU, h,i)=xi；
The depth network by learn low resolution gray level image block and high-resolution gray level image block between mapping relations,
As shown in formula (2)：
X=F (y, Φ) (2)
Wherein, x, y are respectively high-resolution gray level image block and the gray level image block of low resolution, and Φ is super-resolution network
The model parameter learnt, for follow-up high resolution image reconstruction；
In coloured networks, 4th layer reciprocal is warp lamination, and residue is convolutional layer；Each convolutional layer and warp lamination it
It is followed by a Relu active coating；Network inputs are the high-resolution gray level image blocks of super-resolution network output, and the network will be learned
The mapping relations between high-resolution gray level image block and ab color component images block are practised, as shown in formula (3)：
X=f (y, θ) (3)
Wherein, x, y are respectively ab color component images block and high-resolution gray level image block, and θ is what coloured networks learnt
Model parameter.
A kind of 4. satellite image based on multitask deep neural network according to claim 3 super-resolution and coloring simultaneously
Method, it is characterised in that：The loss function of network training uses different in super-resolution network and coloured networks in step 2
Method, in super-resolution network, loss function is represented using mean square error, as shown in formula (4)：
<mrow>
<msub>
<mi>L</mi>
<mn>1</mn>
</msub>
<mo>=</mo>
<munder>
<mi>min</mi>
<mi>&Phi;</mi>
</munder>
<mfrac>
<mn>1</mn>
<mi>N</mi>
</mfrac>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>N</mi>
</munderover>
<mo>|</mo>
<mo>|</mo>
<mi>F</mi>
<mrow>
<mo>(</mo>
<msup>
<mi>y</mi>
<mi>i</mi>
</msup>
<mo>,</mo>
<msub>
<mi>&Phi;</mi>
<mi>i</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>-</mo>
<msup>
<mi>x</mi>
<mi>i</mi>
</msup>
<mo>|</mo>
<msup>
<mo>|</mo>
<mn>2</mn>
</msup>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>4</mn>
<mo>)</mo>
</mrow>
</mrow>
Wherein, N be step 1 gained training set in sample size, xi,yiFor i-th of high-definition picture block and corresponding low resolution
Rate image block；
In coloured networks, loss function intersects entropy loss using multinomial and represented, as shown in formula (5)：
<mrow>
<msub>
<mi>L</mi>
<mn>2</mn>
</msub>
<mrow>
<mo>(</mo>
<mover>
<mi>Z</mi>
<mo>^</mo>
</mover>
<mo>,</mo>
<mi>Z</mi>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mo>-</mo>
<munder>
<mo>&Sigma;</mo>
<mrow>
<mi>h</mi>
<mo>,</mo>
<mi>w</mi>
</mrow>
</munder>
<mi>v</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>Z</mi>
<mrow>
<mi>h</mi>
<mo>,</mo>
<mi>w</mi>
</mrow>
</msub>
<mo>)</mo>
</mrow>
<munder>
<mo>&Sigma;</mo>
<mi>q</mi>
</munder>
<msub>
<mi>Z</mi>
<mrow>
<mi>h</mi>
<mo>,</mo>
<mi>w</mi>
<mo>,</mo>
<mi>q</mi>
</mrow>
</msub>
<mi>log</mi>
<mrow>
<mo>(</mo>
<msub>
<mover>
<mi>Z</mi>
<mo>^</mo>
</mover>
<mrow>
<mi>h</mi>
<mo>,</mo>
<mi>w</mi>
<mo>,</mo>
<mi>q</mi>
</mrow>
</msub>
<mo>)</mo>
</mrow>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>5</mn>
<mo>)</mo>
</mrow>
</mrow>
Wherein,Represent the probability distribution of prediction, and Z then represents real probability distribution, function v be rebalancing because
Son, it is obtained based on the statistics to training set ab color components, and h and w represent the length and width of image respectively, and q is then training set
The classification sum of middle ab color components.
A kind of 5. satellite image based on multitask deep neural network according to claim 4 super-resolution and coloring simultaneously
Method, it is characterised in that：The activation primitive of ReLU active coatings represents as follows with formula (6) in step 2：
F (x)=max (0, x) (6)
Wherein, x is the input of ReLU activation primitives, and f (x) is the output of ReLU activation primitives；
The activation primitive of PReLU active coatings represents as follows with formula (7)：
F (x)=max (0, x)+a*min (0, x) (7)
Wherein, x is the input of PReLU activation primitives, and f (x) is the output of PReLU activation primitives, and a is can learning parameter.
A kind of 6. satellite image based on multitask deep neural network according to claim 5 super-resolution and coloring simultaneously
Method, it is characterised in that：In step 2, except last layer of convolutional layer, the volume of all convolutional layers of the depth network of structure
Product core size is set to 3*3, and the convolution kernel of last layer is dimensioned to 1*1；The convolution kernel of warp lamination is dimensioned to 4*
4。
A kind of 7. satellite image based on multitask deep neural network according to claim 6 super-resolution and coloring simultaneously
Method, it is characterised in that：In super-resolution network, the quantity of the characteristic pattern of preceding 19 convolutional layers is all set to 64, finally
One layer of characteristic pattern quantity is 1；In coloured networks, the quantity of characteristic pattern corresponding to preceding 7 convolutional layers is set to 64,64,
128th, 128,256,256,256, followed by 12 convolutional layers be arranged to 512, the feature of 3 convolutional layers followed by
The quantity of figure is set as 256, and the quantity of the characteristic pattern of last layer of convolution is set to 244, and each convolutional layer and warp lamination obtains
Output such as formula (8) represent：
yi=f (Wixi+bi), i=1,2 ..., 43 (8)
Wherein, WiRepresent i-th layer of weight, biRepresent i-th layer of biasing, xiRepresent i-th layer of input, yiRepresent i-th layer defeated
Go out；
Pass through activation primitive ReLU and PReLU respectively, as a result as shown in formula (9) and (10)：
zi=max (0, yi) (9)
zj=max (0, yj)+a*min(0,yj) (10)
Wherein, yiAnd yjThe respectively output of activation primitive ReLU and PReLU last layers, ziAnd zjRespectively activation primitive ReLU and
PReLU output.
A kind of 8. satellite image based on multitask deep neural network according to claim 7 super-resolution and coloring simultaneously
Method, it is characterised in that：Step 3 is trained using Caffe deep learning platforms to network, first to being built in step 2
Multitask deep neural network weight and biasing initialized, detailed process is：
1) after initializing weight W using MSRA modes in super-resolution network, W meets following Gaussian Profile：
<mrow>
<mi>W</mi>
<mo>~</mo>
<mi>G</mi>
<mo>&lsqb;</mo>
<mn>0</mn>
<mo>,</mo>
<msqrt>
<mfrac>
<mn>2</mn>
<mi>n</mi>
</mfrac>
</msqrt>
<mo>&rsqb;</mo>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>11</mn>
<mo>)</mo>
</mrow>
</mrow>
Wherein, n represents the layer network input block number, i.e. convolutional layer input feature vector figure quantity；
In coloured networks, weights initialisation is all set to 0, i.e. Wi=0；
2) in the entire network, biasing is all initialized as 0, i.e. bi=0.
A kind of 9. satellite image based on multitask deep neural network according to claim 8 super-resolution and coloring simultaneously
Method, it is characterised in that：Step 3 represents as follows using gradient descent method renewal network parameter with formula (12)：
<mrow>
<msub>
<mi>V</mi>
<mrow>
<mi>i</mi>
<mo>+</mo>
<mn>1</mn>
</mrow>
</msub>
<mo>=</mo>
<msub>
<mi>&mu;V</mi>
<mi>i</mi>
</msub>
<mo>-</mo>
<mi>&alpha;</mi>
<mo>&dtri;</mo>
<mi>L</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>W</mi>
<mi>i</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>,</mo>
<msub>
<mi>W</mi>
<mrow>
<mi>i</mi>
<mo>+</mo>
<mn>1</mn>
</mrow>
</msub>
<mo>=</mo>
<msub>
<mi>W</mi>
<mi>i</mi>
</msub>
<mo>+</mo>
<msub>
<mi>V</mi>
<mrow>
<mi>i</mi>
<mo>+</mo>
<mn>1</mn>
</mrow>
</msub>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>12</mn>
<mo>)</mo>
</mrow>
</mrow>
Wherein, Vi+1Represent this weight updated value, and ViLast weight updated value is represented, and μ is last Grad
Weight, α is learning rate,It is gradient；
In the training process, network parameter renewal is carried out by given number of iterations.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711224807.3A CN107833183B (en) | 2017-11-29 | 2017-11-29 | Method for simultaneously super-resolving and coloring satellite image based on multitask deep neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711224807.3A CN107833183B (en) | 2017-11-29 | 2017-11-29 | Method for simultaneously super-resolving and coloring satellite image based on multitask deep neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107833183A true CN107833183A (en) | 2018-03-23 |
| CN107833183B CN107833183B (en) | 2021-05-25 |
Family
ID=61646641
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711224807.3A Active CN107833183B (en) | 2017-11-29 | 2017-11-29 | Method for simultaneously super-resolving and coloring satellite image based on multitask deep neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107833183B (en) |
Cited By (22)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108648144A (en) * | 2018-04-20 | 2018-10-12 | 南开大学 | A kind of FPM high-resolution colour picture method for reconstructing based on deep learning algorithm |
| CN108830912A (en) * | 2018-05-04 | 2018-11-16 | 北京航空航天大学 | A kind of interactive grayscale image color method of depth characteristic confrontation type study |
| CN108876870A (en) * | 2018-05-30 | 2018-11-23 | 福州大学 | A kind of domain mapping GANs image rendering methods considering texture complexity |
| CN108921783A (en) * | 2018-06-01 | 2018-11-30 | 武汉大学 | A kind of satellite image super resolution ratio reconstruction method based on losses by mixture function constraint |
| CN108921932A (en) * | 2018-06-28 | 2018-11-30 | 福州大学 | A method of the black and white personage picture based on convolutional neural networks generates various reasonable coloring in real time |
| CN109063565A (en) * | 2018-06-29 | 2018-12-21 | 中国科学院信息工程研究所 | A kind of low resolution face identification method and device |
| CN109146788A (en) * | 2018-08-16 | 2019-01-04 | 广州视源电子科技股份有限公司 | Super-resolution image reconstruction method and device based on deep learning |
| CN109191411A (en) * | 2018-08-16 | 2019-01-11 | 广州视源电子科技股份有限公司 | A kind of multitask image rebuilding method, device, equipment and medium |
| CN109360148A (en) * | 2018-09-05 | 2019-02-19 | 北京悦图遥感科技发展有限公司 | Based on mixing random down-sampled remote sensing image ultra-resolution ratio reconstructing method and device |
| CN109410114A (en) * | 2018-09-19 | 2019-03-01 | 湖北工业大学 | Compressed sensing image reconstruction algorithm based on deep learning |
| CN109961105A (en) * | 2019-04-08 | 2019-07-02 | 上海市测绘院 | A kind of Classification of High Resolution Satellite Images method based on multitask deep learning |
| CN110163801A (en) * | 2019-05-17 | 2019-08-23 | 深圳先进技术研究院 | A kind of Image Super-resolution and color method, system and electronic equipment |
| CN110288515A (en) * | 2019-05-27 | 2019-09-27 | 宁波大学 | The method and CNN coloring learner intelligently coloured to the microsctructural photograph of electron microscope shooting |
| CN110675462A (en) * | 2019-09-17 | 2020-01-10 | 天津大学 | Gray level image colorizing method based on convolutional neural network |
| CN110880163A (en) * | 2018-09-05 | 2020-03-13 | 南京大学 | Low-light color imaging method based on deep learning |
| CN111429350A (en) * | 2020-03-24 | 2020-07-17 | 安徽工业大学 | Rapid super-resolution processing method for mobile phone photographing |
| CN111627080A (en) * | 2020-05-20 | 2020-09-04 | 广西师范大学 | Gray level image coloring method based on convolution nerve and condition generation antagonistic network |
| CN111986084A (en) * | 2020-08-03 | 2020-11-24 | 南京大学 | Multi-camera low-illumination image quality enhancement method based on multi-task fusion |
| CN112489164A (en) * | 2020-12-07 | 2021-03-12 | 南京理工大学 | Image coloring method based on improved depth separable convolutional neural network |
| CN112508786A (en) * | 2020-12-03 | 2021-03-16 | 武汉大学 | Satellite image-oriented arbitrary-scale super-resolution reconstruction method and system |
| CN113537246A (en) * | 2021-08-12 | 2021-10-22 | 浙江大学 | Gray level image simultaneous coloring and hyper-parting method based on counterstudy |
| CN114463175A (en) * | 2022-01-18 | 2022-05-10 | 哈尔滨工业大学 | Mars image super-resolution method based on deep convolution neural network |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105354273A (en) * | 2015-10-29 | 2016-02-24 | 浙江高速信息工程技术有限公司 | Method for fast retrieving high-similarity image of highway fee evasion vehicle |
| CN106204449A (en) * | 2016-07-06 | 2016-12-07 | 安徽工业大学 | A kind of single image super resolution ratio reconstruction method based on symmetrical degree of depth network |
| US20170148139A1 (en) * | 2015-11-25 | 2017-05-25 | Heptagon Micro Optics Pte. Ltd. | Super-resolution image reconstruction using high-frequency band extraction |
-
2017
- 2017-11-29 CN CN201711224807.3A patent/CN107833183B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105354273A (en) * | 2015-10-29 | 2016-02-24 | 浙江高速信息工程技术有限公司 | Method for fast retrieving high-similarity image of highway fee evasion vehicle |
| US20170148139A1 (en) * | 2015-11-25 | 2017-05-25 | Heptagon Micro Optics Pte. Ltd. | Super-resolution image reconstruction using high-frequency band extraction |
| CN106204449A (en) * | 2016-07-06 | 2016-12-07 | 安徽工业大学 | A kind of single image super resolution ratio reconstruction method based on symmetrical degree of depth network |
Cited By (39)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108648144B (en) * | 2018-04-20 | 2021-12-10 | 南开大学 | FPM high-resolution color image reconstruction method based on deep learning algorithm |
| CN108648144A (en) * | 2018-04-20 | 2018-10-12 | 南开大学 | A kind of FPM high-resolution colour picture method for reconstructing based on deep learning algorithm |
| CN108830912A (en) * | 2018-05-04 | 2018-11-16 | 北京航空航天大学 | A kind of interactive grayscale image color method of depth characteristic confrontation type study |
| CN108876870B (en) * | 2018-05-30 | 2022-12-13 | 福州大学 | Domain mapping GANs image coloring method considering texture complexity |
| CN108876870A (en) * | 2018-05-30 | 2018-11-23 | 福州大学 | A kind of domain mapping GANs image rendering methods considering texture complexity |
| CN108921783A (en) * | 2018-06-01 | 2018-11-30 | 武汉大学 | A kind of satellite image super resolution ratio reconstruction method based on losses by mixture function constraint |
| CN108921783B (en) * | 2018-06-01 | 2022-04-15 | 武汉大学 | Satellite image super-resolution reconstruction method based on mixed loss function constraint |
| CN108921932A (en) * | 2018-06-28 | 2018-11-30 | 福州大学 | A method of the black and white personage picture based on convolutional neural networks generates various reasonable coloring in real time |
| CN108921932B (en) * | 2018-06-28 | 2022-09-23 | 福州大学 | Method for generating multiple reasonable colorings of black and white figure pictures based on convolutional neural network |
| CN109063565A (en) * | 2018-06-29 | 2018-12-21 | 中国科学院信息工程研究所 | A kind of low resolution face identification method and device |
| CN109063565B (en) * | 2018-06-29 | 2021-12-10 | 中国科学院信息工程研究所 | Low-resolution face recognition method and device |
| CN109191411A (en) * | 2018-08-16 | 2019-01-11 | 广州视源电子科技股份有限公司 | A kind of multitask image rebuilding method, device, equipment and medium |
| CN109146788A (en) * | 2018-08-16 | 2019-01-04 | 广州视源电子科技股份有限公司 | Super-resolution image reconstruction method and device based on deep learning |
| CN109191411B (en) * | 2018-08-16 | 2021-05-18 | 广州视源电子科技股份有限公司 | Multitask image reconstruction method, device, equipment and medium |
| CN109360148B (en) * | 2018-09-05 | 2023-11-07 | 北京悦图遥感科技发展有限公司 | Remote sensing image super-resolution reconstruction method and device based on mixed random downsampling |
| CN110880163A (en) * | 2018-09-05 | 2020-03-13 | 南京大学 | Low-light color imaging method based on deep learning |
| CN109360148A (en) * | 2018-09-05 | 2019-02-19 | 北京悦图遥感科技发展有限公司 | Based on mixing random down-sampled remote sensing image ultra-resolution ratio reconstructing method and device |
| CN110880163B (en) * | 2018-09-05 | 2022-08-19 | 南京大学 | Low-light color imaging method based on deep learning |
| CN109410114B (en) * | 2018-09-19 | 2023-08-25 | 湖北工业大学 | Compressed Sensing Image Reconstruction Algorithm Based on Deep Learning |
| CN109410114A (en) * | 2018-09-19 | 2019-03-01 | 湖北工业大学 | Compressed sensing image reconstruction algorithm based on deep learning |
| CN109961105A (en) * | 2019-04-08 | 2019-07-02 | 上海市测绘院 | A kind of Classification of High Resolution Satellite Images method based on multitask deep learning |
| WO2020233129A1 (en) * | 2019-05-17 | 2020-11-26 | 深圳先进技术研究院 | Image super-resolution and coloring method and system, and electronic device |
| CN110163801A (en) * | 2019-05-17 | 2019-08-23 | 深圳先进技术研究院 | A kind of Image Super-resolution and color method, system and electronic equipment |
| CN110163801B (en) * | 2019-05-17 | 2021-07-20 | 深圳先进技术研究院 | Image super-resolution and coloring method, system and electronic equipment |
| CN110288515A (en) * | 2019-05-27 | 2019-09-27 | 宁波大学 | The method and CNN coloring learner intelligently coloured to the microsctructural photograph of electron microscope shooting |
| CN110675462A (en) * | 2019-09-17 | 2020-01-10 | 天津大学 | Gray level image colorizing method based on convolutional neural network |
| CN111429350B (en) * | 2020-03-24 | 2023-02-24 | 安徽工业大学 | Rapid super-resolution processing method for mobile phone photographing |
| CN111429350A (en) * | 2020-03-24 | 2020-07-17 | 安徽工业大学 | Rapid super-resolution processing method for mobile phone photographing |
| CN111627080A (en) * | 2020-05-20 | 2020-09-04 | 广西师范大学 | Gray level image coloring method based on convolution nerve and condition generation antagonistic network |
| CN111627080B (en) * | 2020-05-20 | 2022-11-18 | 广西师范大学 | Gray level image coloring method based on convolution nerve and condition generation antagonistic network |
| CN111986084A (en) * | 2020-08-03 | 2020-11-24 | 南京大学 | Multi-camera low-illumination image quality enhancement method based on multi-task fusion |
| CN111986084B (en) * | 2020-08-03 | 2023-12-12 | 南京大学 | Multi-camera low-illumination image quality enhancement method based on multi-task fusion |
| CN112508786B (en) * | 2020-12-03 | 2022-04-29 | 武汉大学 | Satellite image-oriented arbitrary-scale super-resolution reconstruction method and system |
| CN112508786A (en) * | 2020-12-03 | 2021-03-16 | 武汉大学 | Satellite image-oriented arbitrary-scale super-resolution reconstruction method and system |
| CN112489164A (en) * | 2020-12-07 | 2021-03-12 | 南京理工大学 | Image coloring method based on improved depth separable convolutional neural network |
| CN112489164B (en) * | 2020-12-07 | 2023-07-04 | 南京理工大学 | Image coloring method based on improved depth separable convolutional neural network |
| CN113537246A (en) * | 2021-08-12 | 2021-10-22 | 浙江大学 | Gray level image simultaneous coloring and hyper-parting method based on counterstudy |
| CN114463175A (en) * | 2022-01-18 | 2022-05-10 | 哈尔滨工业大学 | Mars image super-resolution method based on deep convolution neural network |
| CN114463175B (en) * | 2022-01-18 | 2022-11-01 | 哈尔滨工业大学 | Mars image super-resolution method based on deep convolutional neural network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107833183B (en) | 2021-05-25 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107833183A (en) | A kind of satellite image based on multitask deep neural network while super-resolution and the method for coloring | |
| CN108564029B (en) | Face attribute recognition method based on cascade multitask learning deep neural network | |
| CN106920243A (en) | The ceramic material part method for sequence image segmentation of improved full convolutional neural networks | |
| Li et al. | Depth and surface normal estimation from monocular images using regression on deep features and hierarchical crfs | |
| CN109902798A (en) | The training method and device of deep neural network | |
| CN109711413A (en) | Image, semantic dividing method based on deep learning | |
| CN108010049A (en) | Split the method in human hand region in stop-motion animation using full convolutional neural networks | |
| CN109816725A (en) | A kind of monocular camera object pose estimation method and device based on deep learning | |
| CN106981080A (en) | Night unmanned vehicle scene depth method of estimation based on infrared image and radar data | |
| CN109410321A (en) | Three-dimensional rebuilding method based on convolutional neural networks | |
| CN111080511A (en) | End-to-end face exchange method for high-resolution multi-feature extraction | |
| CN107808129A (en) | A kind of facial multi-characteristic points localization method based on single convolutional neural networks | |
| CN110378985A (en) | A kind of animation drawing auxiliary creative method based on GAN | |
| CN105069746A (en) | Video real-time human face substitution method and system based on partial affine and color transfer technology | |
| CN107123089A (en) | Remote sensing images super-resolution reconstruction method and system based on depth convolutional network | |
| CN110163801A (en) | A kind of Image Super-resolution and color method, system and electronic equipment | |
| CN106780543A (en) | A kind of double framework estimating depths and movement technique based on convolutional neural networks | |
| CN106056155A (en) | Super-pixel segmentation method based on boundary information fusion | |
| CN109920012A (en) | Image colorant system and method based on convolutional neural networks | |
| CN109783887A (en) | A kind of intelligent recognition and search method towards Three-dimension process feature | |
| CN112489164B (en) | Image coloring method based on improved depth separable convolutional neural network | |
| Shen et al. | Machine learning assisted urban filling | |
| CN110378208A (en) | A kind of Activity recognition method based on depth residual error network | |
| CN111401156B (en) | Image identification method based on Gabor convolution neural network | |
| Pérez-Benito et al. | Smoothing vs. sharpening of colour images: Together or separated |
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 |