CN111968047A - Adaptive optical image blind restoration method based on generating type countermeasure network - Google Patents
Adaptive optical image blind restoration method based on generating type countermeasure network Download PDFInfo
- Publication number
- CN111968047A CN111968047A CN202010713133.9A CN202010713133A CN111968047A CN 111968047 A CN111968047 A CN 111968047A CN 202010713133 A CN202010713133 A CN 202010713133A CN 111968047 A CN111968047 A CN 111968047A
- Authority
- CN
- China
- Prior art keywords
- image
- network
- convolution
- layer
- clear image
- 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 52
- 230000003287 optical effect Effects 0.000 title claims abstract description 43
- 230000003044 adaptive effect Effects 0.000 title claims abstract description 41
- 238000012549 training Methods 0.000 claims abstract description 36
- 238000010606 normalization Methods 0.000 claims abstract description 7
- 238000007781 pre-processing Methods 0.000 claims abstract description 4
- 238000004422 calculation algorithm Methods 0.000 claims description 15
- 230000015556 catabolic process Effects 0.000 claims description 7
- 238000006731 degradation reaction Methods 0.000 claims description 7
- 238000010586 diagram Methods 0.000 claims description 7
- 238000004088 simulation Methods 0.000 claims description 6
- 238000004364 calculation method Methods 0.000 claims description 4
- 230000004913 activation Effects 0.000 claims description 3
- 230000003042 antagnostic effect Effects 0.000 claims description 3
- 150000001875 compounds Chemical class 0.000 claims description 3
- 230000000593 degrading effect Effects 0.000 claims description 3
- 238000011176 pooling Methods 0.000 claims description 3
- 230000036544 posture Effects 0.000 claims description 3
- 230000001172 regenerating effect Effects 0.000 claims description 3
- 238000009877 rendering Methods 0.000 claims description 3
- 239000000126 substance Substances 0.000 claims description 3
- 230000017105 transposition Effects 0.000 claims description 3
- 230000003247 decreasing effect Effects 0.000 claims description 2
- 230000006870 function Effects 0.000 description 26
- 230000008569 process Effects 0.000 description 10
- 238000003384 imaging method Methods 0.000 description 9
- 238000011084 recovery Methods 0.000 description 4
- 238000012937 correction Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 230000002238 attenuated effect Effects 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 238000001514 detection method Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 238000013136 deep learning model Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000004134 energy conservation Methods 0.000 description 1
- 230000001575 pathological effect Effects 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 238000005316 response function Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a self-adaptive optical image blind restoration method based on a generating countermeasure network, which is implemented according to the following steps: step 1, making a training set of a space target adaptive optical fuzzy image and a corresponding clear image; step 2, constructing a generation confrontation network model for training; step 3, inputting the adaptive optical fuzzy image and the training set of the corresponding clear image in the step 1 into the generation confrontation network model established in the step 2 to obtain a trained generation network model; and 4, carrying out size normalization preprocessing on the blurred image to be restored, inputting the blurred image to the generation network model trained in the step 3, and obtaining the restored sharp image. According to the self-adaptive optical image blind restoration method based on the generative confrontation network, due to the fact that PSF does not need to be estimated, single-frame image restoration efficiency is improved.
Description
Technical Field
The invention belongs to the technical field of optical image processing, and relates to a self-adaptive optical image blind restoration method based on a generating type confrontation network.
Background
When a space target is detected and imaged by the ground-based adaptive optical telescope, the optical wavefront distortion is caused due to the influence of factors such as atmospheric turbulence and the like, so that high-frequency information in the target image is greatly inhibited and attenuated, and the observed target image is seriously degraded and blurred.
When the ground-based optical telescope is used for aerial observation, due to the influence of factors such as atmospheric turbulence disturbance and the like, images acquired by an optical system generally have serious degradation blur. In order to obtain clearer images, most modern large telescopes adopt an Adaptive Optics (AO) system. Due to the adoption of the wavefront correction technology, the image quality of an image acquired by an AO system is improved to a great extent, however, the unit number and the speed of the AO system can not completely meet the correction requirement of atmospheric dynamic disturbance, and a correction residual exists. Therefore, the target high frequency information in the AO image is still suppressed and attenuated to a large extent. Such degraded blurred images are generally difficult to meet the requirement of aerial observation, and must be subjected to sharpening by means of image restoration.
In the field of image processing, the degraded imaging process is generally described as:
where x ═ x, y denotes image space coordinates, i (x) denotes a degraded imaging result, o (x) denotes an ideal sharp image, h (x) denotes a Point Spread Function (PSF) of the optical system, n (x) denotes optical path noise,representing convolution operation, for an ideal imaging system, namely, no degradation exists in the imaging process, the PSF is a unit pulse function, and the process of estimating a clear image o (x) according to a degraded image i (x) is called image deconvolution or image restoration, as shown in FIG. 6; if PSFh (x) is unknown, it is called blind image deconvolutionThe product or image is restored blindly.
In the actual detection imaging process of the space target, the PSF of the imaging optical system is often unknown, and in this case, a blind deconvolution method needs to be adopted to restore an ideal image.
The blind deconvolution of the image is a pathological problem, as shown in formula (1), two unknowns o (x) and h (x)) need to be solved through an equation, so the blind deconvolution technology has greater difficulty and uncertainty in mathematics, and physical constraints need to be continuously applied to the target image and the point spread function in the iteration process during the solution process to ensure the correct and unique solution.
The existing blind restoration technical scheme comprises:
(1) RL-IBD Algorithm
In recent years, foreign research on the ground-based optical image blind deconvolution technology in the field of spatial target detection and identification mainly focuses on how to apply more advanced physical constraints on a target and a PSF (pseudo-static function) to improve deconvolution performance, and no matter single-frame or multi-frame blind deconvolution, a basic framework is generally based on a classic Richardson-Lucy iterative blind deconvolution (RL-IBD) algorithm, the flow of the algorithm is shown in FIGS. 2 and 7, the RL-IBD algorithm absorbs the advantage of simple calculation of the IBD algorithm, and the RL algorithm maintains the characteristics of non-negativity and energy conservation of a restored image, so that a certain restoration effect is achieved. However, in blind deconvolution of an RL-IBD image, the recovery performance is greatly influenced by initial estimation of a point spread function support domain, the recovery effect is obviously reduced due to the fact that the support domain is small, the image recovery calculated amount is increased due to the fact that the support domain is large, and the algorithm efficiency is reduced.
(2) Blind restoration method based on sparse representation
Detail problems that need to be handled in the sparse representation-based image blind restoration method include Dictionary Learning (Dictionary Learning), Smooth Enhancement (Smooth Enhancement), and the like.
The common defects of the existing adaptive optical image blind restoration methods are as follows: 1. the clear image is estimated by estimating the PSF. 2. Before estimating the PSF and the sharp image, physical constraints (sparse constraints) need to be made to provide a priori information. 3. Alternate iteration is needed, the PSF and the clear image are continuously estimated, and the time for restoring one frame of image is long.
Disclosure of Invention
The invention aims to provide an adaptive optical image blind restoration method based on a generative countermeasure network, which improves the single-frame image restoration efficiency because the PSF does not need to be estimated.
The technical scheme adopted by the invention is that a self-adaptive optical image blind restoration method based on a generative confrontation network is implemented according to the following steps:
step 1, making a training set of a space target adaptive optical fuzzy image and a corresponding clear image;
step 2, constructing a generation confrontation network model for training;
step 3, inputting the adaptive optical fuzzy image and the training set of the corresponding clear image in the step 1 into the generation confrontation network model established in the step 2 to obtain a trained generation network model;
and 4, carrying out size normalization preprocessing on the blurred image to be restored, inputting the blurred image to the generation network model trained in the step 3, and obtaining the restored sharp image.
The present invention is also characterized in that,
the step 1 specifically comprises the following steps:
step 1.1, acquiring a real space target simulation 3D model data set;
step 1.2, rendering each target in the 3D model data set obtained in the step 1.1 under different postures to obtain a spatial target clear image data set;
and 1.3, simulating each clear image o (x) of the spatial target clear image data set by adopting a Zernike polynomial method to perform atmospheric turbulence degradation simulation to obtain a corresponding fuzzy observation image i (x) and obtain a training set of the spatial target adaptive optics fuzzy image i (x) and the corresponding clear image o (x).
The calculation method of the fuzzy observation image i (x) comprises the following steps:
wherein the content of the first and second substances,
in the formula (I), the compound is shown in the specification,for atmospheric turbulence degrading wavefront expressed by Zernike polynomial method, akIs a Zernike polynomial of the k termEach term ofReferred to as the wave front mode,1/pi in the unit circle domain, zero outside the circle domain, F is the fourier transform,is the polar axis in the polar coordinate system.
The step 2 specifically comprises the following steps: generating the antagonistic network model comprises generating a network GθGAnd discriminating network DθD;
Generating a network GθGThe structure of (1) is as follows: the method comprises 15 layers, wherein the 1 st layer is an input convolution layer, the size of a convolution kernel is 7 multiplied by 7, and the number of the convolution kernels is 64; the 2 nd layer to the 3 rd layer is a step convolution unit, the size of a convolution kernel is 3 multiplied by 3, the step is 2, and the number of the convolution kernels is 128 and 256 respectively; layers 4-12 are residual convolution units, each residual convolution unit comprises two convolution layers of 3 multiplied by 3; the 13 th layer to the 14 th layer are transposition convolution units, the size of a convolution kernel is 3 multiplied by 3, the step size is 2, and the number of the convolution kernels is 128 and 64 respectively; the 15 th layer is an output convolution layer, the size of the convolution kernel is 7 multiplied by 7, and the number of the convolution kernels is 64; except for the 15 th layer of output convolution, each convolution layer is followed by an example normalization unit and a ReLU activation unit;
discriminating network DθDThe structure of (1) is as follows: bag (bag)The multilayer convolution code comprises 5 layers, wherein the 1 st layer to the 3 rd layer are convolution layers, the size of a convolution kernel is 4 multiplied by 4, the step length is 2, and the number of the convolution kernels is respectively 64, 128 and 256; the 4 th layer and the 5 th layer are convolution layers, the size of the convolution kernel is 4 multiplied by 4, the step length is 1, and the number of the convolution kernels is respectively 512 and 1.
Generating loss functions against a network model by generating content loss functions for a networkPenalty function in connection with discriminant networkWeighting and composition:
where λ is a weight coefficient.
The content loss function for the generating network employs perceptual loss:
wherein phi isi,jIs a feature diagram output by the jth convolutional layer before the ith pooling layer in the convolutional network model structure trained in advance, Wi,jAnd Hi,jIs the dimension of the corresponding feature map.
Judging a countermeasure loss function of the network, adopting Wasserstein distance with gradient penalty as loss, not taking the probability of reconstructing a clear image as output, and calculating the loss according to the following formula:
where N is the number of batch samples.
The step 3 specifically comprises the following steps:
step 3.1, pretraining
The space target obtained in the step 1 is adaptive to opticsThe blurred image i (x) and the sharp image o (x) in the training set corresponding to the sharp image o (x) are sequentially input to the discrimination network D for generating the confrontation network modelθDPresetting a termination condition, and continuously iterating until the termination condition is reached, so that the discrimination network has basic discrimination capability;
step 3.2, formal training
Inputting the spatial target adaptive optics fuzzy image i (x) obtained in the step 1 and the spatial target adaptive optics fuzzy image i (x) in the training set corresponding to the clear image o (x) into a generation network for generating a confrontation model to obtain a restoration result, and respectively inputting the restoration result and the corresponding clear image o (x) into a pre-trained discrimination network DθDWhen judging the network DθDThe clear images o (x) and the restored images can be distinguished, and the training is continued to generate the network GθGWhen discriminating the network DθDIf the clear image o (x) and the restored image cannot be distinguished, the training is stopped, and the network training is finished.
Step 3.2, the restoration result and the corresponding clear image o (x) are respectively input into the pre-trained discrimination network DθDThe discriminator network discriminates the input image, discriminates true when a clear image is input and outputs 1, discriminates false when a restored image is input and outputs 0; when the restoration image is input, the judgment is true, and the output of the discriminator is 0.5;
if the restoration result and the corresponding clear image o (x) are inputted into the pre-trained discrimination network D respectivelyθDJudging true when inputting clear image, judging false when inputting restoration image, returning to continuous training to generate network, updating network parameter by gradient descent algorithm for loss function through appointed iteration number, regenerating restoration image, and inputting restoration result and corresponding clear image o (x) in step 3.2 to pre-trained judgment network DθDThe judgment is true until a clear image is input, and the judgment is true when a restored image is input;
if the restoration result and the corresponding clear image o (x) are inputted into the pre-trained discrimination network D respectivelyθDIn the method, when a clear image is input, the image is judged to be false, and the image is returnedTraining the discrimination network continuously, updating network parameters of the loss function by a gradient descent algorithm through appointed iteration times, returning to the step 3.2, and respectively inputting the recovery result and the corresponding clear image o (x) into the updated discrimination network DθDIn the above description, the determination is true until a clear image is input, and the determination is true when a restored image is input.
The invention has the advantages that
The invention relates to a self-adaptive optical image blind restoration method based on a generative confrontation network, which is an end-to-end blind restoration method. And training the network to achieve convergence by utilizing the countermeasure characteristic of the generated countermeasure network model and combining the content loss information of the generated restoration image. The fuzzy observation image to be restored is input into the generation model, so that a restored clear image with clearer target contour and more specific details can be generated, and the restoration precision of the target image and the restoration efficiency of the single-frame image are effectively improved.
Drawings
FIG. 1 is a schematic diagram illustrating the image degradation and restoration process in an adaptive optical image blind restoration method based on a generative countermeasure network according to the present invention;
FIG. 2 is a flow chart of a prior RL-IBD algorithm;
FIG. 3 is a flow chart of a method for blind restoration of an adaptive optical image based on a generative confrontation network according to the present invention;
FIG. 4 is a schematic diagram of a generated network model architecture in an adaptive optical image blind restoration method based on a generated countermeasure network according to the present invention;
FIG. 5 is a diagram of a fuzzy observation (left) and a blind reconstructed sharp image (right) based on GAN in the adaptive optical image blind reconstruction method based on the generative confrontation network according to the present invention;
FIG. 6 is a diagram illustrating a prior art image degradation and restoration process;
FIG. 7 is a flowchart of a prior art RL-IBD algorithm process.
Detailed Description
The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.
The invention discloses a self-adaptive optical image blind restoration method based on a generative confrontation network, the flow of which is shown in figure 3 and is implemented according to the following steps:
step 1, making a training set of a space target adaptive optical fuzzy image and a corresponding clear image;
as shown in fig. 1, specifically:
step 1.1, acquiring a real space target simulation 3D model data set;
step 1.2, rendering each target in the 3D model data set obtained in the step 1.1 under different postures to obtain a spatial target clear image data set;
and 1.3, simulating each clear image o (x) of the spatial target clear image data set by adopting a Zernike polynomial method to perform atmospheric turbulence degradation simulation to obtain a corresponding fuzzy observation image i (x) and obtain a training set of the spatial target adaptive optics fuzzy image i (x) and the corresponding clear image o (x).
The calculation method of the fuzzy observation image i (x) comprises the following steps:
wherein the content of the first and second substances,
in the formula (I), the compound is shown in the specification,for atmospheric turbulence degrading wavefront expressed by Zernike polynomial method, akIs a Zernike polynomial of the k termEach term ofReferred to as the wave front mode,1/pi in the unit circle domain, zero outside the circle domain, F is the fourier transform,is the polar axis in the polar coordinate system.
Step 2, constructing a generation confrontation network model for training; the method specifically comprises the following steps: generating the antagonistic network model comprises generating a network GθGAnd discriminating network DθD;
As shown in FIG. 4, a network G is generatedθGThe structure of (1) is as follows: the method comprises 15 layers, wherein the 1 st layer is an input convolution layer, the size of a convolution kernel is 7 multiplied by 7, and the number of the convolution kernels is 64; the 2 nd layer to the 3 rd layer is a step convolution unit, the size of a convolution kernel is 3 multiplied by 3, the step is 2, and the number of the convolution kernels is 128 and 256 respectively; layers 4-12 are residual convolution units, each residual convolution unit comprises two convolution layers of 3 multiplied by 3; the 13 th layer to the 14 th layer are transposition convolution units, the size of a convolution kernel is 3 multiplied by 3, the step size is 2, and the number of the convolution kernels is 128 and 64 respectively; the 15 th layer is an output convolution layer, the size of the convolution kernel is 7 multiplied by 7, and the number of the convolution kernels is 64; except for the 15 th layer of output convolution, each convolution layer is followed by an example normalization unit and a ReLU activation unit;
discriminating network DθDThe structure of (1) is as follows: comprises 5 layers, the 1 st layer to the 3 rd layer are convolution layers, the size of the convolution kernel is 4 multiplied by 4, the step length is 2, and the number is respectively 64, 128 and 256; the 4 th layer and the 5 th layer are convolution layers, the size of the convolution kernel is 4 multiplied by 4, the step length is 1, and the number of the convolution kernels is respectively 512 and 1.
Generating loss functions against a network model by generating content loss functions for a networkPenalty function in connection with discriminant networkWeighting and composition:
where λ is a weight coefficient.
The content loss function for the generating network employs perceptual loss:
wherein phi isi,jIs a feature diagram output by the jth convolutional layer before the ith pooling layer in the convolutional network model structure trained in advance, Wi,jAnd Hi,jIs the dimension of the corresponding feature map.
Judging a countermeasure loss function of the network, adopting Wasserstein distance with gradient penalty as loss, not taking the probability of reconstructing a clear image as output, and calculating the loss according to the following formula:
where N is the number of batch samples.
Step 3, inputting the adaptive optical fuzzy image and the training set of the corresponding clear image in the step 1 into the generation confrontation network model established in the step 2 to obtain a trained generation network model; the method specifically comprises the following steps:
step 3.1, pretraining
Sequentially inputting the spatial target adaptive optics fuzzy image i (x) obtained in the step 1 and the clear image o (x) in the training set corresponding to the clear image o (x) into a discrimination network D for generating a confrontation network modelθDPresetting a termination condition, and continuously iterating until the termination condition is reached, so that the discrimination network has basic discrimination capability;
step 3.2, formal training
The spatial target adaptive optics fuzzy image i (x) obtained in the step 1 and the spatial target adaptive optics fuzzy image i (x) in the training set corresponding to the clear image o (x)) Inputting into a generation network for generating a confrontation model to obtain a restoration result, and inputting the restoration result and corresponding clear images o (x) into a pre-trained discrimination network D respectively as shown in FIG. 5θDWhen judging the network DθDThe clear images o (x) and the restored images can be distinguished, and the training is continued to generate the network GθGWhen discriminating the network DθDIf the clear image o (x) and the restored image cannot be distinguished, the training is stopped, and the network training is finished.
Step 3.2, the restoration result and the corresponding clear image o (x) are respectively input into the pre-trained discrimination network DθDThe discriminator network discriminates the input image, discriminates true when a clear image is input and outputs 1, discriminates false when a restored image is input and outputs 0; when the restoration image is input, the judgment is true, and the output of the discriminator is 0.5;
if the restoration result and the corresponding clear image o (x) are inputted into the pre-trained discrimination network D respectivelyθDJudging true when inputting clear image, judging false when inputting restoration image, returning to continuous training to generate network, updating network parameter by gradient descent algorithm for loss function through appointed iteration number, regenerating restoration image, and inputting restoration result and corresponding clear image o (x) in step 3.2 to pre-trained judgment network DθDThe judgment is true until a clear image is input, and the judgment is true when a restored image is input;
if the restoration result and the corresponding clear image o (x) are inputted into the pre-trained discrimination network D respectivelyθDJudging the image as false when inputting clear image, returning to train and judge network, updating network parameter by gradient decreasing algorithm for loss function through appointed iteration number, returning to step 3.2, inputting the restored result and corresponding clear image o (x) to the updated judge network DθDIn the above description, the determination is true until a clear image is input, and the determination is true when a restored image is input.
And 4, carrying out size normalization preprocessing on the blurred image to be restored, inputting the blurred image to the generation network model trained in the step 3, and obtaining the restored sharp image.
Interpretation of terms:
generating a countermeasure network: a Generative Adaptive Networks (GAN) is a deep learning model, and is one of the most promising methods for unsupervised learning in complex distribution in recent years. The model passes through (at least) two modules in the framework: the mutual game learning of the Generative Model (Generative Model) and the Discriminative Model (Discriminative Model) yields a reasonably good output.
Adaptive optics: adaptive Optics (AO) is one of the most promising techniques to compensate in near real-time for Optical wavefront distortions in the imaging process caused by atmospheric turbulence or other factors.
Blind restoration of an image: the image blind restoration is an image processing method for estimating an original image and an imaging point spread function by using an observed degraded blurred image under the condition that the prior knowledge of the related original image and the imaging point spread function is unknown or incompletely known.
Point spread function: a Point Spread Function (PSF) refers to a light field distribution of an output image of an optical system when an input object is a Point light source, and is referred to as a Point Spread Function. Point sources can be represented mathematically by a function (point pulse) and the light field distribution of the output image is called the impulse response, so the point spread function is the impulse response function of the optical system.
Claims (9)
1. A self-adaptive optical image blind restoration method based on a generative confrontation network is characterized by comprising the following steps:
step 1, making a training set of a space target adaptive optical fuzzy image and a corresponding clear image;
step 2, constructing a generation confrontation network model for training;
step 3, inputting the adaptive optical fuzzy image and the training set of the corresponding clear image in the step 1 into the generation confrontation network model established in the step 2 to obtain a trained generation network model;
and 4, carrying out size normalization preprocessing on the blurred image to be restored, inputting the blurred image to the generation network model trained in the step 3, and obtaining the restored sharp image.
2. The adaptive optical image blind restoration method based on the generative countermeasure network according to claim 1, wherein the step 1 specifically comprises:
step 1.1, acquiring a real space target simulation 3D model data set;
step 1.2, rendering each target in the 3D model data set obtained in the step 1.1 under different postures to obtain a spatial target clear image data set;
and 1.3, simulating each clear image o (x) of the spatial target clear image data set by adopting a Zernike polynomial method to perform atmospheric turbulence degradation simulation to obtain a corresponding fuzzy observation image i (x) and obtain a training set of the spatial target adaptive optics fuzzy image i (x) and the corresponding clear image o (x).
3. The adaptive optical image blind restoration method based on the generative countermeasure network as claimed in claim 2, wherein the calculation method of the blurred observation image i (x) is as follows:
wherein the content of the first and second substances,
in the formula (I), the compound is shown in the specification,for atmospheric turbulence degrading wavefront expressed by Zernike polynomial method, akIs a Zernike polynomial of the k termEach term ofReferred to as the wave front mode,1/pi in the unit circle domain, zero outside the circle domain, F is the fourier transform,is the polar axis in the polar coordinate system.
4. The adaptive optical image blind restoration method based on the generative countermeasure network according to claim 2, wherein the step 2 is specifically: generating the antagonistic network model comprises generating a network GθGAnd discriminating network DθD;
The generation network GθGThe structure of (1) is as follows: the method comprises 15 layers, wherein the 1 st layer is an input convolution layer, the size of a convolution kernel is 7 multiplied by 7, and the number of the convolution kernels is 64; the 2 nd layer to the 3 rd layer is a step convolution unit, the size of a convolution kernel is 3 multiplied by 3, the step is 2, and the number of the convolution kernels is 128 and 256 respectively; layers 4-12 are residual convolution units, each residual convolution unit comprises two convolution layers of 3 multiplied by 3; the 13 th layer to the 14 th layer are transposition convolution units, the size of a convolution kernel is 3 multiplied by 3, the step size is 2, and the number of the convolution kernels is 128 and 64 respectively; the 15 th layer is an output convolution layer, the size of the convolution kernel is 7 multiplied by 7, and the number of the convolution kernels is 64; except for the 15 th layer of output convolution, each convolution layer is followed by an example normalization unit and a ReLU activation unit;
said discriminating network DθDThe structure of (1) is as follows: comprises 5 layers, the 1 st layer to the 3 rd layer are convolution layers, the size of the convolution kernel is 4 multiplied by 4, the step length is 2, and the number is respectively 64, 128 and 256; the 4 th layer and the 5 th layer are convolution layers, the size of the convolution kernel is 4 multiplied by 4, the step length is 1, and the number of the convolution kernels is respectively 512 and 1.
5. The adaptive optical image blind restoration method based on generative countermeasure network as claimed in claim 4, wherein the generative countermeasure networkLoss function of network model is generated by generating content loss function of networkPenalty function in connection with discriminant networkWeighting and composition:
where λ is a weight coefficient.
6. The adaptive optical image blind restoration method based on the generative countermeasure network as claimed in claim 5, wherein the content loss function of the generative network employs perceptual loss:
wherein phi isi,jIs a feature diagram output by the jth convolutional layer before the ith pooling layer in the convolutional network model structure trained in advance, Wi,jAnd Hi,jIs the dimension of the corresponding feature map.
7. The adaptive optical image blind restoration method based on the generative confrontation network as claimed in claim 5, wherein the confrontation loss function of the discriminant network adopts Wasserstein distance with gradient penalty as loss, which does not take probability of reconstructing a sharp image as output, and the loss is calculated as follows:
where N is the number of batch samples.
8. The adaptive optical image blind restoration method based on the generative countermeasure network according to claim 5, wherein the step 3 is specifically:
step 3.1, pretraining
Sequentially inputting the spatial target adaptive optics fuzzy image i (x) obtained in the step 1 and the clear image o (x) in the training set corresponding to the clear image o (x) into a discrimination network D for generating a confrontation network modelθDPresetting a termination condition, and continuously iterating until the termination condition is reached, so that the discrimination network has basic discrimination capability;
step 3.2, formal training
Inputting the spatial target adaptive optics fuzzy image i (x) obtained in the step 1 and the spatial target adaptive optics fuzzy image i (x) in the training set corresponding to the clear image o (x) into a generation network for generating a confrontation model to obtain a restoration result, and respectively inputting the restoration result and the corresponding clear image o (x) into a pre-trained discrimination network DθDWhen judging the network DθDThe clear images o (x) and the restored images can be distinguished, and the training is continued to generate the network GθGWhen discriminating the network DθDIf the clear image o (x) and the restored image cannot be distinguished, the training is stopped, and the network training is finished.
9. The adaptive optical image blind restoration method based on the generative confrontation network as claimed in claim 8, wherein the restoration result and the corresponding clear image o (x) are respectively input to the pre-trained discriminative network D in the step 3.2θDThe discriminator network discriminates the input image, discriminates true when a clear image is input and outputs 1, discriminates false when a restored image is input and outputs 0; when the restoration image is input, the judgment is true, and the output of the discriminator is 0.5;
if the restoration result and the corresponding clear image o (x) are inputted into the pre-trained discrimination network D respectivelyθDIf the clear image is input, the judgment is true, if the restored image is input, the judgment is false, and the method returns to the continuous trainerNetworking, updating network parameters of the loss function by a gradient descent algorithm through appointed iteration times, regenerating a restored image, and respectively inputting the restored result and the corresponding clear image o (x) into a pre-trained discrimination network D in the step 3.2θDThe judgment is true until a clear image is input, and the judgment is true when a restored image is input;
if the restoration result and the corresponding clear image o (x) are inputted into the pre-trained discrimination network D respectivelyθDJudging the image as false when inputting clear image, returning to train and judge network, updating network parameter by gradient decreasing algorithm for loss function through appointed iteration number, returning to step 3.2, inputting the restored result and corresponding clear image o (x) to the updated judge network DθDIn the above description, the determination is true until a clear image is input, and the determination is true when a restored image is input.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010713133.9A CN111968047A (en) | 2020-07-22 | 2020-07-22 | Adaptive optical image blind restoration method based on generating type countermeasure network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010713133.9A CN111968047A (en) | 2020-07-22 | 2020-07-22 | Adaptive optical image blind restoration method based on generating type countermeasure network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN111968047A true CN111968047A (en) | 2020-11-20 |
Family
ID=73362449
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010713133.9A Pending CN111968047A (en) | 2020-07-22 | 2020-07-22 | Adaptive optical image blind restoration method based on generating type countermeasure network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN111968047A (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112561828A (en) * | 2020-12-23 | 2021-03-26 | 北京环境特性研究所 | Gas turbulence fuzzy image reconstruction method based on generation countermeasure network |
CN113077540A (en) * | 2021-03-31 | 2021-07-06 | 点昀技术(南通)有限公司 | End-to-end imaging equipment design method and device |
CN113129232A (en) * | 2021-04-15 | 2021-07-16 | 中山大学 | Weak light speckle imaging recovery method based on countermeasure network generated by deep convolution |
CN114742779A (en) * | 2022-04-01 | 2022-07-12 | 中国科学院光电技术研究所 | High-resolution self-adaptive optical image quality evaluation method based on deep learning |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108520504A (en) * | 2018-04-16 | 2018-09-11 | 湘潭大学 | A kind of blurred picture blind restoration method based on generation confrontation network end-to-end |
CN108711141A (en) * | 2018-05-17 | 2018-10-26 | 重庆大学 | The motion blur image blind restoration method of network is fought using improved production |
CN110969589A (en) * | 2019-12-03 | 2020-04-07 | 重庆大学 | Dynamic scene fuzzy image blind restoration method based on multi-stream attention countermeasure network |
CN111275637A (en) * | 2020-01-15 | 2020-06-12 | 北京工业大学 | Non-uniform motion blurred image self-adaptive restoration method based on attention model |
-
2020
- 2020-07-22 CN CN202010713133.9A patent/CN111968047A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108520504A (en) * | 2018-04-16 | 2018-09-11 | 湘潭大学 | A kind of blurred picture blind restoration method based on generation confrontation network end-to-end |
CN108711141A (en) * | 2018-05-17 | 2018-10-26 | 重庆大学 | The motion blur image blind restoration method of network is fought using improved production |
CN110969589A (en) * | 2019-12-03 | 2020-04-07 | 重庆大学 | Dynamic scene fuzzy image blind restoration method based on multi-stream attention countermeasure network |
CN111275637A (en) * | 2020-01-15 | 2020-06-12 | 北京工业大学 | Non-uniform motion blurred image self-adaptive restoration method based on attention model |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112561828A (en) * | 2020-12-23 | 2021-03-26 | 北京环境特性研究所 | Gas turbulence fuzzy image reconstruction method based on generation countermeasure network |
CN113077540A (en) * | 2021-03-31 | 2021-07-06 | 点昀技术(南通)有限公司 | End-to-end imaging equipment design method and device |
CN113077540B (en) * | 2021-03-31 | 2024-03-12 | 点昀技术(南通)有限公司 | End-to-end imaging equipment design method and device |
CN113129232A (en) * | 2021-04-15 | 2021-07-16 | 中山大学 | Weak light speckle imaging recovery method based on countermeasure network generated by deep convolution |
CN114742779A (en) * | 2022-04-01 | 2022-07-12 | 中国科学院光电技术研究所 | High-resolution self-adaptive optical image quality evaluation method based on deep learning |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109859147B (en) | Real image denoising method based on generation of antagonistic network noise modeling | |
CN111968047A (en) | Adaptive optical image blind restoration method based on generating type countermeasure network | |
CN111028163A (en) | Convolution neural network-based combined image denoising and weak light enhancement method | |
Yin et al. | Highly accurate image reconstruction for multimodal noise suppression using semisupervised learning on big data | |
CN111861894B (en) | Image motion blur removing method based on generation type countermeasure network | |
CN110796616A (en) | Fractional order differential operator based L0Norm constraint and adaptive weighted gradient turbulence degradation image recovery method | |
CN109003234A (en) | For the fuzzy core calculation method of motion blur image restoration | |
CN115578262A (en) | Polarization image super-resolution reconstruction method based on AFAN model | |
CN112200752B (en) | Multi-frame image deblurring system and method based on ER network | |
Tian et al. | Underwater sonar image denoising through nonconvex total variation regularization and generalized Kullback–Leibler fidelity | |
CN112184567A (en) | Multi-channel blind identification adaptive optical image restoration method based on alternate minimization | |
CN112330549A (en) | Blind deconvolution network-based blurred image blind restoration method and system | |
Guan et al. | DiffWater: Underwater image enhancement based on conditional denoising diffusion probabilistic model | |
CN111047537A (en) | System for recovering details in image denoising | |
CN116188265A (en) | Space variable kernel perception blind super-division reconstruction method based on real degradation | |
CN114972081A (en) | Blind restoration-based image restoration method under complex optical imaging condition | |
CN114037636A (en) | Multi-frame blind restoration method for correcting image by adaptive optical system | |
CN112330550A (en) | Image restoration method and system based on image blind deconvolution technology | |
CN112001956A (en) | CNN-based schlieren strong laser far-field focal spot measurement image denoising method | |
Chen et al. | GADO-Net: an improved AOD-Net single image dehazing algorithm | |
CN117611456A (en) | Atmospheric turbulence image restoration method and system based on multiscale generation countermeasure network | |
CN114022385B (en) | Image restoration method based on local surface fitting | |
Gupta et al. | Blind restoration method for satellite images using memetic algorithm | |
CN112132758B (en) | Image restoration method based on asymmetric optical system point spread function model | |
CN113066023B (en) | SAR image speckle removing method based on self-calibration 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 |