CN110533610A - The generation method and device of image enhancement model, application method and device - Google Patents
The generation method and device of image enhancement model, application method and device Download PDFInfo
- Publication number
- CN110533610A CN110533610A CN201910769010.4A CN201910769010A CN110533610A CN 110533610 A CN110533610 A CN 110533610A CN 201910769010 A CN201910769010 A CN 201910769010A CN 110533610 A CN110533610 A CN 110533610A
- Authority
- CN
- China
- Prior art keywords
- image
- model
- deep learning
- reinforced
- network
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 82
- 238000012549 training Methods 0.000 claims abstract description 205
- 238000013135 deep learning Methods 0.000 claims abstract description 140
- 230000002708 enhancing effect Effects 0.000 claims description 101
- 230000006870 function Effects 0.000 claims description 36
- 230000008569 process Effects 0.000 claims description 29
- 238000013527 convolutional neural network Methods 0.000 claims description 20
- 238000012545 processing Methods 0.000 claims description 11
- 230000004913 activation Effects 0.000 claims description 5
- 238000010586 diagram Methods 0.000 description 10
- 230000000694 effects Effects 0.000 description 10
- 230000004048 modification Effects 0.000 description 3
- 238000012986 modification Methods 0.000 description 3
- 241000208340 Araliaceae Species 0.000 description 2
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 2
- 235000003140 Panax quinquefolius Nutrition 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 238000002591 computed tomography Methods 0.000 description 2
- 238000004590 computer program Methods 0.000 description 2
- 238000002059 diagnostic imaging Methods 0.000 description 2
- 235000008434 ginseng Nutrition 0.000 description 2
- 238000013507 mapping Methods 0.000 description 2
- 241001269238 Data Species 0.000 description 1
- 230000006978 adaptation Effects 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 238000004422 calculation algorithm Methods 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 210000004218 nerve net Anatomy 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 230000003014 reinforcing effect Effects 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/70—Denoising; Smoothing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0012—Biomedical image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10072—Tomographic images
- G06T2207/10081—Computed x-ray tomography [CT]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Medical Informatics (AREA)
- Radiology & Medical Imaging (AREA)
- Quality & Reliability (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The embodiment of the present invention provides the generation method and device, application method and device of a kind of image enhancement model.In the embodiment of the present invention, by the way that deep learning network model is arranged, deep learning network model is made of the n times iteration of major network;Each input for deep learning network model, the input are the input of major network in iteration for the first time, and the output of major network is the input of major network in i+1 time iteration in i-th iteration, the major network parameter having the same in all iteration;And the value of the number of iterations N and the initial parameter value of major network are set, obtain several groups training data, deep learning network model is trained using training data, the deep learning network model that training finishes is obtained, as image enhancement model, shares network parameter using the network structure of iteration, reduce the quantity of overall network parameter, reduce the quantitative requirement for training data, it is therefore prevented that over-fitting effectively reduces model training difficulty.
Description
Technical field
The present invention relates to the generation method and device of technical field of image processing more particularly to a kind of image enhancement model,
Application method and device.
Background technique
Image enhancement (Image Enhancement) is an important directions in field of image processing, image denoising
(De-noise), super-resolution rebuilding (Super-resolution), deblurring (De-blur), filling (Inpaint) etc. all belong to
In the scope of image enhancement.
Currently, in the related technology, deep learning (Deep Learning) is applied to image enhancement.Its common practice
It is directly to be learnt from image to be reinforced to increasing using convolutional neural networks (Convolutional Neural Network, CNN)
The mapping function of strong image, is then handled image to be reinforced using mapping function, to obtain enhancing image.This technology
In, network unknown parameter is more, and this requires training datas also must be very much, could obtain preferable training effect.However
In practical application, it is frequently encountered the fewer situation of training data.For example, with goldstandard (Ground Truth, GT)
Medical image training data is fewer.In this way, the deficiency of training data easily causes over-fitting (Over-fit) problem,
So that expected requirement is not achieved in training effect, therefore, training is difficult.
Summary of the invention
To overcome the problems in correlation technique, the present invention provides the generation methods and dress of a kind of image enhancement model
It sets, application method and device, to obtain preferable training effect in the case where training data is less, the training for reducing model is difficult
Degree.
According to a first aspect of the embodiments of the present invention, a kind of generation method of image enhancement model, the method packet are provided
It includes:
Deep learning network model is set, and the deep learning network model is made of the n times iteration of major network;For
Each input of the deep learning network model, the input are the input of major network in iteration for the first time, main in i-th iteration
The output of volume grid is the input of major network in i+1 time iteration, the major network parameter having the same in all iteration;
And the value of the number of iterations N and the initial parameter value of the major network are set;
Obtain several groups training data;
The deep learning network model is trained using the training data, obtains the deep learning that training finishes
Network model, using the deep learning network model that the training finishes as image enhancement model;
Wherein, N, i are natural number, and N-1 >=i >=1.
According to a second aspect of the embodiments of the present invention, a kind of application method of image enhancement model, the method packet are provided
It includes:
The value of the number of iterations N of major network in the image enhancement model finished has been trained in adjustment, and the image that is applied increases
Strong model, it is the model generated according to the generation method of any one of first aspect image enhancement model that described image, which enhances model,;
Receive image to be reinforced;
The image to be reinforced is handled using application image enhancing model, obtains the image pair to be reinforced
The enhancing image answered.
According to a third aspect of the embodiments of the present invention, a kind of generating means of image enhancement model, described device packet are provided
It includes:
Setup module, for being arranged deep learning network model, the deep learning network model by major network n times
Iteration composition;Each input for the deep learning network model, the input are the input of major network in iteration for the first time,
The output of major network is the input of major network in i+1 time iteration in i-th iteration, the major network tool in all iteration
There is identical parameter;And the value of the number of iterations N and the initial parameter value of the major network are set;
Training data obtains module, for obtaining several groups training data;
Training module is trained for being trained using the training data to the deep learning network model
The deep learning network model finished;
Wherein, N, i are natural number, and N-1 >=i >=1.
According to a fourth aspect of the embodiments of the present invention, a kind of application apparatus of image enhancement model, described device packet are provided
It includes:
The number of iterations adjusts module, for adjusting the number of iterations for having trained major network in the image enhancement model finished
The value of N, be applied image enhancement model, and it is according to any one of first aspect image enhancement model that described image, which enhances model,
The model that generation method generates;
Image receiver module, for receiving image to be reinforced;
Image processing module is obtained for being handled using application image enhancing model the image to be reinforced
To the corresponding enhancing image of the image to be reinforced.
Technical solution provided in an embodiment of the present invention can include the following benefits:
In the embodiment of the present invention, by the way that deep learning network model is arranged, the deep learning network model is by main net
The n times iteration of network forms;Each input for the deep learning network model, the input are major network in iteration for the first time
Input, the output of major network is the input of major network in i+1 time iteration in i-th iteration, the master in all iteration
Volume grid parameter having the same;And the value of the number of iterations N and the initial parameter value of the major network are set, if obtaining
Dry group training data, is trained the deep learning network model using the training data, obtains the depth that training finishes
It spends learning network model and utilizes the net of iteration using the deep learning network model that the training finishes as image enhancement model
Network structure shares network parameter, reduces the quantity of overall network parameter, so that the quantitative requirement for training data is reduced,
The generation for effectively preventing over-fitting reduces the training difficulty of model, can obtain in the case where training data negligible amounts
Obtain preferable training effect.
It should be understood that above general description and following detailed description be only it is exemplary and explanatory, not
This specification can be limited.
Detailed description of the invention
The drawings herein are incorporated into the specification and forms part of this specification, and shows the reality for meeting this specification
Example is applied, and is used to explain the principle of this specification together with specification.
Fig. 1 is the flow example figure of the generation method of image enhancement model provided in an embodiment of the present invention.
Fig. 2 is the structure chart of deep learning network provided in an embodiment of the present invention.
Fig. 3 is the topology example figure of convolutional neural networks.
Fig. 4 is the flow example figure of the application method of image enhancement model provided in an embodiment of the present invention.
Fig. 5 is the application exemplary diagram of the application method of image enhancement model provided in an embodiment of the present invention.
Fig. 6 is the functional block diagram of the generating means of image enhancement model provided in an embodiment of the present invention.
Fig. 7 is the functional block diagram of the application apparatus of image enhancement model provided in an embodiment of the present invention.
Fig. 8 is a hardware structure diagram of electronic equipment provided in an embodiment of the present invention.
Fig. 9 is another hardware structure diagram of electronic equipment provided in an embodiment of the present invention.
Specific embodiment
Example embodiments are described in detail here, and the example is illustrated in the accompanying drawings.Following description is related to
When attached drawing, unless otherwise indicated, the same numbers in different drawings indicate the same or similar elements.Following exemplary embodiment
Described in embodiment do not represent all embodiments consistented with the present invention.On the contrary, they be only with it is such as appended
The example of the consistent device and method of some aspects be described in detail in claims, the embodiment of the present invention.
It in the term that the embodiment of the present invention uses and is not intended to merely for describing the purpose of the specific embodiment of the present invention
Limit the embodiment of the present invention.In the embodiment of the present invention and the "an" of singular used in the attached claims, " institute
State " and "the" be also intended to including most forms, unless the context clearly indicates other meaning.It is also understood that making herein
Term "and/or" refers to and may combine comprising one or more associated any or all of project listed.
It will be appreciated that though various letters may be described using term first, second, third, etc. in the embodiment of the present invention
Breath, but these information should not necessarily be limited by these terms.These terms are only used to for same type of information being distinguished from each other out.For example,
In the case where not departing from range of embodiment of the invention, the first information can also be referred to as the second information, similarly, the second information
The first information can also be referred to as.Depending on context, word as used in this " if " can be construed to " ...
When " or " when ... " or " in response to determination ".
One typical case scene of image enhancement model is that the image generated to medical imaging device carries out enhancing processing.
CT scan (Computed Tomography, CT) image is carried out for example, can use image enhancement model
Denoising, deblurring processing etc..
Certainly, above-mentioned application scenarios are only an example of the applicable scene of image enhancement model, are not used to figure
The application scenarios of image intensifying model are limited.In addition to above-mentioned application scenarios, generated using method provided in an embodiment of the present invention
Image enhancement model can be also used for other application scenarios, for example, image enhancement model can be also used for biomedical
Microoptic image is handled, and is handled etc. the image that aerospace vehicle is passed back.
The training process and application process of image enhancement model are illustrated respectively below by embodiment.
Fig. 1 is the flow example figure of the generation method of image enhancement model provided in an embodiment of the present invention.The embodiment is retouched
The training process of image enhancement model is stated.In the training process, deep learning network model, the deep learning net are constructed in advance
Network model is made of the n times iteration of major network (Body Network), and the output of major network is i+1 in i-th iteration
The input of major network in secondary iteration, the major network parameter having the same in all iteration, as shown in Figure 1, image enhancement
The generation method of model may include:
S101, is arranged deep learning network model, and deep learning network model is made of the n times iteration of major network;It is right
In each input of deep learning network model, which is the input of major network in iteration for the first time, main body in i-th iteration
The output of network is the input of major network in i+1 time iteration, the major network parameter having the same in all iteration;With
And the value of the number of iterations N and the initial parameter of major network are set.Wherein, N, i are natural number, and N-1 >=i >=1.Value
S102 obtains several groups training data.
S103 is trained deep learning network model using training data, obtains the deep learning net that training finishes
Network model, using the deep learning network model that training finishes as image enhancement model.
In the embodiment of the present invention, deep learning network is that the n times iteration of main volume grid is formed by network.
Fig. 2 is the structure chart of the deep learning network model of the embodiment of the present invention.Fig. 2 is referred to, major network passes through n times
Iteration constitutes deep learning network model.Using same major network when due to each iteration, in all iteration
Major network share identical parameter, thus significantly reduce the number of parameters of whole network, thus reduce for instruction
The quantitative requirement for practicing data, effectively prevents the generation of over-fitting, reduces the training difficulty of model, can be in training data number
Measure it is less in the case where obtain preferable training effect.
For ease of understanding, structure shown in Fig. 2 also can be regarded as the network model of N number of same body cascade composition.
Unlike common cascade structure network model, N number of major network is identical in Fig. 2 institute representation model, i.e. this N number of main body
Network has identical parameter and parameter value.
Here the number of parameters of Fig. 2 institute's representation model and common cascade structure network model is illustrated.Assuming that
Each major network has n parameter, then the parameter of Fig. 2 institute representation model includes the parameter (n) and major network of major network
The number of iterations N, in the case where N value determines, the number of Fig. 2 institute representation model is n.If it is common cascade structure network mould
The number of parameters of type, general network model made of N number of major network cascade is n × N.
For example, common cascade structure network model is by three different subjects network As (x), B (x), C (x) sequentially cascaded series
At then common cascade structure network model Y1=C (B (A (x))).And by three same body network F in Fig. 2 institute representation model
(x) when cascading, Fig. 2 institute's representation model is Y2=F (F (F (x))).Assuming that A (x), B (x), C (x), F (x) have n ginseng
Number, then the number of parameters of network model Y1 is 3n, then the number of parameters of network model Y2 is n.
As it can be seen that compared to common cascade structure network model, the ginseng of the deep learning network model of the embodiment of the present invention
Number quantity substantially reduces.The amount of training data as needed for training pattern is directly proportional to the number of parameters of model, in parameter
In the case that quantity substantially reduces, amount of training data needed for training pattern is greatly decreased.
Further, iteration structure used by the deep learning network model of the embodiment of the present invention also makes model training
The convergence rate of process faster, so as to the time needed for reducing training process.
N can be set according to specific requirements.
It in one example, can be according to the difference of image to be reinforced in training data known enhancing image corresponding with its
Size determines the value of N.For example, when the diversity ratio of image to be reinforced known enhancing image corresponding with its is larger, can take compared with
Value of the big value as N;Conversely, when the comparison in difference of image to be reinforced known enhancing image corresponding with its is small, Ke Yiqu
Value of the lesser value as N.
In another example, the value of N can be determined according to the structure of major network.For example, when major network uses
When the convolutional neural networks of cascade structure shown in Fig. 3, if cascade structure quantity (can be indicated with the number of plies of convolutional layer) compared with
It is few, biggish value can be taken as the value of N, if cascade structure quantity is more, lesser value can be taken as the value of N.
Wherein, major network can use convolutional neural networks, can also be using recurrent neural network, circulation nerve net
Network, adaptive learning algorithm network etc..
When major network uses convolutional neural networks, the structure of convolutional neural networks can be as shown in Figure 3.Refer to figure
3, convolutional neural networks can be cascaded by multiple convolutional layers and activation primitive.Certainly, as the convolutional Neural of major network
Network can also use the other structures different from Fig. 3, specific knot of the present invention to the convolutional neural networks as major network
Structure is with no restriction.
In one example, every group of training data in several groups training data includes image to be reinforced and image to be reinforced
Corresponding first enhancing image;
Deep learning network model is trained using training data, obtains the deep learning network model that training finishes
It is trained, comprising:
In the training process, the parameter value of major network is in the corresponding deep learning network model of the 1st group of training data
The initial parameter value, the parameter value of major network is through jth -1 in the corresponding deep learning network model of jth group training data
The parameter value that adjusts after group training data training, j are natural number, and j >=2;For every group of training data, perform the following operations:
Image to be reinforced in this group of training data is inputted into the corresponding deep learning network model of this group of training data, is obtained
To the corresponding second enhancing image of this group of training data;
Obtain the difference value of the first enhancing image and the second enhancing image;
If the difference value is greater than preset threshold, main body in the deep learning network model is adjusted according to the difference value
The parameter value of network;If the difference value is less than or equal to the preset threshold, deconditioning is corresponding with this group of training data
The deep learning network model that deep learning network model is finished as training.
As it can be seen that in the training process, the parameter value of deep learning network model is all when each group of training data is trained
Parameter value after being completed upper one group of training data training of training.That is, depth when each group of training data training
The parameter value for practising network model is all different, n-th of deep learning net of the image to be reinforced input in n-th group training data
The parameter value of network model is by determining after (n-1)th group of training data training.Under the premise of N value is set, the present invention is implemented
The parameter value of deep learning network model refers to the parameter value of major network in deep learning network model in the training process of example.
The parameter value of major network is arranged in step s101 in the 1st deep learning network model in training process, in training
In the process, on the basis of can be using the parameter value that training data is arranged herein, constantly the parameter value of major network be adjusted
It is whole, until meeting the preset condition of convergence, complete training.
On aforementioned base, during an illustrative realization, the method can also include: preparatory building loss
Function, the loss function include corresponding with first corresponding first variable of enhancing image and with the second enhancing image
The second variable;
Obtain the difference value of the first enhancing image and the second enhancing image, comprising:
Using the first enhancing image as the value of the first variable in the loss function, made with the second enhancing image
For bivariate value in the loss function, the value of loss function is calculated, using the value of loss function as first enhancing
The difference value of image and the second enhancing image.
During an illustrative realization, main body in the deep learning network model is adjusted according to the difference value
The parameter value of network, comprising:
Difference value is returned into the every of deep learning network model by back propagation (Back Propagation, BP)
One layer, the parameter value of major network in deep learning network model is modified.
Training process shown in step S103 is exemplified below.
Assuming that training data group includes (Iin1, Igt1)、(Iin2, Igt2)、……(Iin100, Igt100), in this example, IinFor
Image to be reinforced, IgtFor image I to be reinforcedinCorresponding first enhancing image, IoutImage I to be reinforcedinInput deep learning net
The second enhancing image that model exports after network model, the number of iterations N=2, major network is in n-th of deep learning network model
Fn(x), then training process is as follows:
(1) by the image I to be reinforced in the 1st group of training datain1The 1st deep learning network model is inputted, I is obtainedin1
Corresponding second enhancing image Iout1, process is: by Iin1Input major network F1(x), image F is obtained1(Iin1), due to repeatedly
Generation number N=2, it is therefore desirable to by image F1(Iin1) major network F is inputted again1(x), image I is obtainedout1=F1(F1
(Iin1));
(2) according to preset loss function, image I is calculatedout1With image Igt1Difference value δ1;
(3) judge δ1Whether threshold value m is less than or equal to, if δ1Greater than threshold value m, by δ1Back to the 1st deep learning net
Network model F1(F1(x)) each layer, to F1(x) parameter value is adjusted, and major network becomes F after adjustment2(x);
(4) by the image I to be reinforced in the 2nd group of training datain2The 2nd deep learning network model is inputted, I is obtainedin2
Corresponding second enhancing image Iout2, process is: by Iin2Input major network F2(x), image F is obtained2(Iin2), due to repeatedly
Generation number N=2, it is therefore desirable to by image F2(Iin2) major network F is inputted again2(x), image I is obtainedout2=F2(F2
(Iin2));
(5) according to preset loss function, image I is calculatedout2With image Igt2Difference value δ2;
(6) judge δ2Whether threshold value m is less than or equal to, if δ2Greater than threshold value m, by δ2Back to the 2nd deep learning net
Network model F2(F2(x)) each layer, to F2(x) parameter value is adjusted, and major network becomes F after adjustment3(x);
……
In this way, repeating step (1)~(3), until δ is less than or equal to threshold value m, terminate training.
Wherein, the parameter value of major network when difference value is less than or equal to preset threshold, that is, the master determined at the end of training
The parameter value of volume grid.So far, the parameter value of whole parameters (parameter and N including major network) of deep learning network model
It all determines, is image enhancement model with the deep learning network model that these parameter values determine.
Embodiment illustrated in fig. 1, by the way that deep learning network model is arranged, the deep learning network model is by major network
N times iteration composition;Each input for the deep learning network model, the input are major network in iteration for the first time
It inputs, the output of major network is the input of major network in i+1 time iteration in i-th iteration, the main body in all iteration
Network parameter having the same;And the value of the number of iterations N and the initial parameter value of the major network are set, it obtains several
Group training data, is trained the deep learning network model using the training data, obtains the depth that training finishes
Learning network model utilizes the network of iteration using the deep learning network model that the training finishes as image enhancement model
Structure shares network parameter, reduces the quantity of overall network parameter, to reduce the quantitative requirement for training data, has
Effect prevents the generation of over-fitting, reduces the training difficulty of model, can obtain in the case where training data negligible amounts
Preferable training effect.
Fig. 4 is the flow example figure of the application method of image enhancement model provided in an embodiment of the present invention.The embodiment is retouched
The application process of image enhancement model is stated.As shown in figure 4, the application method of image enhancement model may include:
S401 adjusts the value of the number of iterations N of major network in trained image enhancement model, and be applied image
Enhance model, image enhancement model is the model generated according to the generation method of the image enhancement model of the embodiment of the present invention.
S402 receives image to be reinforced.
S403 handles image to be reinforced using application image enhancing model, obtains the corresponding increasing of image to be reinforced
Strong image.
In step S401, image enhancement model is the generation side using image enhancement model provided in an embodiment of the present invention
The model that method generates, such as the image enhancement model that embodiment illustrated in fig. 1 generates.
Wherein, major network can be convolutional neural networks, such as the convolutional neural networks can use knot shown in Fig. 3
Structure.
It should be noted that in application process, the parameter value of major network is fixed in image enhancement model, i.e., it is logical
The major network parameter that the training process crossed in the generation method of the image enhancement model of the embodiment of the present invention determines, still, figure
The number of iterations of major network can be identical as the number of iterations of the image enhancement model trained in image intensifying model, can also be with
It is different.By default, the number of iterations of major network can use and training mould in image enhancement model in application process
The identical number of the number of iterations when type.For example, the number of iterations when training pattern is 5, then by default, when application
The number of iterations of major network is also 5 in image enhancement model.
If in use, image to be reinforced from training when training data in picture quality to be reinforced it is different,
In application, in adjustable pre-generated image enhancement model the number of iterations N of major network value.
Therefore, during an illustrative realization, in step S401, adjustment is in trained image enhancement model
The value of the number of iterations N of major network, be applied image enhancement model, may include:
Using image to be reinforced to be processed as the first image to be reinforced, trained number used when with training image enhancing model
Image to be reinforced in compares the image matter of the first image to be reinforced and the second image to be reinforced as the second image to be reinforced
Amount;
If the picture quality of the first image to be reinforced is better than the picture quality of the second image to be reinforced, setting the first iteration time
Number is less than secondary iteration number;If the picture quality of the second image to be reinforced is better than the picture quality of the first image to be reinforced, if
The first the number of iterations is set greater than secondary iteration number;If the picture quality of the second image to be reinforced and the described first image to be reinforced
Picture quality it is identical, setting the first the number of iterations be equal to secondary iteration number;Wherein, the first the number of iterations is application image increasing
The value of the number of iterations N of major network in strong model, secondary iteration number is to have trained main body in the image enhancement model finished
The value of the number of iterations N of network.
By this example, model when institute can be enhanced according to image to be reinforced and training image to be processed in application process
The value of the number of iterations N is adjusted with the quality difference of the image to be reinforced in training data, so that image enhancement model is not
It is only applicable to the image to be reinforced identical in quality with the image to be reinforced in training data used when training image enhancing model,
The to be reinforced image different from the quality of image to be reinforced in training data used when training image enhancing model is applied also for,
Therefore the use scope of image enhancement model is expanded.
It should be noted that although N value takes the natural number more than or equal to 2 in the training process, in application process
In, N value can be set to 1.
As it can be seen that using when image enhancement model major network use the number of iterations can be with iteration when training pattern
Number is different.For example, the number of iterations when training pattern is 5, and by setting, main body in image enhancement model when can will apply
The number of iterations of network is revised as 2,3,4,6,7 etc..
When the number of iterations difference, enhancing image is treated using image enhancement model and does not carry out the effect of image enhancement also not
Together.
For example, it is assumed that the number of iterations when training pattern is 5, in the application, if low to the quality requirement of enhancing image
The number of iterations can be set below 5 value, both met to increasing in this way by the quality for having enhanced image when training pattern
The quality requirement of strong image, and calculation amount can be reduced by reducing the number of iterations, to reduce the processing load of system.
It for another example, can be with if the quality for having enhanced image when being higher than training pattern to the quality requirement of enhancing image
The number of iterations is set above to 5 value, re -training model is not needed in this way and is just able to satisfy to the quality requirement for enhancing image,
Avoid the trouble of re -training model.
In step S402, image to be reinforced can be original image untreated in concrete application scene.For example,
Image to be reinforced can be the medical image that medical imaging device scanning obtains, such as CT image.
In step S402, image to be reinforced input application image enhancing model is handled, application image enhances mould
The output result obtained after type processing is the corresponding enhancing image of image to be reinforced.
It illustrates below and the application process of image enhancement model in embodiment illustrated in fig. 4 is illustrated.
Assuming that utilizing training data group (I according to aforementioned exemplaryin1, Igt1)、(Iin2, Igt2)、……(Iin100, Igt100) instruction
Major network is F in the image enhancement model practised98(x), N=2.
By the image I to be reinforced in applicationdWith the image I to be reinforced in training data groupin1、Iin2……Iin100Compared
Compared with discovery image IdQuality be substantially less than image Iin1、Iin2……Iin100Quality, therefore will application image enhance model in
The value of N is set as 3, and the major network in application image enhancing model is the major network F in the image enhancement model trained98
(x), then application image enhances model to image IdTreatment process it is as follows:
(1) by image IdInput the major network F in application image enhancing model98(x), F is obtained98(Id);
(2) by F98(Id) input data is used as to input major network F again98(x), the 2nd iteration is carried out, F is obtained98(F98
(Id));
(3) by F98(F98(Id)) input data is used as to input major network F again98(x), the 3rd iteration is carried out, is obtained
F98(F98(F98(Id))), so far obtain image IdCorresponding enhancing image F98(F98(F98(Id)))。
Fig. 5 is the application exemplary diagram of the application method of image enhancement model provided in an embodiment of the present invention.In this example, benefit
Denoising is carried out to human body low-dose CT image with the application method of image enhancement model provided in an embodiment of the present invention.It please join
See Fig. 5, number first image in the left side is the human body low-dose CT image before denoising in Fig. 5, and number second figure in the left side is left side number
The enhancing image that first image obtains after the slight denoising of 1 iteration, left side number third image is left side number first
A image is the enhancing image obtained after the reinforcing denoising of 3 iteration.As seen from Figure 5, this example is by adjusting iteration time
Number, realizes different degrees of image enhancement effects.
Embodiment illustrated in fig. 4, by adjusting the number of iterations N of major network in trained image enhancement model
Value, be applied image enhancement model, and image enhancement model is the generation side according to the image enhancement model of the embodiment of the present invention
The model that method generates, receives image to be reinforced, is handled using application image enhancing model image to be reinforced, is obtained wait increase
The corresponding enhancing image of strong image can realize different degrees of image enhancement effects by adjusting the number of iterations, and can
Image enhancement processing is carried out to the image to be reinforced of different quality, to improve the flexibility and the scope of application of method.
Based on above-mentioned embodiment of the method, the embodiment of the present invention also provides corresponding devices, equipment and storage medium are real
Apply example.
Fig. 6 is the functional block diagram of the generating means of image enhancement model provided in an embodiment of the present invention.In the present embodiment,
The generating means of image enhancement model construct deep learning network model in advance, deep learning network model by major network N
Secondary iteration composition, the output of major network is the input of major network in i+1 time iteration in i-th iteration, in all iteration
Major network parameter having the same, as shown in fig. 6, in the present embodiment, the generating means of image enhancement model may include:
Setup module 610, for deep learning network model to be arranged, the deep learning network model is by major network
N times iteration composition;Each input for the deep learning network model, the input be for the first time in iteration major network it is defeated
Enter, the output of major network is the input of major network in i+1 time iteration in i-th iteration, the main net in all iteration
Network parameter having the same;And the value of the number of iterations N and the initial parameter value of the major network are set;
Training data obtains module 620, for obtaining several groups training data;
Training module 630 is instructed for being trained using the training data to the deep learning network model
Practice the deep learning network model finished;
Wherein, N, i are natural number, and N-1 >=i >=1.
During an illustrative realization, every group of training data in the several groups training data includes to be reinforced
Image and the corresponding first enhancing image of the image to be reinforced;
The training module 630 is being used to be trained the deep learning network model using the training data,
When obtaining that the deep learning network model finished is trained to be trained, it is specifically used for:
In the training process, the parameter value of major network is in the corresponding deep learning network model of the 1st group of training data
The initial parameter value, the parameter value of major network is through jth -1 in the corresponding deep learning network model of jth group training data
The parameter value that adjusts after group training data training, j are natural number, and j >=2;For every group of training data, perform the following operations:
Image to be reinforced in this group of training data is inputted into the corresponding deep learning network model of this group of training data, is obtained
To the corresponding second enhancing image of this group of training data;
Obtain the difference value of the first enhancing image and the second enhancing image;
If the difference value is greater than preset threshold, main body in the deep learning network model is adjusted according to the difference value
The parameter value of network;If the difference value is less than or equal to the preset threshold, deconditioning is corresponding with this group of training data
The deep learning network model that deep learning network model is finished as training.
During an illustrative realization, further includes:
Function constructs module, and for constructing loss function in advance, the loss function includes and the first enhancing image
Corresponding first variable and the second variable corresponding with the second enhancing image;
The training module 630 is in the difference value for obtaining the first enhancing image and the second enhancing image
When, it is specifically used for:
Using the first enhancing image as the value of the first variable in the loss function, made with the second enhancing image
For bivariate value in the loss function, the value of loss function is calculated, using the value of loss function as first enhancing
The difference value of image and the second enhancing image.
During an illustrative realization, the training module 630 is for according to difference value adjustment
In deep learning network model when the parameter value of major network, it is specifically used for:
Each layer that the difference value is returned to deep learning network model by back propagation, to deep learning net
The parameter value of major network is modified in network model.
During an illustrative realization, the major network is convolutional neural networks.
During an illustrative realization, the convolutional neural networks are to be cascaded by multiple convolutional layers and activation primitive
Made of network.
Fig. 7 is the functional block diagram of the application apparatus of image enhancement model provided in an embodiment of the present invention.As shown in fig. 7,
In the present embodiment, the application apparatus of image enhancement model may include:
The number of iterations adjusts module 710, for adjusting the iteration time of major network in trained image enhancement model
The value of number N, be applied image enhancement model, and image enhancement model is according to any image enhancement model of the embodiment of the present invention
Generation method generate model;
Image receiver module 720, for receiving image to be reinforced;
Image processing module 730 is obtained for being handled using application image enhancing model image to be reinforced wait increase
The corresponding enhancing image of strong image.
During an illustrative realization, the number of iterations adjustment module 710 includes:
Comparing unit, for enhancing mould with training image using image to be reinforced to be processed as the first image to be reinforced
Image to be reinforced when type in training data used compares the first image to be reinforced with second wait increase as the second image to be reinforced
The picture quality of strong image;
Setting unit, if being better than the picture quality of the second image to be reinforced for the picture quality of the first image to be reinforced,
First the number of iterations is set less than secondary iteration number;If the picture quality of the second image to be reinforced is better than the first image to be reinforced
Picture quality, setting the first the number of iterations be greater than secondary iteration number;If the picture quality and first of the second image to be reinforced
The picture quality of image to be reinforced is identical, and the first the number of iterations of setting is equal to secondary iteration number;Wherein, the first the number of iterations is
Application image enhances the value of the number of iterations N of major network in model, and secondary iteration number is to have trained the image enhancement finished
The value of the number of iterations N of major network in model.
The embodiment of the invention also provides a kind of electronic equipment.Fig. 8 is the one of electronic equipment provided in an embodiment of the present invention
A hardware structure diagram.As shown in figure 8, electronic equipment includes: internal bus 801, and the memory connected by internal bus
802, processor 803 and external interface 804, the electronic equipment construct deep learning network model in advance, the deep learning net
Network model is made of the n times iteration of major network (Body Network), and the output of major network is i+1 in i-th iteration
The input of major network in secondary iteration, the major network parameter having the same in all iteration, wherein
The processor 803 for reading the machine readable instructions on memory 802, and executes described instruction to realize
Following operation:
Deep learning network model is set, and the deep learning network model is made of the n times iteration of major network;For
Each input of the deep learning network model, the input are the input of major network in iteration for the first time, main in i-th iteration
The output of volume grid is the input of major network in i+1 time iteration, the major network parameter having the same in all iteration;
And the value of the number of iterations N and the initial parameter value of the major network are set;
Obtain several groups training data;
The deep learning network model is trained using the training data, obtains the deep learning that training finishes
Network model, using the deep learning network model that the training finishes as image enhancement model;
Wherein, N, i are natural number, and N-1 >=i >=1.
During an illustrative realization, every group of training data in the several groups training data includes to be reinforced
Image and the corresponding first enhancing image of the image to be reinforced;
It is described that the deep learning network model is trained using the training data, obtain the depth that training finishes
Learning network model is trained, comprising:
In the training process, the parameter value of major network is in the corresponding deep learning network model of the 1st group of training data
The initial parameter value, the parameter value of major network is through jth -1 in the corresponding deep learning network model of jth group training data
The parameter value that adjusts after group training data training, j are natural number, and j >=2;For every group of training data, perform the following operations:
Image to be reinforced in this group of training data is inputted into the corresponding deep learning network model of this group of training data, is obtained
To the corresponding second enhancing image of this group of training data;
Obtain the difference value of the first enhancing image and the second enhancing image;
If the difference value is greater than preset threshold, main body in the deep learning network model is adjusted according to the difference value
The parameter value of network;If the difference value is less than or equal to the preset threshold, deconditioning is corresponding with this group of training data
The deep learning network model that deep learning network model is finished as training.
During an illustrative realization, further includes:
Building loss function in advance, the loss function include with first corresponding first variable of enhancing image and with
Described second enhances corresponding second variable of image;
The difference value for obtaining the first enhancing image and the second enhancing image, comprising:
Using the first enhancing image as the value of the first variable in the loss function, made with the second enhancing image
For bivariate value in the loss function, the value of loss function is calculated, using the value of loss function as first enhancing
The difference value of image and the second enhancing image.
It is described to be adjusted in the deep learning network model according to the difference value during an illustrative realization
The parameter value of major network, comprising:
Each layer that the difference value is returned to deep learning network model by back propagation, to deep learning net
The parameter value of major network is modified in network model.
During an illustrative realization, the major network is convolutional neural networks.
During an illustrative realization, the convolutional neural networks are to be cascaded by multiple convolutional layers and activation primitive
Made of network.
The embodiment of the invention also provides a kind of electronic equipment.Fig. 9 is the another of electronic equipment provided in an embodiment of the present invention
One hardware structure diagram.As shown in figure 9, electronic equipment includes: internal bus 901, and the storage connected by internal bus
Device 902, processor 903 and external interface 904, wherein
The processor 903 for reading the machine readable instructions on memory 902, and executes described instruction to realize
Following operation:
The value of the number of iterations N of major network in the image enhancement model finished has been trained in adjustment, and the image that is applied increases
Strong model, it is the model generated according to the generation method of aforementioned any image enhancement model that described image, which enhances model,;
Receive image to be reinforced;
The image to be reinforced is handled using application image enhancing model, obtains the image pair to be reinforced
The enhancing image answered.
During an illustrative realization, major network in the image enhancement model finished has been trained in the adjustment
The value of the number of iterations N, be applied image enhancement model, comprising:
Using image to be reinforced to be processed as the first image to be reinforced, instruction used when training described image enhancing model
Practice the image to be reinforced in data as the second image to be reinforced, first image to be reinforced and described second to be reinforced
The picture quality of image;
If the picture quality of first image to be reinforced is better than the picture quality of the described second image to be reinforced, setting the
One the number of iterations is less than secondary iteration number;If the picture quality of second image to be reinforced is better than the described first figure to be reinforced
The picture quality of picture, the first the number of iterations of setting are greater than secondary iteration number;If the picture quality of second image to be reinforced
Identical as the picture quality of the described first image to be reinforced, the first the number of iterations of setting is equal to secondary iteration number;Wherein, described
First the number of iterations is the value that application image enhances the number of iterations N of major network in model, and the secondary iteration number is
The value of the number of iterations N of major network in the image enhancement model that training finishes.
The embodiment of the present invention also provides a kind of computer readable storage medium, is stored thereon with computer program, wherein institute
It states and realizes following operation when program is executed by processor:
Deep learning network model is set, and the deep learning network model is made of the n times iteration of major network;For
Each input of the deep learning network model, the input are the input of major network in iteration for the first time, main in i-th iteration
The output of volume grid is the input of major network in i+1 time iteration, the major network parameter having the same in all iteration;
And the value of the number of iterations N and the initial parameter value of the major network are set;
Obtain several groups training data;
The deep learning network model is trained using the training data, obtains the deep learning that training finishes
Network model, using the deep learning network model that the training finishes as image enhancement model;
Wherein, N, i are natural number, and N-1 >=i >=1.
During an illustrative realization, every group of training data in the several groups training data includes to be reinforced
Image and the corresponding first enhancing image of the image to be reinforced;
It is described that the deep learning network model is trained using the training data, obtain the depth that training finishes
Learning network model is trained, comprising:
In the training process, the parameter value of major network is in the corresponding deep learning network model of the 1st group of training data
The initial parameter value, the parameter value of major network is through jth -1 in the corresponding deep learning network model of jth group training data
The parameter value that adjusts after group training data training, j are natural number, and j >=2;For every group of training data, perform the following operations:
Image to be reinforced in this group of training data is inputted into the corresponding deep learning network model of this group of training data, is obtained
To the corresponding second enhancing image of this group of training data;
Obtain the difference value of the first enhancing image and the second enhancing image;
If the difference value is greater than preset threshold, main body in the deep learning network model is adjusted according to the difference value
The parameter value of network;If the difference value is less than or equal to the preset threshold, deconditioning is corresponding with this group of training data
The deep learning network model that deep learning network model is finished as training.
During an illustrative realization, further includes:
Building loss function in advance, the loss function include with first corresponding first variable of enhancing image and with
Described second enhances corresponding second variable of image;
The difference value for obtaining the first enhancing image and the second enhancing image, comprising:
Using the first enhancing image as the value of the first variable in the loss function, made with the second enhancing image
For bivariate value in the loss function, the value of loss function is calculated, using the value of loss function as first enhancing
The difference value of image and the second enhancing image.
It is described to be adjusted in the deep learning network model according to the difference value during an illustrative realization
The parameter value of major network, comprising:
Each layer that the difference value is returned to deep learning network model by back propagation, to deep learning net
The parameter value of major network is modified in network model.
During an illustrative realization, the major network is convolutional neural networks.
During an illustrative realization, the convolutional neural networks are to be cascaded by multiple convolutional layers and activation primitive
Made of network.
The embodiment of the present invention also provides a kind of computer readable storage medium, is stored thereon with computer program, wherein institute
It states and realizes following operation when program is executed by processor:
The value of the number of iterations N of major network in the image enhancement model finished has been trained in adjustment, and the image that is applied increases
Strong model, it is the model generated according to the generation method of aforementioned any image enhancement model that described image, which enhances model,;
Receive image to be reinforced;
The image to be reinforced is handled using application image enhancing model, obtains the image pair to be reinforced
The enhancing image answered.
During an illustrative realization, major network in the image enhancement model finished has been trained in the adjustment
The value of the number of iterations N, be applied image enhancement model, comprising:
Using image to be reinforced to be processed as the first image to be reinforced, instruction used when training described image enhancing model
Practice the image to be reinforced in data as the second image to be reinforced, first image to be reinforced and described second to be reinforced
The picture quality of image;
If the picture quality of first image to be reinforced is better than the picture quality of the described second image to be reinforced, setting the
One the number of iterations is less than secondary iteration number;If the picture quality of second image to be reinforced is better than the described first figure to be reinforced
The picture quality of picture, the first the number of iterations of setting are greater than secondary iteration number;If the picture quality of second image to be reinforced
Identical as the picture quality of the described first image to be reinforced, the first the number of iterations of setting is equal to secondary iteration number;Wherein, described
First the number of iterations is the value that application image enhances the number of iterations N of major network in model, and the secondary iteration number is
The value of the number of iterations N of major network in the image enhancement model that training finishes.
For device and apparatus embodiments, since it corresponds essentially to embodiment of the method, so related place referring to
The part of embodiment of the method illustrates.The apparatus embodiments described above are merely exemplary, wherein described be used as is divided
Module from part description may or may not be physically separated, the component shown as module can be or
It may not be physical module, it can it is in one place, or may be distributed on multiple network modules.It can basis
The actual purpose for needing to select some or all of the modules therein to realize this specification scheme.Ordinary skill people
Member can understand and implement without creative efforts.
It is above-mentioned that this specification specific embodiment is described.Other embodiments are in the scope of the appended claims
It is interior.In some cases, the movement recorded in detail in the claims or step can be come according to the sequence being different from embodiment
It executes and desired result still may be implemented.In addition, process depicted in the drawing not necessarily require show it is specific suitable
Sequence or consecutive order are just able to achieve desired result.In some embodiments, multitasking and parallel processing be also can
With or may be advantageous.
Those skilled in the art will readily occur to this specification after considering specification and practicing the invention applied here
Other embodiments.This specification is intended to cover any variations, uses, or adaptations of this specification, these modifications,
Purposes or adaptive change follow the general principle of this specification and do not apply in the art including this specification
Common knowledge or conventional techniques.The description and examples are only to be considered as illustrative, the true scope of this specification and
Spirit is indicated by the following claims.
It should be understood that this specification is not limited to the precise structure that has been described above and shown in the drawings,
And various modifications and changes may be made without departing from the scope thereof.The range of this specification is only limited by the attached claims
System.
The foregoing is merely the preferred embodiments of this specification, all in this explanation not to limit this specification
Within the spirit and principle of book, any modification, equivalent substitution, improvement and etc. done should be included in the model of this specification protection
Within enclosing.
Claims (10)
1. a kind of generation method of image enhancement model, which is characterized in that the described method includes:
Deep learning network model is set, and the deep learning network model is made of the n times iteration of major network;For described
Each input of deep learning network model, the input are the input of major network in iteration for the first time, main net in i-th iteration
The output of network is the input of major network in i+1 time iteration, the major network parameter having the same in all iteration;And
The value of the number of iterations N and the initial parameter value of the major network are set;
Obtain several groups training data;
The deep learning network model is trained using the training data, obtains the deep learning network that training finishes
Model, using the deep learning network model that the training finishes as image enhancement model;
Wherein, N, i are natural number, and N-1 >=i >=1.
2. the method according to claim 1, wherein every group of training data packet in the several groups training data
Include image to be reinforced and the corresponding first enhancing image of the image to be reinforced;
It is described that the deep learning network model is trained using the training data, obtain the deep learning that training finishes
Network model is trained, comprising:
In the training process, the parameter value of major network is described in the corresponding deep learning network model of the 1st group of training data
Initial parameter value, the parameter value of major network is through -1 group instruction of jth in the corresponding deep learning network model of jth group training data
Practice the parameter value that adjusts after data training, j is natural number, and j >=2;For every group of training data, perform the following operations:
Image to be reinforced in this group of training data is inputted into the corresponding deep learning network model of this group of training data, is somebody's turn to do
The corresponding second enhancing image of group training data;
Obtain the difference value of the first enhancing image and the second enhancing image;
If the difference value is greater than preset threshold, major network in the deep learning network model is adjusted according to the difference value
Parameter value;If the difference value is less than or equal to the preset threshold, deconditioning, with the corresponding depth of this group of training data
The deep learning network model that learning network model is finished as training.
3. according to the method described in claim 2, it is characterized by further comprising:
Building loss function in advance, the loss function include with first corresponding first variable of enhancing image and with it is described
Corresponding second variable of second enhancing image;
The difference value for obtaining the first enhancing image and the second enhancing image, comprising:
Using the first enhancing image as the value of the first variable in the loss function, using the second enhancing image as institute
Bivariate value in loss function is stated, the value of loss function is calculated, using the value of loss function as the first enhancing image
With the difference value of the second enhancing image.
4. according to the method described in claim 2, it is characterized in that, described adjust the deep learning net according to the difference value
The parameter value of major network in network model, comprising:
Each layer that the difference value is returned to deep learning network model by back propagation, to deep learning network mould
The parameter value of major network is modified in type.
5. the method according to claim 1, wherein the major network is convolutional neural networks.
6. according to the method described in claim 5, it is characterized in that, the convolutional neural networks are by multiple convolutional layers and activation
Network made of function cascaded.
7. a kind of application method of image enhancement model, which is characterized in that the described method includes:
The value of the number of iterations N of major network in the image enhancement model finished has been trained in adjustment, and be applied Image Enhancement Based
Type, it is the model generated according to the generation method of any one of claim 1~6 image enhancement model that described image, which enhances model,;
Receive image to be reinforced;
The image to be reinforced is handled using application image enhancing model, it is corresponding to obtain the image to be reinforced
Enhance image.
8. the method according to the description of claim 7 is characterized in that master in the image enhancement model finished has been trained in the adjustment
The value of the number of iterations N of volume grid, be applied image enhancement model, comprising:
Using image to be reinforced to be processed as the first image to be reinforced, trained number used when training described image enhancing model
Image to be reinforced in is as the second image to be reinforced, first image to be reinforced and the described second image to be reinforced
Picture quality;
If the picture quality of first image to be reinforced is better than the picture quality of the described second image to be reinforced, setting first changes
Generation number is less than secondary iteration number;If the picture quality of second image to be reinforced is better than the described first image to be reinforced
Picture quality, the first the number of iterations of setting are greater than secondary iteration number;If picture quality and the institute of second image to be reinforced
The picture quality for stating the first image to be reinforced is identical, and the first the number of iterations of setting is equal to secondary iteration number;Wherein, described first
The number of iterations is the value that application image enhances the number of iterations N of major network in model, and the secondary iteration number is to have trained
The value of the number of iterations N of major network in the image enhancement model finished.
9. a kind of generating means of image enhancement model, which is characterized in that described device includes:
Setup module, for being arranged deep learning network model, the deep learning network model by major network n times iteration
Composition;Each input for the deep learning network model, the input are the input of major network in iteration for the first time, i-th
The output of major network is the input of major network in i+1 time iteration in secondary iteration, and the major network in all iteration has
Identical parameter;And the value of the number of iterations N and the initial parameter value of the major network are set;
Training data obtains module, for obtaining several groups training data;
Training module is obtained training and finished for being trained using the training data to the deep learning network model
Deep learning network model, using the deep learning network model that the training finishes as image enhancement model;
Wherein, N, i are natural number, and N-1 >=i >=1.
10. a kind of application apparatus of image enhancement model, which is characterized in that described device includes:
The number of iterations adjusts module, for adjusting the number of iterations N for having trained major network in the image enhancement model finished
Value, be applied image enhancement model, and it is according to any one of claim 1~6 image enhancement model that described image, which enhances model,
Generation method generate model;
Image receiver module, for receiving image to be reinforced;
Image processing module obtains institute for handling using application image enhancing model the image to be reinforced
State the corresponding enhancing image of image to be reinforced.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910769010.4A CN110533610A (en) | 2019-08-20 | 2019-08-20 | The generation method and device of image enhancement model, application method and device |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910769010.4A CN110533610A (en) | 2019-08-20 | 2019-08-20 | The generation method and device of image enhancement model, application method and device |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110533610A true CN110533610A (en) | 2019-12-03 |
Family
ID=68663699
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910769010.4A Pending CN110533610A (en) | 2019-08-20 | 2019-08-20 | The generation method and device of image enhancement model, application method and device |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110533610A (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111179235A (en) * | 2019-12-23 | 2020-05-19 | 沈阳先进医疗设备技术孵化中心有限公司 | Image detection model generation method and device, and application method and device |
CN111311531A (en) * | 2020-01-22 | 2020-06-19 | 沈阳先进医疗设备技术孵化中心有限公司 | Image enhancement method and device, console equipment and medical imaging system |
CN112213673A (en) * | 2020-09-07 | 2021-01-12 | 上海东软医疗科技有限公司 | Dynamic magnetic resonance imaging method, device, reconstruction computer and magnetic resonance system |
CN113240113A (en) * | 2021-06-04 | 2021-08-10 | 北京富通东方科技有限公司 | Method for enhancing network prediction robustness |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103531042A (en) * | 2013-10-25 | 2014-01-22 | 吉林大学 | Rear-end collision pre-warning method based on driver types |
CN106873595A (en) * | 2017-03-13 | 2017-06-20 | 同济大学 | A kind of is recognition methods with garage based on Timed Automata |
CN108460726A (en) * | 2018-03-26 | 2018-08-28 | 厦门大学 | A kind of magnetic resonance image super-resolution reconstruction method based on enhancing recurrence residual error network |
CN108647577A (en) * | 2018-04-10 | 2018-10-12 | 华中科技大学 | A kind of pedestrian's weight identification model that adaptive difficult example is excavated, method and system |
CN108830816A (en) * | 2018-06-27 | 2018-11-16 | 厦门美图之家科技有限公司 | Image enchancing method and device |
CN109978778A (en) * | 2019-03-06 | 2019-07-05 | 浙江工业大学 | Convolutional neural networks medicine CT image denoising method based on residual error study |
-
2019
- 2019-08-20 CN CN201910769010.4A patent/CN110533610A/en active Pending
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103531042A (en) * | 2013-10-25 | 2014-01-22 | 吉林大学 | Rear-end collision pre-warning method based on driver types |
CN106873595A (en) * | 2017-03-13 | 2017-06-20 | 同济大学 | A kind of is recognition methods with garage based on Timed Automata |
CN108460726A (en) * | 2018-03-26 | 2018-08-28 | 厦门大学 | A kind of magnetic resonance image super-resolution reconstruction method based on enhancing recurrence residual error network |
CN108647577A (en) * | 2018-04-10 | 2018-10-12 | 华中科技大学 | A kind of pedestrian's weight identification model that adaptive difficult example is excavated, method and system |
CN108830816A (en) * | 2018-06-27 | 2018-11-16 | 厦门美图之家科技有限公司 | Image enchancing method and device |
CN109978778A (en) * | 2019-03-06 | 2019-07-05 | 浙江工业大学 | Convolutional neural networks medicine CT image denoising method based on residual error study |
Non-Patent Citations (1)
Title |
---|
SHENGYU ZHAO等: "Recursive Cascaded Networks for Unsupervised Medical Image Registration", 《ARXIV》 * |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111179235A (en) * | 2019-12-23 | 2020-05-19 | 沈阳先进医疗设备技术孵化中心有限公司 | Image detection model generation method and device, and application method and device |
CN111179235B (en) * | 2019-12-23 | 2024-03-08 | 东软医疗***股份有限公司 | Image detection model generation method and device, and application method and device |
CN111311531A (en) * | 2020-01-22 | 2020-06-19 | 沈阳先进医疗设备技术孵化中心有限公司 | Image enhancement method and device, console equipment and medical imaging system |
CN111311531B (en) * | 2020-01-22 | 2024-03-08 | 东软医疗***股份有限公司 | Image enhancement method, device, console device and medical imaging system |
CN112213673A (en) * | 2020-09-07 | 2021-01-12 | 上海东软医疗科技有限公司 | Dynamic magnetic resonance imaging method, device, reconstruction computer and magnetic resonance system |
CN112213673B (en) * | 2020-09-07 | 2022-11-22 | 上海东软医疗科技有限公司 | Dynamic magnetic resonance imaging method, device, reconstruction computer and magnetic resonance system |
CN113240113A (en) * | 2021-06-04 | 2021-08-10 | 北京富通东方科技有限公司 | Method for enhancing network prediction robustness |
CN113240113B (en) * | 2021-06-04 | 2024-05-28 | 北京富通东方科技有限公司 | Method for enhancing network prediction robustness |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110533610A (en) | The generation method and device of image enhancement model, application method and device | |
CN109658344A (en) | Image de-noising method, device, equipment and storage medium based on deep learning | |
Venkatakrishnan et al. | Plug-and-play priors for model based reconstruction | |
DE69021089T2 (en) | HEURISTIC PROCESSOR. | |
CN113902921B (en) | Image processing method, device, equipment and storage medium | |
CN109949215B (en) | Low-dose CT image simulation method | |
CN110598843B (en) | Training method for generating countermeasure network organization structure based on discriminant sharing | |
CN112215755B (en) | Image super-resolution reconstruction method based on back projection attention network | |
CN111383741B (en) | Method, device and equipment for establishing medical imaging model and storage medium | |
JP2017151438A (en) | Method and apparatus for generating grapheme-to-phoneme conversion model based on artificial intelligence | |
DE102021107050A1 (en) | GRADIENT COMPRESSION FOR DISTRIBUTED TRAINING | |
DE102021107509A1 (en) | MULTI-MODEL TRAINING PIPELINE IN DISTRIBUTED SYSTEMS | |
CN107595312A (en) | Model generating method, image processing method and medical imaging devices | |
CN110415198A (en) | A kind of Method of Medical Image Fusion based on laplacian pyramid Yu parameter adaptive Pulse Coupled Neural Network | |
CN109345604B (en) | Picture processing method, computer device and storage medium | |
CN105719276A (en) | Liver parenchymal segmentation method based on CT images | |
Li et al. | A fully trainable network with RNN-based pooling | |
CN114494003B (en) | Ancient character generation method combining shape transformation and texture transformation | |
DE102021124537A1 (en) | ENERGY-BASED VARIATIONAL AUTOENCODER | |
CN112288046B (en) | Mixed granularity-based joint sparse method for neural network | |
CN109993701A (en) | A method of the depth map super-resolution rebuilding based on pyramid structure | |
CN113674156A (en) | Method and system for reconstructing image super-resolution | |
US20210319326A1 (en) | Method and apparatus for learning stochastic inference models between multiple random variables with unpaired data | |
CN113298895A (en) | Convergence guarantee-oriented unsupervised bidirectional generation automatic coding method and system | |
CN112602121B (en) | Method, device and system for reconstructing point cloud model |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20191203 |