CN109872288A - For the network training method of image denoising, device, terminal and storage medium - Google Patents
For the network training method of image denoising, device, terminal and storage medium Download PDFInfo
- Publication number
- CN109872288A CN109872288A CN201910100276.XA CN201910100276A CN109872288A CN 109872288 A CN109872288 A CN 109872288A CN 201910100276 A CN201910100276 A CN 201910100276A CN 109872288 A CN109872288 A CN 109872288A
- Authority
- CN
- China
- Prior art keywords
- image
- denoising
- convolutional neural
- neural networks
- loss
- 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
-
- 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
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The present invention is applicable in technical field of image processing, provide a kind of network training method for image denoising, device, terminal and storage medium, this method comprises: the noise image of preset quantity is inputted preset convolutional neural networks, initialize the parameter of convolutional neural networks, the denoising characteristics of image of noise image is obtained by the denoising layer of convolutional neural networks, it is combined by the combination layer of preset convolutional neural networks by characteristics of image is denoised, to obtain the denoising image after noise image denoising, image will be denoised and the corresponding noise-free picture of noise image inputs preset loss of detail model, obtain loss of detail when convolutional neural networks denoising, when loss of detail model is not converged, by loss of detail anti-pass to convolutional neural networks, and convolutional neural networks parameter is updated according to loss of detail, to continue to convolutional neural networks It is trained, so that by constantly denoising, adjusting parameter reduces loss of detail when network, and then improve the denoising effect of the network.
Description
Technical field
The invention belongs to technical field of image processing more particularly to a kind of network training methods for image denoising, dress
It sets, terminal and storage medium.
Background technique
With the development of various electric terminal photographs technologies, it is more commonly used in today's society that image, which is developing progressively,
Information carrier, but usually will receive in the acquisition, transmission or storage process of image various noises interference and influence and make figure
As degrading.Image denoising is a critically important research direction in field of image processing, and the purpose is to by making an uproar in noise image
Sound removal, to obtain clean images.
Currently, most of image denoisings all rely on the convolutional neural networks after having carried out deep learning, and convolutional Neural
For network in the process of deep learning, more situation is with mean square error (MSE, mean square error) as optimization net
The loss function of network, although the loss function can help network to find one in the standard that objectively evaluates i.e. Y-PSNR
(peak signal-to-noise ratio, PSNR) shows good solution, but simultaneously, which is calculating figure
When the loss of detail of picture, MSE can drive network to look for a homogenizing scheme, and the feature that can not be caught well between grabgraf picture is poor
Away from the feature gap grabbed is excessively balanced, so that the image after restoring is more fuzzy.
Summary of the invention
The purpose of the present invention is to provide a kind of for the network training method of image denoising, device, terminal and storage
Medium, it is intended to solve that a kind of effective image denoising network training method can not be provided due to the prior art, lead to conventional images
Denoise the undesirable problem of the image denoising effect of network.
On the one hand, the present invention provides a kind of network training method for image denoising, the method includes following steps
It is rapid:
The noise image of preset quantity is inputted into preset convolutional neural networks, initializes the ginseng of the convolutional neural networks
Number obtains the denoising characteristics of image of noise image by the denoising layer of the convolutional neural networks；
The denoising characteristics of image is combined by the combination layer of the preset convolutional neural networks, to obtain
Denoising image after stating noise image denoising；
The denoising image and the corresponding noise-free picture of the noise image are inputted into preset loss of detail model, led to
It crosses the loss of detail model and obtains loss of detail when the convolutional neural networks denoise the noise image；
When the loss of detail model is not converged, by gradient anti-pass mode by the loss of detail anti-pass to the volume
Neural network is accumulated, and updates the parameter of the convolutional neural networks according to the loss of detail, to continue to the convolutional Neural
Network is trained.
On the other hand, the present invention provides a kind of network training device for image denoising, described device includes:
Feature denoises unit, for the noise image of preset quantity to be inputted preset convolutional neural networks, initializes institute
The parameter for stating convolutional neural networks obtains the denoising characteristics of image of noise image by the denoising layer of the convolutional neural networks；
Feature assembled unit, for the combination layer by the preset convolutional neural networks by the denoising characteristics of image
It is combined, to obtain the denoising image after the noise image denoising；
Acquiring unit is lost, it is default for inputting the denoising image and the corresponding noise-free picture of the noise image
Loss of detail model, obtained when the convolutional neural networks denoise the noise image by the loss of detail model
Loss of detail；And
Anti-pass updating unit is used for when the loss of detail model is not converged, will be described thin by gradient anti-pass mode
Section loses anti-pass to the convolutional neural networks, and the parameter of the convolutional neural networks is updated according to the loss of detail, with
Continue to be trained the convolutional neural networks.
On the other hand, the present invention also provides a kind of computing terminal, including memory, processor and it is stored in described deposit
In reservoir and the computer program that can run on the processor, the processor are realized such as when executing the computer program
The step of above-mentioned network training method for image denoising.
On the other hand, the present invention also provides a kind of computer readable storage medium, the computer readable storage mediums
It is stored with computer program, such as the above-mentioned network training for image denoising is realized when the computer program is executed by processor
The step of method.
The noise image of preset quantity is first inputted preset convolutional neural networks by the present invention, initializes convolutional neural networks
Parameter, by convolutional neural networks denoising layer obtain noise image denoising characteristics of image, pass through preset convolutional Neural
The combination layer of network is combined characteristics of image is denoised, and to obtain the denoising image after noise image denoising, then will denoise figure
Picture and the corresponding noise-free picture of noise image input preset loss of detail model, obtain thin when convolutional neural networks denoise
Section loss, when loss of detail model is not converged, by loss of detail anti-pass to convolutional neural networks, and updates according to loss of detail
Convolutional neural networks parameter, to continue to be trained convolutional neural networks, so that by constantly denoising, adjusting parameter reduces
Loss of detail when network, and then improve the denoising effect of the network.
Detailed description of the invention
Fig. 1 is the implementation flow chart for the network training method for image denoising that the embodiment of the present invention one provides；
Fig. 2 is the structural schematic diagram for the image denoising network that the embodiment of the present invention one provides；
Fig. 3 is the implementation flow chart for the acquisition loss of detail that the embodiment of the present invention one provides；
Fig. 4 is the schematic diagram of a layer structure for the VGG19 network that the embodiment of the present invention one provides；And
Fig. 5 is the structural schematic diagram of the network training device provided by Embodiment 2 of the present invention for image denoising；
Fig. 6 is the structural schematic diagram for the network training device for image denoising that the embodiment of the present invention three provides；And
Fig. 7 is a kind of structural schematic diagram for terminal that the embodiment of the present invention four provides.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to the accompanying drawings and embodiments, right
The present invention is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, and
It is not used in the restriction present invention.
Specific implementation of the invention is described in detail below in conjunction with specific embodiment:
Embodiment one:
Fig. 1 shows the implementation process of the network training method for image denoising of the offer of the embodiment of the present invention one, is
Convenient for explanation, only parts related to embodiments of the present invention are shown, details are as follows:
In step s101, the noise image of preset quantity is inputted into preset convolutional neural networks, initialization convolution mind
Parameter through network obtains the denoising characteristics of image of noise image by the denoising layer of convolutional neural networks.
The embodiment of the present invention is suitable for computing terminal, which can load, run the network for being used for image denoising, is
It is convenient for subsequent descriptions, which is known as image denoising network, it can be right by the image denoising network
Image carries out noise remove.The preferable image denoising network of effect is denoised in order to obtain, is first carried out using convolutional neural networks deep
Therefore the noise image of preset quantity in embodiments of the present invention, is first inputted preset convolutional neural networks by degree denoising study
Among, then Initialize installation is carried out to the parameter of the convolutional neural networks, then obtained by the denoising layer of convolutional neural networks
The denoising characteristics of image of noise image.
Preferably, the noise image of preset quantity is noise source approximation or identical image, for example, can be according to environment
Difference instructs the image of image or cloudy day shooting or the image of fine day shooting that the haze sky of preset quantity is shot respectively
Practice, obtained image denoising network can the image respectively to haze sky, cloudy day and fine day environment reach good denoising effect；?
When being initialized, Initialize installation can be carried out to the parameter of the convolutional neural networks by way of being randomly provided parameter；?
When denoising to the noise image of input, noise pattern is extracted by the first default convolutional layer and active coating of convolutional neural networks
The noise image feature of picture denoises noise image feature by the residual block of preset convolutional neural networks, to obtain
Characteristics of image is denoised, to denoise according to current image denoising network parameter to the noise image of input, the residual block
The second default convolutional layer and active coating including convolutional neural networks, wherein for ease of description, quantity will be pre-set
It convolutional layer and pre-sets the active coating of quantity and according to pre-set sequence carries out arrangement and be formed by network structure being known as
First default convolutional layer and active coating or the second default convolutional layer and active coating.
Fig. 2 shows the structure charts for the image denoising network that the embodiment of the present invention one provides, wherein (a) is image denoising
The structural schematic diagram of network；It (b) is the structural schematic diagram of the residual block of image denoising network.
In embodiments of the present invention, the convolutional neural networks denoising layer by preset quantity convolutional layer and active coating structure
At after receiving noise image, by 1 convolutional layer (k9n64s1) and 1 active coating (PReLU) to the spy of noise image
Sign extracts, and the noise image feature is then input to residual block, denoises in residual block to the noise image feature
When, 1 residual block denoises noise image feature by 2 convolutional layers (k3n64s1) and 1 active coating (PReLU), is obtained
Feature and the noise image feature of input be overlapped, while the input as next residual block, to the last one it is residual
Noise image feature is superimposed to denoising characteristics of image after obtaining denoising characteristics of image by poor block output denoising characteristics of image
In, while being input to the combination layer of the image denoising network.It disappears in the gradient of prevention convolutional neural networks and gradient explosion is asked
When topic, BN (Batch Normalization, batch normalize) layer of convolutional neural networks is deleted in embodiments of the present invention
It removes, so that characteristics of image be effectively prevent to become unintelligible, convolutional neural networks is retained in image denoising more thin
Section.
Wherein, PReLU indicates the title of activation primitive, and k indicates that convolutional layer corresponds to convolution kernel size, and n indicates characteristic pattern number
Amount, s indicates convolution paces size, for example, k9n64s1, expression convolution kernel size is 9*9, and output characteristic pattern quantity 64 is opened, convolution
The paces of core sliding are 1.
Further, before it will denoise characteristics of image and be combined, it is special that noise image feature is superimposed to denoising image
In sign, while it being input to the combination layer of the image denoising network, to avoid the problem that gradient disappears and gradient is exploded.
In step s 102, it is combined by the combination layer of preset convolutional neural networks by characteristics of image is denoised, with
Denoising image after obtaining noise image denoising.
In embodiments of the present invention, denoising image is that the convolutional neural networks that the noise image inputted is predetermined are denoised
Image generated after operation passes through 2 convolutional layers (k3n256s1), 1 convolutional layer after getting denoising characteristics of image
(k9n3s1) and 2 active coatings (PReLU) carry out feature combination to the denoising characteristics of image of acquisition, thus being somebody's turn to do after being denoised
The denoising image of noise image.
In step s 103, denoising image and the corresponding noise-free picture of noise image are inputted into preset loss of detail mould
Type obtains loss of detail when convolutional neural networks denoise noise image.
In embodiments of the present invention, loss of detail model is pre-set for calculating the corresponding noiseless of noise image
The model of the characteristic loss of image, after obtaining denoising image, by denoising image and the corresponding noise-free picture of noise image
Preset loss of detail model is inputted, loss of detail when convolutional neural networks denoise noise image, the loss of detail are obtained
Illustrate the denoising effect of the current parameter of convolutional neural networks.
Fig. 3 shows the implementation process of the acquisition loss of detail of the offer of the embodiment of the present invention one, is obtained by following steps
Loss of detail preferably supplemental training can go out to denoise the preferable image denoising network of effect:
In step S301, the characteristic pattern and noise-free picture for denoising image are obtained by preset VGG neural network
The first Euclidean distance between characteristic pattern.
In embodiments of the present invention, Fig. 4 shows the structure chart of the VGG19 network of the offer of the embodiment of the present invention one, VGG mind
It is preparatory trained 19 layers of VGG network (VGG19 network) through network, after having input denoising image and noise-free picture, point
Image will not denoised and noise-free picture imports in trained VGG19 network in advance, then respectively from the 5th pond of VGG19 network
Change the characteristic pattern that the 4th convolutional layer (conv5-4) in layer (pool5) extracts denoising image and noise-free picture, according to VGG19 net
The loss of detail that the characteristic pattern that the 4th convolutional layer extracts in the 5th pond layer of network obtains can preferably assist convolution nerve net
Network restores the details of noise image, is then obtained in the 5th pond layer (pool5) of VGG19 network by Euclidean distance formula
The 4th convolutional layer (conv5-4) extract denoising image and the characteristic pattern of noise-free picture between Euclidean distanceFor ease of description, that the Euclidean distance obtained at this time is known as first is European
Distance, at this point, first Euclidean distance indicates perception loss when noise image is denoised in convolutional neural networks, by this
A kind of loss of detail for the loss of detail that perception loss generates during denoising as image, so that image denoising network
Retaining the details in more noise-free pictures when denoising to image, wherein x indicates noise-free picture,Indicate noise pattern
Picture, functionIndicate the characteristic pattern that the 4th convolutional layer extracts in the 5th pond layer of VGG19 network, W5,4Indicate that characteristic pattern exists
Dimension in width direction, H5,4Indicate the dimension of characteristic pattern in the height direction, (i, j) indicates the pixel coordinate of characteristic pattern.
In step s 302, the frequency image information of denoising image and the high frequency of noise-free picture are obtained by High frequency filter
The second Euclidean distance between image information.
In embodiments of the present invention, the high frequency letter that high-pass filter obtains denoising image and noise-free picture respectively is first passed through
Breath, the high-pass filter can be the single order high-pass filter using Sobel operator (Sobel Operator), which includes
The detailed information such as edge and texture, then by Euclidean distance formula obtain denoising image and noise-free picture high-frequency information it
Between Euclidean distanceFor ease of description, the Euclidean distance obtained at this time is become the
Two Euclidean distances, which indicates high frequency loss when noise image is denoised in convolutional neural networks, by the height
Frequently another loss of detail of the loss of detail generated during denoising as image is lost, so that the image obtained be avoided to go
Network of making an uproar generates high-frequency artifacts when denoising to image, wherein function phi indicates the high frequency obtained by high-pass filter
Information image, W and H respectively indicate high-frequency information image in the direction of the width with the dimension in short transverse, and (i, j) indicates feature
The pixel coordinate of figure.
In step S303, according to formulaObtain loss of detail.
In embodiments of the present invention, the perception loss when getting noise image and being denoised in convolutional neural networks
After high frequency loss, according to loss of detail formulaWhen obtaining convolutional neural networks to noise image denoising
Loss of detail, for ease of description by formulaReferred to as loss of detail formula, wherein ldetailExpression is made an uproar
The loss of detail that acoustic image is generated when being denoised, lperIndicate the first Euclidean distance, lhfIndicate the second Euclidean distance,For
The weight of second Euclidean distance.
In step S104, when loss of detail model is not converged, loss of detail anti-pass is arrived by gradient anti-pass mode
Convolutional neural networks, and according to the parameter of loss of detail update convolutional neural networks, to continue to instruct convolutional neural networks
Practice.
In embodiments of the present invention, when loss of detail model is not converged, show the parameter of current convolutional neural networks simultaneously
It cannot reach expected and denoise effect, at this point, by gradient anti-pass mode by loss of detail anti-pass to convolutional neural networks, and root
The parameter that convolutional neural networks are updated according to loss of detail, to continue to be trained convolutional neural networks, until the loss of detail
When model is restrained, the parameter of updated convolutional neural networks is exported, to obtain the image denoising network of training completion, by not
Disconnected training image denoising network denoises orientation type image, to obtain the denoising effect for making the orientation type image
The parameter of preferable convolutional neural networks.
In embodiments of the present invention, the noise image of preset quantity is inputted into preset convolutional neural networks, initialization volume
The parameter of product neural network obtains the denoising characteristics of image of noise image by the denoising layer of convolutional neural networks, by default
Convolutional neural networks combination layer by denoise characteristics of image be combined, with obtain noise image denoising after denoising image,
Image will be denoised and the corresponding noise-free picture of noise image inputs preset loss of detail model, convolutional neural networks are obtained and go
Loss of detail when making an uproar, when loss of detail model is not converged, by loss of detail anti-pass to convolutional neural networks, and according to details
Loss updates convolutional neural networks parameter, to continue to be trained convolutional neural networks, thus by constantly denoising, adjusting ginseng
Number improves the denoising effect of the network to reduce loss of detail when network.
Embodiment two:
Fig. 5 shows the structure of the network training device provided by Embodiment 2 of the present invention for image denoising, in order to just
In explanation, only parts related to embodiments of the present invention are shown, including:
Feature denoises unit 51, for the noise image of preset quantity to be inputted preset convolutional neural networks, initialization
The parameter of convolutional neural networks obtains the denoising characteristics of image of noise image by the denoising layer of convolutional neural networks；
Feature assembled unit 52 carries out group for characteristics of image is denoised for the combination layer by preset convolutional neural networks
It closes, to obtain the denoising image after noise image denoising；
Acquiring unit 53 is lost, inputs preset details for image and the corresponding noise-free picture of noise image will to be denoised
Loss model obtains loss of detail when convolutional neural networks denoise noise image；And
Anti-pass updating unit 54, for when loss of detail model is not converged, by gradient anti-pass mode by loss of detail
Anti-pass to convolutional neural networks, and according to loss of detail update convolutional neural networks parameter, to continue to convolutional neural networks
It is trained.
In embodiments of the present invention, the noise image of preset quantity is inputted into preset convolutional neural networks, initialization volume
The parameter of product neural network obtains the denoising characteristics of image of noise image by the denoising layer of convolutional neural networks, by default
Convolutional neural networks combination layer by denoise characteristics of image be combined, with obtain noise image denoising after denoising image,
Image will be denoised and the corresponding noise-free picture of noise image inputs preset loss of detail model, convolutional neural networks are obtained and go
Loss of detail when making an uproar, when loss of detail model is not converged, by loss of detail anti-pass to convolutional neural networks, and according to details
Loss updates convolutional neural networks parameter, to continue to be trained convolutional neural networks, thus by constantly denoising, adjusting ginseng
Number improves the denoising effect of the network to reduce loss of detail when network.
It in embodiments of the present invention, can be by corresponding hardware or soft for each unit of the network training device of image denoising
Part unit realizes that each unit can be independent soft and hardware unit, also can integrate as a soft and hardware unit, does not have to herein
To limit the present invention.The specific embodiment of each unit can refer to the description of embodiment one, and details are not described herein.
Embodiment three:
Fig. 6 shows the structure of the network training device for image denoising of the offer of the embodiment of the present invention three, in order to just
In explanation, only parts related to embodiments of the present invention are shown, including:
Feature denoises unit 61, for the noise image of preset quantity to be inputted preset convolutional neural networks, initialization
The parameter of convolutional neural networks obtains the denoising characteristics of image of noise image by the denoising layer of convolutional neural networks；
Feature assembled unit 62 carries out group for characteristics of image is denoised for the combination layer by preset convolutional neural networks
It closes, to obtain the denoising image after noise image denoising；
Acquiring unit 63 is lost, inputs preset details for image and the corresponding noise-free picture of noise image will to be denoised
Loss model obtains loss of detail when convolutional neural networks denoise noise image；And
Anti-pass updating unit 64, for when loss of detail model is not converged, by gradient anti-pass mode by loss of detail
Anti-pass to convolutional neural networks, and according to loss of detail update convolutional neural networks parameter, to continue to convolutional neural networks
It is trained.
Wherein, feature denoising unit 61 includes:
Feature extraction unit 611, for the first default convolutional layer and active coating extraction noise by convolutional neural networks
The noise image feature of image；And
Feature denoise subelement 612, for the residual block by preset convolutional neural networks to noise image feature into
Row denoising, to obtain denoising characteristics of image, residual block includes the second default convolutional layer and active coating of convolutional neural networks.
Losing acquiring unit 63 includes:
First distance acquiring unit 631, for obtaining the characteristic pattern and nothing of denoising image by preset VGG neural network
The first Euclidean distance between the characteristic pattern of noise image；
Second distance acquiring unit 632 is made an uproar for obtaining the frequency image information for denoising image and nothing by High frequency filter
The second Euclidean distance between the frequency image information of acoustic image；And
Loss obtains subelement 633, for according to formulaObtain loss of detail, ldetailExpression is made an uproar
The loss of detail that acoustic image is generated when being denoised, lperIndicate the first Euclidean distance, lhfIndicate the second Euclidean distance,For
The weight of second Euclidean distance.
In embodiments of the present invention, the noise image of preset quantity is inputted into preset convolutional neural networks, initialization volume
The parameter of product neural network obtains the denoising characteristics of image of noise image by the denoising layer of convolutional neural networks, by default
Convolutional neural networks combination layer by denoise characteristics of image be combined, with obtain noise image denoising after denoising image,
Image will be denoised and the corresponding noise-free picture of noise image inputs preset loss of detail model, convolutional neural networks are obtained and go
Loss of detail when making an uproar, when loss of detail model is not converged, by loss of detail anti-pass to convolutional neural networks, and according to details
Loss updates convolutional neural networks parameter, to continue to be trained convolutional neural networks, until the loss of detail model is restrained
When, the parameter of updated convolutional neural networks is exported, to obtain the image denoising network of training completion, thus by constantly going
Make an uproar, adjusting parameter reduces loss of detail when network, and then improve the denoising effect of the network.
It in embodiments of the present invention, can be by corresponding hardware or soft for each unit of the network training device of image denoising
Part unit realizes that each unit can be independent soft and hardware unit, also can integrate as a soft and hardware unit, does not have to herein
To limit the present invention.The specific embodiment of each unit can refer to the description of embodiment one, and details are not described herein.
Example IV:
Fig. 7 shows the structure of the computing terminal of the offer of the embodiment of the present invention four, for ease of description, illustrates only and this
The relevant part of inventive embodiments, including:
The computing terminal 7 of the embodiment of the present invention includes processor 71, memory 72 and is stored in memory 72 and can
The computer program 73 run on processor 71.The processor 71 is realized when executing computer program 73 and above-mentioned is gone for image
The step in network training method embodiment made an uproar, for example, step S101 to S104 shown in FIG. 1 and step shown in Fig. 3
S301 to S303.Alternatively, processor 71 realizes above-mentioned each network training for image denoising when executing computer program 73
The function of each unit in Installation practice, for example, the function of unit 61 to 64 shown in unit 51 to 54 and Fig. 6 shown in Fig. 5.
In embodiments of the present invention, when which executes computer program, the noise image of preset quantity is inputted pre-
If convolutional neural networks, initialize the parameter of convolutional neural networks, noise pattern obtained by the denoising layers of convolutional neural networks
The denoising characteristics of image of picture is combined, to obtain by characteristics of image is denoised by the combination layer of preset convolutional neural networks
Denoising image and the corresponding noise-free picture of noise image are inputted preset details and damaged by the denoising image after noise image denoising
Model is lost, loss of detail when convolutional neural networks denoising is obtained, when loss of detail model is not converged, by loss of detail anti-pass
Convolutional neural networks parameter is updated to convolutional neural networks, and according to loss of detail, to continue to instruct convolutional neural networks
Practice, so that by constantly denoising, adjusting parameter reduces loss of detail when network, and then improve the denoising effect of the network.
The processor is realized in the above-mentioned network training method embodiment for image denoising when executing computer program
Step can refer to the description of embodiment one, and details are not described herein.
Embodiment five:
In embodiments of the present invention, a kind of computer readable storage medium is provided, which deposits
Computer program is contained, which realizes above-mentioned each network training side for image denoising when being executed by processor
Step in method embodiment, for example, step S101 to S104 shown in FIG. 1 and step S301 to S303 shown in Fig. 3.Or
Person, the computer program are realized each in above-mentioned each network training Installation practice for image denoising when being executed by processor
The function of unit, for example, the function of unit 61 to 64 shown in unit 51 to 54 and Fig. 6 shown in Fig. 5.
In embodiments of the present invention, after computer program is executed by processor, the noise image of preset quantity is inputted
Preset convolutional neural networks initialize the parameter of convolutional neural networks, obtain noise by the denoising layer of convolutional neural networks
The denoising characteristics of image of image is combined by the combination layer of preset convolutional neural networks by characteristics of image is denoised, with
Denoising image and the corresponding noise-free picture of noise image are inputted preset details by the denoising image to after noise image denoising
Loss model obtains loss of detail when convolutional neural networks denoising, when loss of detail model is not converged, loss of detail is anti-
Convolutional neural networks are passed to, and convolutional neural networks parameter is updated according to loss of detail, to continue to carry out convolutional neural networks
Training, so that by constantly denoising, adjusting parameter reduces loss of detail when network, and then improve the denoising effect of the network
Fruit.
The computer program is realized when being executed by processor in the above-mentioned network training method embodiment for image denoising
The step of can refer to the description of embodiment one, details are not described herein.
The computer readable storage medium of the embodiment of the present invention may include can carry computer program code any
Entity or device, storage medium, for example, the memories such as ROM/RAM, disk, CD, flash memory.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all in essence of the invention
Made any modifications, equivalent replacements, and improvements etc., should all be included in the protection scope of the present invention within mind and principle.
Claims (10)
1. a kind of network training method for image denoising, which is characterized in that the method includes the following steps:
The noise image of preset quantity is inputted into preset convolutional neural networks, initializes the parameter of the convolutional neural networks,
The denoising characteristics of image of noise image is obtained by the denoising layer of the convolutional neural networks；
The denoising characteristics of image is combined by the combination layer of the preset convolutional neural networks, to obtain described make an uproar
Denoising image after acoustic image denoising；
The denoising image and the corresponding noise-free picture of the noise image are inputted into preset loss of detail model, pass through institute
It states loss of detail model and obtains loss of detail when the convolutional neural networks denoise the noise image；
It is by gradient anti-pass mode that the loss of detail anti-pass is refreshing to the convolution when the loss of detail model is not converged
Through network, and the parameter of the convolutional neural networks is updated according to the loss of detail, to continue to the convolutional neural networks
It is trained.
2. the method as described in claim 1, which is characterized in that obtain the convolutional Neural net by the loss of detail model
The step of loss of detail when network denoises the noise image, comprising:
It is obtained by preset VGG neural network between the characteristic pattern of the denoising image and the characteristic pattern of the noise-free picture
The first Euclidean distance；
The frequency image information of the denoising image and the frequency image information of the noise-free picture are obtained by High frequency filter
Between the second Euclidean distance；
According to formulaObtain the loss of detail, wherein ldetailIndicate the loss of detail, lperIt indicates
First Euclidean distance, lhfIndicate second Euclidean distance,Indicate the weight of second Euclidean distance.
3. the method as described in right wants 1, which is characterized in that obtain noise image by the denoising layer of the convolutional neural networks
Denoising characteristics of image the step of, comprising:
The noise image for extracting the noise image by the first default convolutional layer and active coating of the convolutional neural networks is special
Sign；
The noise image feature is denoised by the residual block of the preset convolutional neural networks, to obtain described go
It makes an uproar characteristics of image, the residual block includes the second default convolutional layer and active coating of the convolutional neural networks.
4. method as claimed in claim 3, which is characterized in that obtain noise pattern by the denoising layer of the convolutional neural networks
The step of denoising characteristics of image of picture, further includes:
The noise image feature is superimposed in the denoising characteristics of image.
5. the method as described in claim 1, which is characterized in that the method also includes:
When loss of detail model convergence, the parameter of the updated convolutional neural networks is exported, to obtain having trained
At image denoising network.
6. a kind of network training device for image denoising, which is characterized in that described device includes:
Feature denoises unit, for the noise image of preset quantity to be inputted preset convolutional neural networks, initializes the volume
Product neural network parameter, by the convolutional neural networks denoising layer obtain noise image denoising characteristics of image, with after
It is continuous that the convolutional neural networks are trained；
Feature assembled unit, for being carried out the denoising characteristics of image by the combination layer of the preset convolutional neural networks
Combination, to obtain the denoising image after the noise image denoising；
Acquiring unit is lost, for the denoising image and the corresponding noise-free picture input of the noise image is preset thin
Loss model is saved, the details when convolutional neural networks denoise the noise image is obtained by the loss of detail model
Loss；And
Anti-pass updating unit, for being damaged the details by gradient anti-pass mode when the loss of detail model is not converged
Anti-pass is lost to the convolutional neural networks, and updates the parameter of the convolutional neural networks according to the loss of detail.
7. device as claimed in claim 6, which is characterized in that the loss acquiring unit includes:
First distance acquiring unit, for obtaining the denoising characteristic pattern of image and described by preset VGG neural network
The first Euclidean distance between the characteristic pattern of noise-free picture；
Second distance acquiring unit, for by High frequency filter obtain it is described denoising image frequency image information and the nothing make an uproar
The second Euclidean distance between the frequency image information of acoustic image；And
Loss obtains subelement, for according to formulaObtain the loss of detail, wherein ldetailIt indicates
The loss of detail, lperIndicate first Euclidean distance, lhfIndicate second Euclidean distance,Indicate second Europe
The weight of formula distance.
8. device as claimed in claim 6, which is characterized in that described device further include:
Feature extraction unit, for extracting the noise by the first default convolutional layer and active coating of the convolutional neural networks
The noise image feature of image；And
Feature denoise subelement, for the residual block by the preset convolutional neural networks to the noise image feature into
Row denoising, to obtain the denoising characteristics of image, the residual block includes the second default convolutional layer of the convolutional neural networks
And active coating.
9. a kind of computing terminal, including memory, processor and storage are in the memory and can be on the processor
The computer program of operation, which is characterized in that the processor realizes such as claim 1 to 5 when executing the computer program
The step of item the method.
10. a kind of computer readable storage medium, the computer-readable recording medium storage has computer program, and feature exists
In realization is such as the step of claim 1 to 5 the method when the computer program is executed by processor.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910100276.XA CN109872288B (en) | 2019-01-31 | 2019-01-31 | Network training method, device, terminal and storage medium for image denoising |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910100276.XA CN109872288B (en) | 2019-01-31 | 2019-01-31 | Network training method, device, terminal and storage medium for image denoising |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109872288A true CN109872288A (en) | 2019-06-11 |
| CN109872288B CN109872288B (en) | 2023-05-23 |
Family
ID=66918487
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910100276.XA Active CN109872288B (en) | 2019-01-31 | 2019-01-31 | Network training method, device, terminal and storage medium for image denoising |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109872288B (en) |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110349103A (en) * | 2019-07-01 | 2019-10-18 | 昆明理工大学 | It is a kind of based on deep neural network and jump connection without clean label image denoising method |
| CN110443758A (en) * | 2019-07-05 | 2019-11-12 | 广东省人民医院(广东省医学科学院) | A kind of medical imaging denoising method and device |
| CN110569961A (en) * | 2019-08-08 | 2019-12-13 | 合肥图鸭信息科技有限公司 | neural network training method and device and terminal equipment |
| CN110728636A (en) * | 2019-09-17 | 2020-01-24 | 杭州群核信息技术有限公司 | Monte Carlo rendering image denoising model, method and device based on generative confrontation network |
| CN110738616A (en) * | 2019-10-12 | 2020-01-31 | 成都考拉悠然科技有限公司 | image denoising method with detail information learning capability |
| CN111047537A (en) * | 2019-12-18 | 2020-04-21 | 清华大学深圳国际研究生院 | System for recovering details in image denoising |
| CN111161180A (en) * | 2019-12-26 | 2020-05-15 | 华南理工大学 | Deep learning ultrasonic image denoising method based on migration and structure prior |
| CN111369456A (en) * | 2020-02-28 | 2020-07-03 | 深圳市商汤科技有限公司 | Image denoising method and device, electronic device and storage medium |
| CN111626950A (en) * | 2020-05-19 | 2020-09-04 | 上海集成电路研发中心有限公司 | Online training device and method for image denoising model |
| CN112308785A (en) * | 2019-08-01 | 2021-02-02 | 武汉Tcl集团工业研究院有限公司 | Image denoising method, storage medium and terminal device |
| CN112330575A (en) * | 2020-12-03 | 2021-02-05 | 华北理工大学 | Convolution neural network medical CT image denoising method |
| CN112598599A (en) * | 2020-12-29 | 2021-04-02 | 南京大学 | Hyperspectral image denoising model training method and denoising method |
| CN112801888A (en) * | 2021-01-06 | 2021-05-14 | 杭州海康威视数字技术股份有限公司 | Image processing method, image processing device, computer equipment and storage medium |
| CN112967195A (en) * | 2021-03-04 | 2021-06-15 | 浙江大华技术股份有限公司 | Image denoising method and device and computer readable storage medium |
| CN113096023A (en) * | 2020-01-08 | 2021-07-09 | 字节跳动有限公司 | Neural network training method, image processing method and device, and storage medium |
| CN113538281A (en) * | 2021-07-21 | 2021-10-22 | 深圳大学 | Image denoising method and device, computer equipment and storage medium |
| US20220138910A1 (en) * | 2020-11-05 | 2022-05-05 | Leica Microsystems Cms Gmbh | Methods and systems for training convolutional neural networks |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108171217A (en) * | 2018-01-29 | 2018-06-15 | 深圳市唯特视科技有限公司 | A kind of three-dimension object detection method based on converged network |
| CN108230280A (en) * | 2018-04-11 | 2018-06-29 | 哈尔滨工业大学 | Image speckle noise minimizing technology based on tensor model and compressive sensing theory |
| CN108304921A (en) * | 2018-02-09 | 2018-07-20 | 北京市商汤科技开发有限公司 | The training method and image processing method of convolutional neural networks, device |
| CN108335306A (en) * | 2018-02-28 | 2018-07-27 | 北京市商汤科技开发有限公司 | Image processing method and device, electronic equipment and storage medium |
| CN108765320A (en) * | 2018-05-16 | 2018-11-06 | 哈尔滨工业大学 | Image restoration system based on multilevel wavelet convolutional neural networks |
| CN109063584A (en) * | 2018-07-11 | 2018-12-21 | 深圳大学 | Facial characteristics independent positioning method, device, equipment and the medium returned based on cascade |
| CN109118435A (en) * | 2018-06-15 | 2019-01-01 | 广东工业大学 | A kind of depth residual error convolutional neural networks image de-noising method based on PReLU |
-
2019
- 2019-01-31 CN CN201910100276.XA patent/CN109872288B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108171217A (en) * | 2018-01-29 | 2018-06-15 | 深圳市唯特视科技有限公司 | A kind of three-dimension object detection method based on converged network |
| CN108304921A (en) * | 2018-02-09 | 2018-07-20 | 北京市商汤科技开发有限公司 | The training method and image processing method of convolutional neural networks, device |
| CN108335306A (en) * | 2018-02-28 | 2018-07-27 | 北京市商汤科技开发有限公司 | Image processing method and device, electronic equipment and storage medium |
| CN108230280A (en) * | 2018-04-11 | 2018-06-29 | 哈尔滨工业大学 | Image speckle noise minimizing technology based on tensor model and compressive sensing theory |
| CN108765320A (en) * | 2018-05-16 | 2018-11-06 | 哈尔滨工业大学 | Image restoration system based on multilevel wavelet convolutional neural networks |
| CN109118435A (en) * | 2018-06-15 | 2019-01-01 | 广东工业大学 | A kind of depth residual error convolutional neural networks image de-noising method based on PReLU |
| CN109063584A (en) * | 2018-07-11 | 2018-12-21 | 深圳大学 | Facial characteristics independent positioning method, device, equipment and the medium returned based on cascade |
Non-Patent Citations (1)
| Title |
|---|
| 吴从中，等: "结合深度残差学习和感知损失的图像去噪", 《中国图像图形学报》 * |
Cited By (28)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110349103A (en) * | 2019-07-01 | 2019-10-18 | 昆明理工大学 | It is a kind of based on deep neural network and jump connection without clean label image denoising method |
| CN110443758A (en) * | 2019-07-05 | 2019-11-12 | 广东省人民医院(广东省医学科学院) | A kind of medical imaging denoising method and device |
| CN110443758B (en) * | 2019-07-05 | 2023-08-25 | 广东省人民医院(广东省医学科学院) | Medical image denoising method and device |
| CN112308785A (en) * | 2019-08-01 | 2021-02-02 | 武汉Tcl集团工业研究院有限公司 | Image denoising method, storage medium and terminal device |
| CN112308785B (en) * | 2019-08-01 | 2024-05-28 | 武汉Tcl集团工业研究院有限公司 | Image denoising method, storage medium and terminal equipment |
| CN110569961A (en) * | 2019-08-08 | 2019-12-13 | 合肥图鸭信息科技有限公司 | neural network training method and device and terminal equipment |
| CN110728636A (en) * | 2019-09-17 | 2020-01-24 | 杭州群核信息技术有限公司 | Monte Carlo rendering image denoising model, method and device based on generative confrontation network |
| US20220335574A1 (en) * | 2019-09-17 | 2022-10-20 | Hangzhou Qunhe Information Technology Co., Ltd. | A monte carlo rendering image denoising model, method and device based on generative adversarial network |
| CN110738616B (en) * | 2019-10-12 | 2020-06-26 | 成都考拉悠然科技有限公司 | Image denoising method with detail information learning capability |
| CN110738616A (en) * | 2019-10-12 | 2020-01-31 | 成都考拉悠然科技有限公司 | image denoising method with detail information learning capability |
| CN111047537A (en) * | 2019-12-18 | 2020-04-21 | 清华大学深圳国际研究生院 | System for recovering details in image denoising |
| CN111161180B (en) * | 2019-12-26 | 2023-09-26 | 华南理工大学 | Deep learning ultrasonic image de-noising method based on migration and structure priori |
| CN111161180A (en) * | 2019-12-26 | 2020-05-15 | 华南理工大学 | Deep learning ultrasonic image denoising method based on migration and structure prior |
| CN113096023B (en) * | 2020-01-08 | 2023-10-27 | 字节跳动有限公司 | Training method, image processing method and device for neural network and storage medium |
| CN113096023A (en) * | 2020-01-08 | 2021-07-09 | 字节跳动有限公司 | Neural network training method, image processing method and device, and storage medium |
| CN111369456A (en) * | 2020-02-28 | 2020-07-03 | 深圳市商汤科技有限公司 | Image denoising method and device, electronic device and storage medium |
| CN111626950A (en) * | 2020-05-19 | 2020-09-04 | 上海集成电路研发中心有限公司 | Online training device and method for image denoising model |
| US20220138910A1 (en) * | 2020-11-05 | 2022-05-05 | Leica Microsystems Cms Gmbh | Methods and systems for training convolutional neural networks |
| EP3996035A1 (en) * | 2020-11-05 | 2022-05-11 | Leica Microsystems CMS GmbH | Methods and systems for training convolutional neural networks |
| US12020405B2 (en) | 2020-11-05 | 2024-06-25 | Leica Microsystems Cms Gmbh | Methods and systems for training convolutional neural networks |
| CN112330575A (en) * | 2020-12-03 | 2021-02-05 | 华北理工大学 | Convolution neural network medical CT image denoising method |
| CN112598599A (en) * | 2020-12-29 | 2021-04-02 | 南京大学 | Hyperspectral image denoising model training method and denoising method |
| CN112598599B (en) * | 2020-12-29 | 2024-04-09 | 南京大学 | Denoising model training method and denoising method for hyperspectral image |
| CN112801888A (en) * | 2021-01-06 | 2021-05-14 | 杭州海康威视数字技术股份有限公司 | Image processing method, image processing device, computer equipment and storage medium |
| CN112967195A (en) * | 2021-03-04 | 2021-06-15 | 浙江大华技术股份有限公司 | Image denoising method and device and computer readable storage medium |
| CN112967195B (en) * | 2021-03-04 | 2024-04-23 | 浙江大华技术股份有限公司 | Image denoising method, device and computer readable storage medium |
| CN113538281A (en) * | 2021-07-21 | 2021-10-22 | 深圳大学 | Image denoising method and device, computer equipment and storage medium |
| CN113538281B (en) * | 2021-07-21 | 2023-07-11 | 深圳大学 | Image denoising method, image denoising device, computer equipment and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109872288B (en) | 2023-05-23 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109872288A (en) | For the network training method of image denoising, device, terminal and storage medium | |
| CN114140353B (en) | Swin-Transformer image denoising method and system based on channel attention | |
| Almeida et al. | Blind and semi-blind deblurring of natural images | |
| Guo et al. | Toward convolutional blind denoising of real photographs | |
| Wang et al. | An experimental-based review of image enhancement and image restoration methods for underwater imaging | |
| CN110599409B (en) | Convolutional neural network image denoising method based on multi-scale convolutional groups and parallel | |
| Ahn et al. | Block-matching convolutional neural network for image denoising | |
| Choudhury et al. | The trilateral filter for high contrast images and meshes | |
| CN111275643B (en) | Real noise blind denoising network system and method based on channel and space attention | |
| CN112270654A (en) | Image denoising method based on multi-channel GAN | |
| CN111028163A (en) | Convolution neural network-based combined image denoising and weak light enhancement method | |
| CN110807742B (en) | Low-light-level image enhancement method based on integrated network | |
| CN106204482B (en) | Based on the mixed noise minimizing technology that weighting is sparse | |
| CN110390646B (en) | Detail-preserving image denoising method | |
| CN111861939A (en) | Single image defogging method based on unsupervised learning | |
| CN112019704B (en) | Video denoising method based on prior information and convolutional neural network | |
| Paul et al. | Modified convolutional neural network with pseudo-CNN for removing nonlinear noise in digital images | |
| Ahn et al. | Block-matching convolutional neural network (BMCNN): improving CNN-based denoising by block-matched inputs | |
| CN111640077A (en) | Simple and efficient fuzzy text picture sharpening processing method | |
| CN113888405B (en) | Denoising and demosaicing method based on clustering self-adaptive expansion convolutional neural network | |
| KR102466061B1 (en) | Apparatus for denoising using hierarchical generative adversarial network and method thereof | |
| Tun et al. | Joint Training of Noisy Image Patch and Impulse Response of Low-Pass Filter in CNN for Image Denoising | |
| CN112634176B (en) | Mobile phone shot image denoising method based on multi-perception countermeasure generation network | |
| Gayathri et al. | A Survey on Image Denoising Algorithms (IDA) | |
| CN114663285A (en) | Old film super-resolution system based on convolutional neural network |
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 |