CN108765338A - Spatial target images restored method based on convolution own coding convolutional neural networks - Google Patents
Spatial target images restored method based on convolution own coding convolutional neural networks Download PDFInfo
- Publication number
- CN108765338A CN108765338A CN201810523868.8A CN201810523868A CN108765338A CN 108765338 A CN108765338 A CN 108765338A CN 201810523868 A CN201810523868 A CN 201810523868A CN 108765338 A CN108765338 A CN 108765338A
- Authority
- CN
- China
- Prior art keywords
- convolution
- image
- neural network
- layer
- input
- 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 35
- 238000013527 convolutional neural network Methods 0.000 title claims abstract description 19
- 238000003062 neural network model Methods 0.000 claims abstract description 11
- 238000004422 calculation algorithm Methods 0.000 claims description 66
- 238000013528 artificial neural network Methods 0.000 claims description 57
- 230000004913 activation Effects 0.000 claims description 26
- 230000006870 function Effects 0.000 claims description 18
- 238000005457 optimization Methods 0.000 claims description 15
- 230000009466 transformation Effects 0.000 claims description 13
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 11
- 230000008569 process Effects 0.000 claims description 10
- 238000010606 normalization Methods 0.000 claims description 8
- 230000000694 effects Effects 0.000 claims description 7
- 210000002569 neuron Anatomy 0.000 claims description 6
- 230000002441 reversible effect Effects 0.000 claims description 6
- 238000004364 calculation method Methods 0.000 claims description 5
- 238000000605 extraction Methods 0.000 claims description 5
- 238000000844 transformation Methods 0.000 claims description 5
- 238000005070 sampling Methods 0.000 claims description 4
- 230000005540 biological transmission Effects 0.000 claims description 3
- 230000008859 change Effects 0.000 claims description 3
- 230000007547 defect Effects 0.000 claims description 3
- 230000002401 inhibitory effect Effects 0.000 claims description 3
- 238000012886 linear function Methods 0.000 claims description 3
- 238000012216 screening Methods 0.000 claims description 3
- 238000005096 rolling process Methods 0.000 claims 1
- 238000012549 training Methods 0.000 abstract description 32
- 238000010586 diagram Methods 0.000 abstract description 7
- 230000008901 benefit Effects 0.000 abstract description 5
- 230000007850 degeneration Effects 0.000 abstract description 4
- 238000002474 experimental method Methods 0.000 description 15
- 238000013507 mapping Methods 0.000 description 8
- 238000011084 recovery Methods 0.000 description 8
- 238000011156 evaluation Methods 0.000 description 7
- 230000000875 corresponding effect Effects 0.000 description 4
- 238000013461 design Methods 0.000 description 4
- 235000013399 edible fruits Nutrition 0.000 description 4
- 230000001537 neural effect Effects 0.000 description 4
- 238000012360 testing method Methods 0.000 description 4
- 238000004458 analytical method Methods 0.000 description 3
- 230000007423 decrease Effects 0.000 description 3
- 210000005036 nerve Anatomy 0.000 description 3
- 238000011160 research Methods 0.000 description 3
- 230000003044 adaptive effect Effects 0.000 description 2
- 230000000739 chaotic effect Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000012795 verification Methods 0.000 description 2
- 230000000007 visual effect Effects 0.000 description 2
- 241000208340 Araliaceae Species 0.000 description 1
- 208000025679 Bell phenomenon Diseases 0.000 description 1
- 241000626238 Cepora Species 0.000 description 1
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 1
- 235000003140 Panax quinquefolius Nutrition 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 230000006835 compression Effects 0.000 description 1
- 238000007906 compression Methods 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 230000002596 correlated effect Effects 0.000 description 1
- 238000006731 degradation reaction Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 235000008434 ginseng Nutrition 0.000 description 1
- 230000010365 information processing Effects 0.000 description 1
- 238000007477 logistic regression Methods 0.000 description 1
- 210000004218 nerve net Anatomy 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 230000008447 perception Effects 0.000 description 1
- 230000002028 premature Effects 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 230000001869 rapid Effects 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 230000008439 repair process Effects 0.000 description 1
- 230000003362 replicative effect Effects 0.000 description 1
- 229920003987 resole Polymers 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of spatial target images restored methods based on convolution own coding convolutional neural networks, including：The degraded image with different degree of degenerations is built as input data, to learn and build more healthy and stronger CAE neural network models；Have the characteristics that high similarity between priori and model using known to extraterrestrial target limited amount, department pattern, constructs analogous diagram image set of the different type from different fog-levels, and for training in convolutional network.The advantage of the invention is that：Marginal texture clearly image can be restored；It goes turbulent flow to obscure ability with outstanding, and restores the image border contrast height, anti-noise ability is outstanding, and restored image internal structure, which is shown, more to be stablized clearly, more efficient.
Description
Technical field
The present invention relates to technical field of image processing, more particularly to the space based on convolution own coding convolutional neural networks
Target image restored method.
Background technology
The whole concept of neural network is to define an effective object module and one to weigh object module pros and cons
Loss function by successive optimization object module and minimizes the loss of object module, study input data and prediction data it
Between internal relation, to make neural network model complete various tasks.In the learning method for image restoration, it is assumed that
There are local correlations between figure for figure, on this basis by learning the feature of the degradation model and degraded image of image,
It will realize the recovery to degraded image.Wherein mainly there are the method based on rarefaction representation and the side based on deep neural network
Method, and based on the method for deep neural network due to its powerful nonlinear fitting ability, at present in super-resolution research work
Have been achieved for breakthrough progress.
Wherein paper (Dong C, Loy C C, He K, et al.Image Super-Resol ution Using Deep
Convolutional Networks[J].IEEE Transactions on Pattern Analysis and Machine
Intelligence,2016,38(2):295-307.) research team that Li Tangxiao gulls professor leads proposes based on 3 layers of convolution
The image super-resolution network (SRCNN, Super-Resolution Convolutional Neural Network) of layer,
Utilize the non-linear mapping capability of network, study low resolution to the end-to-end mapping between high-definition picture.Divide in article
The equivalence between SR methods and convolutional network based on traditional sparse coding has been analysed, and has proved to construct a kind of simple
Convolutional network restores the thought of sparse coding.Coding, mapping, decoded side are handled respectively different from traditional sparse coding
Method, neural network will automatically learn to sparse coding table, and optimize the work of 3 parts jointly.SRC NN neural network structures
Lightly, and in super-resolution reparation it improves quality better than traditional sparse coding method with medium trained degree, and also ensures
The degree that reality uses online can be reached in efficiency.The disadvantage is that structure is single, if using 3 layers of neural network of SRCNN, it will
No calligraphy learning is to the turbulent flow blurred picture of degree of degeneration complexity and the sparse expression mapping relations of clear image.
And in paper (Hradis M, Kotera J, Zemcik P, et al.Convolutional Neural
Networks for Direct Text Deblurring.[C].british machine visi on conference,
2015.) in research work, Michal Hradis et al. be directed to text invention shelves picture quality resume work on, instruction
The deeper 15 layers of convolutional network (L1 5-CNN) practised, by enough fuzzy and noise text image data training,
Will high quality graphic directly be reconstructed from low quality input picture, be carried out by the real pictures that various equipment are shot real
It tests, demonstrating the convolutional network of deep layer can learn and repair out of focus fuzzy with text invention shelves caused by camera shake.But
Difficulty is trained also can for the make of convolutional layer with reason in article and caused by indefinite and too deep network
It is doubled and redoubled, and is easy to bring the study consequence of over-fitting, and for different page orientations, font style and text language
On, the performance of neural network is not fine.
Invention content
The present invention in view of the drawbacks of the prior art, provides a kind of space based on convolution own coding convolutional neural networks
Target image restored method, can effectively solve the problem that the above-mentioned problems of the prior art.
In order to realize the above goal of the invention, the technical solution adopted by the present invention is as follows：
A kind of spatial target images restored method based on convolution own coding convolutional neural networks, including：
The CAE neural network models built, if f1, f2, f3, f4, f5 are the respective convolution kernel size of five convolutional layers,
If n1, n2, n3, n4, n5 are the convolution nuclear volume of five convolutional layers.CAE neural network models have the neuron for including 9 layers altogether,
And 1~5 layer is coding convolution, 6~9 layers are decoding convolution.The gray-scale map that the CAE network inputs sizes of structure are 32 × 32
Picture, wherein convolution working method are indicated with following weighted sum formula：
N represents the number of current layer neuron, XjRepresent j-th output knot of the present node to preceding i input data
Fruit, xiFor input image data, wijFor the convolution kernel of corresponding j-th of output, * is convolution operation, bjFor bias term.ReLu
(Rectified Linear Units) is the activation primitive of present invention structure Web vector graphic, and ReLu activation primitives are by positive and negative
The piecewise linear function that two parts are formed, all negative values are modified to 0 by it, and keep positive constant.The effect of ReLu
It is the unilateral transmission for inhibiting gradient, in order to ensure that the coding convolutional layer of CAE networks can be corresponded with decoding convolutional layer, and
Coding-decoded image can revert to the same size of input figure, need to grasp into row bound zero padding to being convolved input picture
Make, to ensure that characteristic pattern size is identical as input image size after convolution.
The calculating process of entire neural network is as follows：
Level 1 volume product will export the characteristic pattern of n1 a 32 × 32 to inputting after figure progress convolution, and letter is being activated by ReLu
After number corrects the maximum pondization operation with 2 × 2, then an image characteristics extraction and screening are completed, it will output behind pond
N1 16 × 16 characteristic patterns.Pond is handled usually as the Feature Selection after convolution operation, the purpose in maximum pond be in order to
More significant local feature statistic, and the size of energy compressive features figure are obtained, and reduces calculation amount.3rd layer of convolution is by the 1st
For the characteristic pattern of layer convolution and Chi Huahou as input value, which carries out convolution operation using the convolution kernel of n2 f2 × f2, and leads to
ReLu activation primitive linear transformations are crossed, convolution mode is same as above.2 × 2 maximum ponds are carried out in the characteristic pattern that convolution obtains, in this way
The size of convolution characteristic pattern reduces half, the characteristic pattern that output is n2 8 × 8 again.3rd layer of convolution uses the volume of n3 f3 × f3
Product core carries out convolution operation, and passes through ReLu activation primitive linear transformations.Such 1~5 layer of convolution will be extracted with pondization
Low-level image feature in artwork completes the coding Encode processes to inputting figure.
Followed by 6~9 parts decoding Decode, primary anti-pond is carried out first, or is known as up-sampling operation, is led to
The vertical and horizontal value of duplication is crossed, 8 × 8 characteristic pattern is expanded to 16 × 16.Then the 4th convolutional layer makees the characteristic pattern behind anti-pond
For input value, which does convolution operation using the convolution kernel of n4 f4 × f4, and carries out linear transformation using activation primitive.So
It is followed by the output of 4 layers of convolution, then carries out primary anti-pondization operation, 16 × 16 characteristic pattern is expanded to 32 × 32.5th layer of volume
Product is passing through the convolution nuclear convolution of f5 × f5 and the linear transformation of activation primitive using the characteristic pattern behind anti-pond as input value
Afterwards, decoded restored map will finally be obtained.
Further, select loss functions of the MSE as neural network, MSE that will correctly assess output figure and prognostic chart
Correspondence between pixel, formula are as follows：
M indicates that number of samples, x are input pictures, and y is output image, the wherein formula of mean square deviation MSE and peak value noise
Calculating than PSNR is inversely.PSNR values show that the image fault after repairing is smaller, closer to original graph, therefore it is excellent
The target for changing function is exactly that MSE is allowed to get minimum value as far as possible.
Further, present invention employs Adam optimization algorithms to carry out reverse train network weight.
Its function mode is similar with momentum.Parameter more new formula is：
Since Adam further improves algorithm speed, convergence rate faster, and is avoided that in other optimization algorithms and exists
Learning rate loss, the excessive defect of parameter update variance, in the performance that compared Different Optimization device, present invention employs
Adam optimization algorithms carry out reverse train network weight.
Further, the present invention selects following simple normalization algorithm to handle inputoutput data；
Simplest normalization algorithm, formula are as follows：
Y=(x-min) × (max-min) (4)
X is input data, and min and max is the minimum value and maximum value of x respectively, and y is the result after normalization.
Compared with prior art the advantage of the invention is that：
Can be utilized has height between priori and model known to extraterrestrial target limited amount, department pattern
The characteristics of spending similitude constructs analogous diagram image set of the different type from different fog-levels, and for being instructed in convolutional network
Practice.Neural network after training up will restore marginal texture and clearly scheme from the degraded image in test set
Picture.
It goes turbulent flow to obscure ability with outstanding, and restores the image border contrast height, anti-noise ability is better than existing
There is technology, restored image internal structure, which is shown, more to be stablized clearly.
More high efficiency will effectively learn the low of turbulent flow image by the network of end-to-end mapping learning training
Dimensional feature.
Description of the drawings
Fig. 1 is the convolution own coding Artificial Neural Network Structures of the embodiment of the present invention；
Fig. 2 is the training set schematic diagram of a part of neural network of the embodiment of the present invention；
Fig. 3 is psnr curve graphs trained under the different CAE Artificial Neural Network Structures of the embodiment of the present invention；
Fig. 4 is psnr curve graphs trained under the different convolution nuclear volumes of the embodiment of the present invention；
Fig. 5 is the part convolution nuclear shape signal in the 1st layer in the convolution own coding neural network of the embodiment of the present invention
Figure；
Fig. 6 is one group of restoration result schematic diagram in the moderate turbulent flow Degenerate Graphs of the emulation of the embodiment of the present invention；
Fig. 7 is one group of restoration result schematic diagram in the severe turbulent flow Degenerate Graphs of the emulation of the embodiment of the present invention.
Specific implementation mode
To make the objectives, technical solutions, and advantages of the present invention more comprehensible, below in conjunction with attached drawing and implementation is enumerated
Example, is described in further details the present invention.
Convolution autoencoder network is by encoding the low-dimensional data compressed in extraction image collection, and then decoded back arrives
Original image, to learn the correlated characteristic in image pattern automatically.Convolution autoencoder network first converts the data of input
To the space of low-dimensional, is then extended again and revert to approximate original image.This unsupervised learning method is usually used in acquisition one
The internal feature of serial associated data set removes the existence of redundant of input data, to obtain the figure with certain robustness
As low-dimensional feature.CAE neural network models such as Fig. 1 that the present invention is built.
Wherein f1, f2, f3, f4, f5 represent the respective convolution kernel size of five convolutional layers, and n1, n2, n3, n4, n5 are represented
The convolution nuclear volume of five convolutional layers.It is used in the present embodiment fuzzy after atmospheric turbulance interferes under short-time exposure
Image carries out image restoration experiment, many minutias of this seriously polluted missing image, but remains observed object
General outline, in order to extract this important information, experiment uses convolution own coding neural network model, allows neural network
Therefrom study filters out turbulent flow pollution part, and reconstruct the figure for being not affected by Turbulent Flow Effects to the low-dimensional feature of original image
Picture.Wherein, the characteristics of image after first half is encoded in Fig. 1, latter half decoded back go out original image.The present invention is set
The CAE neural network models of meter have the neuron for including 9 layers altogether, and 1~5 layer is coding convolution, and 6~9 layers are decoding convolution.
The gray level image that the CAE network inputs sizes that the present invention is built are 32 × 32, the wherein following weighted sum of convolution working method
Formula indicates：
N represents the number of current layer neuron, XjRepresent j-th output knot of the present node to preceding i input data
Fruit, xiFor input image data, wijFor the convolution kernel of corresponding j-th of output, * is convolution operation, bjFor bias term.ReLu
(Rectified Linear Units) is the activation primitive of present invention structure Web vector graphic, and ReLu activation primitives are by positive and negative
The piecewise linear function that two parts are formed, all negative values are modified to 0 by it, and keep positive constant.The effect of ReLu
It is the unilateral transmission for inhibiting gradient, it compared to other activation primitives there is stronger gradient to decline ability.Due to the ladder of ReLu
Degree has enough gradient magnitudes in non-negative region, so there is no what gradient disappeared to ask forever without carrying out Compression Correction
Topic, this advantage can help the convergence rate of network model to be maintained at relatively stable state.And letter is activated for others
Number, since the gradient near 0 is very small, causes the error propagation between predicted value and actual value that can constantly decay, deeper
Neural network will be more difficult to train or even premature with regard to deconditioning.Therefore in building network, hidden layer would generally select
With ReLu activation primitives, this can transmit the gradient for ensureing neural network always down.Since the principle of convolution is to convolution
Value in core frame does product and sums, and after finishing convolution, the characteristic pattern size of output can be reduced a convolution kernel, be
Ensure the coding convolutional layers of CAE networks and decodes that convolutional layer can correspond and coding-decoded image can revert to
The same size of input figure needs to operate into row bound zero padding to being convolved input picture, to ensure characteristic pattern ruler after convolution
It is very little identical as input image size.
The calculating process of entire neural network is as follows：Level 1 volume product to input scheme to carry out after convolution will output n1 32 ×
32 characteristic pattern then completes a characteristics of image after the maximum pondization operation by the amendment of ReLu activation primitives with 2 × 2
Extraction and screening, will n1 16 × 16 characteristic patterns of output behind pond.It is sieved usually as the feature after convolution operation in pond
Choosing is handled, the purpose in maximum pond be in order to obtain more significant local feature statistic, and the size of energy compressive features figure,
And reduce calculation amount.3rd layer of convolution using level 1 volume product and Chi Huahou characteristic pattern be used as input value, the layer use n2 f2 ×
The convolution kernel of f2 carries out convolution operation, and by ReLu activation primitive linear transformations, convolution mode is same as above.It is obtained in convolution
Characteristic pattern carries out 2 × 2 maximum ponds, and the size of such convolution characteristic pattern reduces half, the characteristic pattern that output is n2 8 × 8 again.
3rd layer of convolution carries out convolution operation using the convolution kernel of n3 f3 × f3, and passes through ReLu activation primitive linear transformations.Such 1
~5 layers of convolution will extract the low-level image feature in artwork with pondization, complete the coding Encode processes to inputting figure.It connects
Be 6~9 the parts decoding Decode, primary anti-pond is carried out first, or be known as up-sampling operation, by replicating anyhow
Value, 16 × 16 are expanded to by 8 × 8 characteristic pattern.Then the 4th convolutional layer, should using the characteristic pattern behind anti-pond as input value
Layer does convolution operation using the convolution kernel of n4 f4 × f4, and carries out linear transformation using activation primitive.Then 4 layers of convolution are connect
Output, then carry out primary anti-pondization and operate, 16 × 16 characteristic pattern is expanded to 32 × 32.5th layer of convolution will be behind anti-pond
Characteristic pattern will finally be obtained after the linear transformation by the convolution nuclear convolution of f5 × f5 and activation primitive as input value
Decoded restored map.
Allowable loss function
In order to calculate the fault tolerances value of individualized training sample, need that rational loss function is arranged for neural network.Damage
It is the calculation formula for assessing the extent of deviation between output valve and predicted value to lose function, and is entire neural network
Training objective function.For the neural network of classification problem, logistic regression loss function is mainly used for obtaining in training sample most
Big possible prediction distribution value.In the figure to figure map neural network that the present invention discusses, MSE is generally selected as nerve net
The loss function of network, MSE will correctly assess the correspondence between output figure and prognostic chart pixel, and formula is as follows：
M indicates that number of samples, x are input pictures, and y is output image, the wherein formula of mean square deviation MSE and peak value noise
Calculating than PSNR is inversely.PSNR values show that the image fault after repairing is smaller, closer to original graph, therefore it is excellent
The target for changing function is exactly that MSE is allowed to get minimum value as far as possible.
Design optimization device
Optimizer is used to update the weight of neural network model, selects correct optimizer that neural network can be allowed using most
Small frequency of training quickly searches out optimal solution, that is, converges on global minimum.Back-propagation algorithm in optimizer is nerve
Input signal is converted to the core concept of output signal by network, while being also the basic skills of training complex nonlinear function.
Back-propagation algorithm by the prediction error of output toward travel back, by calculating the gradient of error function and predicted value, and in ladder
Every layer of weight parameter is updated on degree negative direction.Wherein the more new direction Yu gradient direction of weight are on the contrary, neural in this way
Network will drop to local minimum according to gradient direction.
Present invention employs Adam optimization algorithms to carry out reverse train network weight.
Adaptive moments estimation algorithm (Adam, Adaptive Moment Estimation) is in the mode of renewal learning rate
On be improved (Kingma D P, Ba J L.Adam:A Meth od for Stochastic Optimization[J]
.international conference on learnin g representations, 2015.), Adam algorithms not only can
The average attenuation of gradient is stored, and previous gradient decaying can be calculated simultaneously, function mode is similar with momentum.Ginseng
Counting more new formula is：
Since Adam further improves algorithm speed, convergence rate faster, and is avoided that in other optimization algorithms and exists
Learning rate loss, the excessive defect of parameter update variance, in the performance that compared Different Optimization device, present invention employs
Adam optimization algorithms carry out reverse train network weight.
The pretreatment of data set
The value range very little of neural network output layer activation primitive, input data is too big or too small can all lead to nerve
Network output valve deviates normal range (NR).If input value is excessive, the convergence rate of neural network will reduce, therefore network is instructed
Practice target data to need to be remapped to a unified range, wherein it is most common one that data, which are normalized,
Kind preprocess method.
If using sigmoid and ReLu activation primitives in the output layer of neural network, due to sigmoid functions
Output valve is between [0,1], then the training data needs of neural network normalize in [0,1] section；If bis- using tanh
Curve activation primitive, training data will then normalize between [- 1,1].Wherein linear transformation formula is simplest normalization
Algorithm, formula are as follows：
Y=(x-min) × (max-min) (4)
X is input data, and min and max is the minimum value and maximum value of x respectively, and y is after normalizing as a result, above-mentioned public affairs
Image data is standardized as between section [0,1] range by formula.Neural network design based on figure to figure is upper usually using ReLu
As the activation primitive of output layer, otherwise training process is extremely slow, therefore for the convolution autoencoder network of the invention designed
Training select above simple normalization algorithm to handle inputoutput data.
Experiment and analysis
Experimental setup and parameter optimization
Experimental selection of the present invention is studied in Ubuntu systems, CPU frequency 3.3GHz, the RAM 8GB of computer,
GPU is GTX Titan X, and video memory size is 12GB, Keras development of neural networks kit of the experimental selection based on Theano.
The data set structure of experiment has used 233 satellite mappings that satellite modeling software generates, and zooms in and out rotation to them and expand
Data set is opened up, paper is used in combination[48]The Turbulence-degraded Images calculation formula provided calculates the rapids generated after degenerating on Matlab
It flows blurred picture, is cut into effective image part (black color part is background, needs to be deleted when constructing training set)
And the fuzzy graph of up to ten thousand 32 × 32 sizes is obtained, training set of the corresponding clear artwork as experiment will be fuzzy and clear
Clear fritter cutting image respectively as input data (input) and label data (label), be put into convolutional network constantly into
Row right value update.Wherein Fig. 2 is a part for the training set that the present invention constructs, and in corresponding one group of segment, left is label
Data correspond to original image；Right is input data, corresponds to blurred picture such as Fig. 2 of emulation.
Although the neural network constructed in experiment is only trained in 32 × 32 sizes, when the weights whole network
It moves on various sizes of image in use, only needing the image size for changing input terminal that can be carried out using convolutional network
Deblurring works, it should be noted that network has done pond down-sampling operation twice, and the image length and width of input must be 4 times
Number.
How size in order to study the convolutional layer under same turbulent flow Blur scale selects, to every layer of convolution in experiment
Core size f and convolution nuclear volume n are adjusted, and are compared in several groups of different neural network algorithms.Complete primary instruction
The neural network for practicing (1 epoch, which is equal to, has recycled primary entire training set) carries out image restoration on verification collection image, and
The average peak signal to noise ratio of the upper restored image of output verification collection, obtained Y-PSNR curve such as Fig. 3.
Wherein f and n corresponds to the convolution nuclear volume of the convolution kernel size and five convolutional layers of five convolutional layers in Fig. 1,
Middle n is defined as taking the same number entirely.It can find out from curve graph, when the convolution kernel number of neural network is set as 128,
Curve is that 300 or so can basically reach maximum value, and shake in follow-up training process obvious in epoch；And convolution kernel
After number is designed as 64, curve convergence it is more gentle, more stable PSNR can just be obtained by reaching 800 or so in epoch
Maximum value.And possessing the neural network model of different convolution kernel sizes, final PSNR maximum values also can be variant, in convolution
Core size f is respectively 13,3,3,3,13, and convolution kernel number n is the CAE model structures under 128, after completely training
PSNR values show best, and Y-PSNR is higher than remaining 5 kinds of model structure.Thus, it is possible to draw a conclusion：Construct bigger
The recovery effect of convolution kernel size f and convolution nuclear volume n, neural network can theoretically obtain better promotion, but promote degree
Can't be too big, after giving enough training times, the PSNR score values of 6 kinds of model structures can finally converge on it is close
26.0 up and down.
Next experiment needs to analyze in the case of fixed convolution kernel size, nerve under more convolution kernel numbers
Can network also have room for promotion, training curve such as Fig. 4 under the 4 kinds of CAE model structures constructed.It can find out when network convolution
Nuclear volume is promoted from 32 to after 128, and the convergent speed of neural network faster, and will eventually pass through 1000 circuit training
In neural network, the model structure PSNR maximum values under convolution kernel number n=128 are substantially higher in other two network
As a result；And after convolution kernel number n is promoted to 256, neural metwork training number can then decline always after more than 200, this says
It is bright in n=256, neural network overfitting training data causes network generalization to weaken, is testing instead
It is showed worse and worse on collection image.
Show to change by the convolution size and every layer of convolution nuclear volume of increase filter by multigroup contrast test
The Quality of recovery of kind image, the characteristic pattern estimation that convolutional network can be included in more surrounding pixels Encoder processes end are worked as
In, while the utilization information of the final image reduction of Decoder also can be more.But when the design of convolution nuclear volume is more than certain
After value, network can not can preferably reconstruct original image because of overlearning instead.
Compare the result such as table 1 of recovery time under different CAE model structures.
Table 1 restores individual figure and averagely takes
It can find out on recovery time, complicated network will spend the longer training time, and neural network reducing power
It is not obviously improved (Fig. 2), and network is susceptible to over-fitting (Fig. 3), bigger artificial neural is on individual figure
Treatment effeciency decreases
It is 9,3,3,3,9 that experiment, which is extracted in training convolutional core size f, when convolution nuclear volume n is respectively 128, CAE nets
A part of convolution kernel that the 1st layer of network, it is seen that after training up, neural network has learnt to good convolution karyomorphism
Shape, structure and the dictionary image learnt in sparse coding are very similar, such as Fig. 5.The convolution of neural network learning of the present invention
Core distribution is sparse and regular, and possesses different sizes and the form of different directions, this demonstrate that the network structure that the present invention designs
There is the ability of the characteristic information of good extraction image.
It is worth noting that, the blurred picture only constructed in the above experiment under smaller Blur scale is used as training
Collection, figure is to can be regarded as simple one-to-one mapping relations between figure.And under actual conditions, the air under short-time exposure
Turbulent flow degeneration convolution kernel scale can not be accurately known, and result between degraded image and real image that there are many-ones in this way
Mapping relations be difficult to normal if the fuzzy set that constructs can not include all possible turbulent flow Blur scale in experiment
Restore the degraded image being disturbed.Therefore experiment also needs to structure bigger and more diversified data set, allows neural network
Image low-dimensional feature can be correctly extracted under different scale, promote the generalization ability of network.Simultaneously in training set image
Noise jamming is added, the generation of over-fitting in network training process can be prevented.
Experimental result and analysis
The CAE neural networks that the present invention is built are trained on the image set of noise pollution.In contrast experiment,
It selects on blind restored image algorithm, adds two groups of neural network algorithms.Wherein there are Jia Jiaya team (Zhang J, Pan
J,Lai W,et al.Learning Fully Convolutional Networks for Iterative Non-Blind
Deconvolution [J].computer vision and pattern recognition,2016:3817-3825.) band
The fuzzy neural network algorithm (DCNN) that deconvolutes out of focus come, Michal et al.[72]What is proposed goes text to obscure convolutional Neural
Network algorithm (L15-CNN), restores to having with reference to Turbulence-degraded Images, wherein two groups of neural network algorithms all employ
Data set constructed by the present invention has carried out transfer training.Two groups of network knots of the CAE network structures that the present invention is built and comparison
Structure (X u L, Ren J S, Liu C, et al.Deep Convolutional Neural Network fo r Image
Deconvolution[C].neural information processing systems,2 014:1790-1798.)
(Hradis M,Kotera J,Zemcik P,et al.Convolutio nal Neural Networks for Direct
Text Deblurring. [C] .british machine vision conference, 2015.) such as table 2, wherein input figure
Gray image as being length and width m × n.
The structure table of 23 kinds of neural networks of table
This 5 kinds of mean absolute difference, signal-to-noise ratio, Y-PSNR, fidelity, mean square deviation of Experimental comparison has with reference to evaluating
Index.In order to accurately evaluate the restorability between various algorithms, experiment will be imitated more concerned with the vision in comparison restored image
Fruit, therefore not only to analyze the evaluation index of image, it is also necessary to observe the whole perception of image restoration result, comprehensive evaluation
The quality of each algorithm.
Wherein the restored image result of medium degree of degeneration and having with reference to evaluation index such as Fig. 6 and table 3.
The restoration result of 3 one groups of moderate turbulent flow Degenerate Graphs of table has with reference to evaluation result
Fig. 6 be subject in analogous diagram (b) moderate turbulent flow in short-term it is fuzzy with the knot for carrying out restoring experiment after poisson noise
Fruit.It can find out from figure (a), turbulent flow blurred picture has been lost most internal structural information, and is done by noise
After disturbing, edge contour acutance declines serious.From the visual effect of restored image, Jan algorithms produce in (d) figure
Serious shake bell phenomenon, noise is serious and picture structure is chaotic；(c) restored image of (e) (f) algorithm has in various degree
The noise of noise, algorithm (c) is the most serious, and noise (f) is showed in blocky；Although (e) (f) can be carried to a certain extent
High rim contrast, but noise resisting ability and bad；(g) algorithm can remove most noise, but blocky effect is serious, mould
Contour edge information is pasted；There is torsional deformation at the edge of restored image in DCNN algorithms (h), and internal noise is more tight
Weight, general image result are partially dark；Inventive algorithm (j) and the relatively other algorithms of L15-CNN algorithms (i), can not only enhance image
Contrast on border, and anti-noise ability is outstanding.
Evaluation index in contrast table 3 between 8 kinds of algorithms, it is seen that the index of 3 kinds of neural network algorithms all shows very
It is good.The algorithm index of wherein Dilip and Jan is worst, and secondly with BDTV algorithms, L0SR is then calculated in 5 groups of non-neural networks BDLIP
Peak value to-noise ratio index is best in method, it was demonstrated that its noise resisting ability is higher than remaining algorithm；And it is evaluated in 3 groups of neural network algorithms
In index, the performance of DCNN is worst, and since the selection of its convolution kernel is excessive, what is showed in turbulent flow recovery operation is not so good；And
Inventive algorithm and L15-CNN algorithm peak values to-noise ratio are all close to 25.5 or so, and inventive algorithm index outline is higher than
The index of L15-CNN algorithms.
On one group of heavy-degraded degree turbulent flow Degenerate Graphs, recovery effect and evaluation index such as Fig. 7 and the table 4 of each algorithm.
The restoration result of 4 one groups of severe turbulent flow Degenerate Graphs of table has with reference to evaluation result
Fig. 7 is the fuzzy recovery experimental result with poisson noise of turbulent flow in short-term for being subject to severe in analogous diagram (b).?
It can be seen that blurred picture is even more serious relative to Fig. 6 (a) fog-levels on Fig. 7 (a), the wheel of blurred picture can be only seen
It is wide.From Integral Restoration result, (d) algorithm of Jan is entirely ineffective in above 8 kinds of methods, and image result can not be distinguished
Recognize；There is more serious blocky phenomenon in L0SR algorithms (g), can not accurately describe object edge structure；Dilip algorithms (c)
It is serious with BDLIP algorithms (e) noise phenomenon；Although algorithm (f) has eliminated noise, also there is blocky phenomenon, illustrate it
The denoising mode of algorithm is not so good；And neural network algorithm is compared, algorithm (h) scalloping of DCNN restores too
Positive wing plate structure is slightly chaotic；Algorithm (i) clear-cut margin but internal structure of L15 is excessively smooth, and part details letter has been fallen in effacement
Breath；The present invention arithmetic result (j) ensure details do not lack in the case of, general image clear-cut margin, visual effect compared with
It is good.
In the comparison index of table 4, it is seen that 3 groups of neural network algorithms are higher than 5 groups of non-neural network algorithms on the whole
Index, show that neural network has certain advantage repairing the Turbulence-degraded Images with noise jamming, anti-noise ability is opposite
Non- neural network performs better than.The algorithm index of wherein Jan is worst, and the signal-to-noise ratio of L0SR is higher than the ranking Dilip of next and calculates
Method 10% or so, noise resisting ability is preferable；In the comparison index of neural network algorithm, the Y-PSNR of L15-CNN is most
It is low；L15-CNN and DCNN algorithms Y-PSNR is all close to 23 or so；Inventive algorithm index will be higher by DCNN algorithms 6%
More than, mean absolute difference is minimum on all comparison algorithms, shows the error smaller of restored image.
Those of ordinary skill in the art will understand that the embodiments described herein, which is to help reader, understands this
The implementation of invention, it should be understood that protection scope of the present invention is not limited to such specific embodiments and embodiments.This
The those of ordinary skill in field can make according to the technical disclosures disclosed by the invention various does not depart from of the invention essence
Various other specific variations and combinations, these variations and combinations are still within the scope of the present invention.
Those of ordinary skill in the art will understand that the embodiments described herein, which is to help reader, understands this
The implementation of invention, it should be understood that protection scope of the present invention is not limited to such specific embodiments and embodiments.This
The those of ordinary skill in field can make according to the technical disclosures disclosed by the invention various does not depart from of the invention essence
Various other specific variations and combinations, these variations and combinations are still within the scope of the present invention.
Claims (4)
1. a kind of spatial target images restored method based on convolution own coding convolutional neural networks, which is characterized in that including：
The CAE neural network models built, if f1, f2, f3, f4, f5 are the respective convolution kernel size of five convolutional layers, if n1,
N2, n3, n4, n5 are the convolution nuclear volume of five convolutional layers；CAE neural network models, which have altogether, includes 9 layers of neuron, and 1~5
Layer is coding convolution, and 6~9 layers are decoding convolution；The gray level image that the CAE network inputs sizes of structure are 32 × 32, wherein rolling up
Product working method is indicated with following weighted sum formula：
N represents the number of current layer neuron, XjPresent node is represented to j-th of output of preceding i input data as a result, xiFor
Input image data, wijFor the convolution kernel of corresponding j-th of output, * is convolution operation, bjFor bias term；ReLu is the present invention
The activation primitive of Web vector graphic is built, the piecewise linear function that ReLu activation primitives are made of positive and negative two parts, it is by institute
There is negative value to be modified to 0, and keeps positive constant；The effect of ReLu is the unilateral transmission for inhibiting gradient, in order to ensure CAE networks
Coding convolutional layer can be corresponded and coding-decoded image can revert to the same ruler of input figure with decoding convolutional layer
It is very little, it needs to operate into row bound zero padding to being convolved input picture, to ensure characteristic pattern size and input picture ruler after convolution
It is very little identical；
The calculating process of entire neural network is as follows：
Level 1 volume product will export the characteristic pattern of n1 a 32 × 32 to inputting after figure progress convolution, be repaiied by ReLu activation primitives
Just with after 2 × 2 maximum pondization operation, then an image characteristics extraction and screening are completed, it will output n1 behind pond
16 × 16 characteristic patterns；Pond is handled usually as the Feature Selection after convolution operation, and the purpose in maximum pond is to obtain
More significant local feature statistic, and the size of energy compressive features figure, and reduce calculation amount；3rd layer of convolution accumulates level 1 volume
And the characteristic pattern of Chi Huahou, as input value, which carries out convolution operation using the convolution kernel of n2 f2 × f2, and passes through ReLu
Activation primitive linear transformation, convolution mode are same as above；2 × 2 maximum ponds, such convolution feature are carried out in the characteristic pattern that convolution obtains
The size of figure reduces half, the characteristic pattern that output is n2 8 × 8 again；3rd layer of convolution is rolled up using the convolution kernel of n3 f3 × f3
Product operation, and pass through ReLu activation primitive linear transformations；Such 1~5 layer of convolution will extract the bottom in artwork with pondization
Layer feature completes the coding Encode processes to inputting figure；
Followed by 6~9 parts decoding Decode, primary anti-pond is carried out first, or is known as up-sampling operation, passes through duplication
Vertical and horizontal value, 16 × 16 are expanded to by 8 × 8 characteristic pattern；Then the 4th convolutional layer is using the characteristic pattern behind anti-pond as input
Value, which does convolution operation using the convolution kernel of n4 f4 × f4, and carries out linear transformation using activation primitive；Then 4 layers are connect
The output of convolution, then primary anti-pondization operation is carried out, 16 × 16 characteristic pattern is expanded to 32 × 32；5th layer of convolution is by anti-pond
Characteristic pattern after change is as input value, after the linear transformation by the convolution nuclear convolution of f5 × f5 and activation primitive, finally will
Obtain decoded restored map.
2. according to the method described in claim 1, it is characterized in that：Select loss functions of the MSE as neural network, MSE will
The correspondence between output figure and prognostic chart pixel can be correctly assessed, formula is as follows：
M indicates that number of samples, x are input pictures, and y is output image, the wherein formula of mean square deviation MSE and Y-PSNR PSNR
Calculating inversely；PSNR values show that the image fault after repairing is smaller, closer to original graph, therefore majorized function
Target is exactly that MSE is allowed to get minimum value as far as possible.
3. according to the method described in claim 2, it is characterized in that：It uses Adam optimization algorithms and carrys out reverse train network weight
Value；
Its function mode is similar with momentum；Parameter more new formula is：
Since Adam further improves algorithm speed, convergence rate faster, and is avoided that present in other optimization algorithms and learns
The excessive defect of rate loss, parameter update variance is practised, in the performance that compared Different Optimization device, it is excellent that present invention employs Adam
Change algorithm and carrys out reverse train network weight.
4. according to the method described in claim 3, it is characterized in that：Select following simple normalization algorithm to input and output number
According to being handled；
Simplest normalization algorithm, formula are as follows：
Y=(x-min) × (max-min) (4)
X is input data, and min and max is the minimum value and maximum value of x respectively, and y is the result after normalization.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810523868.8A CN108765338A (en) | 2018-05-28 | 2018-05-28 | Spatial target images restored method based on convolution own coding convolutional neural networks |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810523868.8A CN108765338A (en) | 2018-05-28 | 2018-05-28 | Spatial target images restored method based on convolution own coding convolutional neural networks |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108765338A true CN108765338A (en) | 2018-11-06 |
Family
ID=64003103
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810523868.8A Pending CN108765338A (en) | 2018-05-28 | 2018-05-28 | Spatial target images restored method based on convolution own coding convolutional neural networks |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108765338A (en) |
Cited By (24)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109509159A (en) * | 2018-11-20 | 2019-03-22 | 湖南湖工电气有限公司 | A kind of end-to-end restored method of UAV Fuzzy image based on deep learning |
| CN109543822A (en) * | 2018-11-29 | 2019-03-29 | 北京理工大学 | A kind of one-dimensional signal data recovery method based on convolutional neural networks |
| CN109671026A (en) * | 2018-11-28 | 2019-04-23 | 浙江大学 | Gray level image noise-reduction method based on empty convolution and automatic encoding and decoding neural network |
| CN109727209A (en) * | 2018-12-13 | 2019-05-07 | 北京爱奇艺科技有限公司 | A kind of method and device of determining incomplete historical relic complete image |
| CN109919864A (en) * | 2019-02-20 | 2019-06-21 | 重庆邮电大学 | A kind of compression of images cognitive method based on sparse denoising autoencoder network |
| CN110007355A (en) * | 2019-04-15 | 2019-07-12 | 中国科学院电子学研究所 | The detection method and device of a kind of convolution self-encoding encoder and interior of articles exception |
| CN110070498A (en) * | 2019-03-12 | 2019-07-30 | 浙江工业大学 | A kind of image enchancing method based on convolution self-encoding encoder |
| CN110111251A (en) * | 2019-04-22 | 2019-08-09 | 电子科技大学 | A kind of combination depth supervision encodes certainly and perceives the image super-resolution rebuilding method of iterative backprojection |
| CN110211017A (en) * | 2019-05-15 | 2019-09-06 | 北京字节跳动网络技术有限公司 | Image processing method, device and electronic equipment |
| CN110599416A (en) * | 2019-09-02 | 2019-12-20 | 太原理工大学 | Non-cooperative target image blind restoration method based on space target image database |
| CN110674334A (en) * | 2019-09-16 | 2020-01-10 | 南京信息工程大学 | Near-repetitive image retrieval method based on consistency region deep learning features |
| CN110831106A (en) * | 2019-11-14 | 2020-02-21 | 西安邮电大学 | Clustering method based on convolution |
| CN110974217A (en) * | 2020-01-03 | 2020-04-10 | 苏州大学 | Dual-stage electrocardiosignal noise reduction method based on convolution self-encoder |
| CN111402175A (en) * | 2020-04-07 | 2020-07-10 | 华中科技大学 | High-speed scanning imaging system and method |
| CN111598964A (en) * | 2020-05-15 | 2020-08-28 | 厦门大学 | Quantitative magnetic susceptibility image reconstruction method based on space adaptive network |
| CN111667526A (en) * | 2019-03-07 | 2020-09-15 | 西门子医疗有限公司 | Method and apparatus for determining size and distance of multiple objects in an environment |
| WO2020186888A1 (en) * | 2019-03-21 | 2020-09-24 | 深圳先进技术研究院 | Method and apparatus for constructing image processing model, and terminal device |
| CN112241668A (en) * | 2019-07-18 | 2021-01-19 | 杭州海康威视数字技术股份有限公司 | Image processing method, device and equipment |
| CN112261415A (en) * | 2020-10-23 | 2021-01-22 | 青海民族大学 | Image compression coding method based on overfitting convolution self-coding network |
| CN112669240A (en) * | 2021-01-22 | 2021-04-16 | 深圳市格灵人工智能与机器人研究院有限公司 | High-definition image restoration method and device, electronic equipment and storage medium |
| WO2021093718A1 (en) * | 2019-11-15 | 2021-05-20 | 北京金山云网络技术有限公司 | Video processing method, video repair method, apparatus and device |
| CN114152596A (en) * | 2021-11-30 | 2022-03-08 | 西华大学 | Method and device for measuring atmospheric turbulence generalized index parameter based on steepness parameter |
| CN114580285A (en) * | 2022-03-07 | 2022-06-03 | 哈尔滨理工大学 | Hyperbolic system model reduction method based on CAE network |
| CN116152054A (en) * | 2022-11-01 | 2023-05-23 | 海飞科(南京)信息技术有限公司 | Image super-resolution method for improving storage capacity and recall precision by using time iteration mode |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105512680A (en) * | 2015-12-02 | 2016-04-20 | 北京航空航天大学 | Multi-view SAR image target recognition method based on depth neural network |
| CN106056564A (en) * | 2016-05-27 | 2016-10-26 | 西华大学 | Edge sharp image fusion method based on joint thinning model |
| CN106251303A (en) * | 2016-07-28 | 2016-12-21 | 同济大学 | A kind of image denoising method using the degree of depth full convolutional encoding decoding network |
| CN106997581A (en) * | 2017-03-01 | 2017-08-01 | 杭州电子科技大学 | A kind of method that utilization deep learning rebuilds high spectrum image |
| CN107067396A (en) * | 2017-04-26 | 2017-08-18 | 中国人民解放军总医院 | A kind of nuclear magnetic resonance image processing unit and method based on self-encoding encoder |
| CN107180248A (en) * | 2017-06-12 | 2017-09-19 | 桂林电子科技大学 | Strengthen the hyperspectral image classification method of network based on associated losses |
| CN107251053A (en) * | 2015-02-13 | 2017-10-13 | 北京市商汤科技开发有限公司 | A kind of method and device for the compression artefacts for reducing lossy compression method image |
| US20180129974A1 (en) * | 2016-11-04 | 2018-05-10 | United Technologies Corporation | Control systems using deep reinforcement learning |
-
2018
- 2018-05-28 CN CN201810523868.8A patent/CN108765338A/en active Pending
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107251053A (en) * | 2015-02-13 | 2017-10-13 | 北京市商汤科技开发有限公司 | A kind of method and device for the compression artefacts for reducing lossy compression method image |
| CN105512680A (en) * | 2015-12-02 | 2016-04-20 | 北京航空航天大学 | Multi-view SAR image target recognition method based on depth neural network |
| CN106056564A (en) * | 2016-05-27 | 2016-10-26 | 西华大学 | Edge sharp image fusion method based on joint thinning model |
| CN106251303A (en) * | 2016-07-28 | 2016-12-21 | 同济大学 | A kind of image denoising method using the degree of depth full convolutional encoding decoding network |
| US20180129974A1 (en) * | 2016-11-04 | 2018-05-10 | United Technologies Corporation | Control systems using deep reinforcement learning |
| CN106997581A (en) * | 2017-03-01 | 2017-08-01 | 杭州电子科技大学 | A kind of method that utilization deep learning rebuilds high spectrum image |
| CN107067396A (en) * | 2017-04-26 | 2017-08-18 | 中国人民解放军总医院 | A kind of nuclear magnetic resonance image processing unit and method based on self-encoding encoder |
| CN107180248A (en) * | 2017-06-12 | 2017-09-19 | 桂林电子科技大学 | Strengthen the hyperspectral image classification method of network based on associated losses |
Non-Patent Citations (4)
| Title |
|---|
| NATHAN HUBENS: "Deep inside: Autoencoders", 《HTTPS://TOWARDSDATASCIENCE.COM/DEEP-INSIDE-AUTOENCODERS-7E41F319999F》 * |
| XIAO-JIAO MAO等: "Image Restoration Using Convolutional Auto-encoders with Symmetric Skip Connections", 《ARXIV:1606.08921V3》 * |
| 刘超等: "超低照度下微光图像的深度卷积自编码网络复原", 《光学精密工程》 * |
| 张喆: "雾霾天气下的集装箱自动识别系统关键技术研究", 《计算机产品与流通》 * |
Cited By (36)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109509159A (en) * | 2018-11-20 | 2019-03-22 | 湖南湖工电气有限公司 | A kind of end-to-end restored method of UAV Fuzzy image based on deep learning |
| CN109671026A (en) * | 2018-11-28 | 2019-04-23 | 浙江大学 | Gray level image noise-reduction method based on empty convolution and automatic encoding and decoding neural network |
| CN109543822A (en) * | 2018-11-29 | 2019-03-29 | 北京理工大学 | A kind of one-dimensional signal data recovery method based on convolutional neural networks |
| CN109727209A (en) * | 2018-12-13 | 2019-05-07 | 北京爱奇艺科技有限公司 | A kind of method and device of determining incomplete historical relic complete image |
| CN109919864A (en) * | 2019-02-20 | 2019-06-21 | 重庆邮电大学 | A kind of compression of images cognitive method based on sparse denoising autoencoder network |
| CN111667526A (en) * | 2019-03-07 | 2020-09-15 | 西门子医疗有限公司 | Method and apparatus for determining size and distance of multiple objects in an environment |
| CN110070498A (en) * | 2019-03-12 | 2019-07-30 | 浙江工业大学 | A kind of image enchancing method based on convolution self-encoding encoder |
| WO2020186888A1 (en) * | 2019-03-21 | 2020-09-24 | 深圳先进技术研究院 | Method and apparatus for constructing image processing model, and terminal device |
| CN110007355A (en) * | 2019-04-15 | 2019-07-12 | 中国科学院电子学研究所 | The detection method and device of a kind of convolution self-encoding encoder and interior of articles exception |
| CN110111251A (en) * | 2019-04-22 | 2019-08-09 | 电子科技大学 | A kind of combination depth supervision encodes certainly and perceives the image super-resolution rebuilding method of iterative backprojection |
| CN110111251B (en) * | 2019-04-22 | 2023-04-28 | 电子科技大学 | Image super-resolution reconstruction method combining depth supervision self-coding and perception iterative back projection |
| CN110211017A (en) * | 2019-05-15 | 2019-09-06 | 北京字节跳动网络技术有限公司 | Image processing method, device and electronic equipment |
| CN110211017B (en) * | 2019-05-15 | 2023-12-19 | 北京字节跳动网络技术有限公司 | Image processing method and device and electronic equipment |
| CN112241668A (en) * | 2019-07-18 | 2021-01-19 | 杭州海康威视数字技术股份有限公司 | Image processing method, device and equipment |
| CN110599416A (en) * | 2019-09-02 | 2019-12-20 | 太原理工大学 | Non-cooperative target image blind restoration method based on space target image database |
| CN110599416B (en) * | 2019-09-02 | 2022-10-11 | 太原理工大学 | Non-cooperative target image blind restoration method based on spatial target image database |
| CN110674334B (en) * | 2019-09-16 | 2020-08-11 | 南京信息工程大学 | Near-repetitive image retrieval method based on consistency region deep learning features |
| CN110674334A (en) * | 2019-09-16 | 2020-01-10 | 南京信息工程大学 | Near-repetitive image retrieval method based on consistency region deep learning features |
| CN110831106A (en) * | 2019-11-14 | 2020-02-21 | 西安邮电大学 | Clustering method based on convolution |
| CN110831106B (en) * | 2019-11-14 | 2021-08-20 | 西安邮电大学 | Clustering method based on convolution |
| WO2021093718A1 (en) * | 2019-11-15 | 2021-05-20 | 北京金山云网络技术有限公司 | Video processing method, video repair method, apparatus and device |
| CN110974217A (en) * | 2020-01-03 | 2020-04-10 | 苏州大学 | Dual-stage electrocardiosignal noise reduction method based on convolution self-encoder |
| CN110974217B (en) * | 2020-01-03 | 2022-08-09 | 苏州大学 | Dual-stage electrocardiosignal noise reduction method based on convolution self-encoder |
| CN111402175B (en) * | 2020-04-07 | 2022-04-08 | 华中科技大学 | High-speed scanning imaging system and method |
| CN111402175A (en) * | 2020-04-07 | 2020-07-10 | 华中科技大学 | High-speed scanning imaging system and method |
| CN111598964A (en) * | 2020-05-15 | 2020-08-28 | 厦门大学 | Quantitative magnetic susceptibility image reconstruction method based on space adaptive network |
| CN111598964B (en) * | 2020-05-15 | 2023-02-14 | 厦门大学 | Quantitative magnetic susceptibility image reconstruction method based on space adaptive network |
| CN112261415A (en) * | 2020-10-23 | 2021-01-22 | 青海民族大学 | Image compression coding method based on overfitting convolution self-coding network |
| CN112669240A (en) * | 2021-01-22 | 2021-04-16 | 深圳市格灵人工智能与机器人研究院有限公司 | High-definition image restoration method and device, electronic equipment and storage medium |
| CN112669240B (en) * | 2021-01-22 | 2024-05-10 | 深圳市格灵人工智能与机器人研究院有限公司 | High-definition image restoration method and device, electronic equipment and storage medium |
| CN114152596A (en) * | 2021-11-30 | 2022-03-08 | 西华大学 | Method and device for measuring atmospheric turbulence generalized index parameter based on steepness parameter |
| CN114152596B (en) * | 2021-11-30 | 2024-01-12 | 西华大学 | Method and device for measuring generalized index parameter of atmospheric turbulence based on sharpness parameter |
| CN114580285B (en) * | 2022-03-07 | 2022-11-01 | 哈尔滨理工大学 | Hyperbolic system model reduction method based on CAE network |
| CN114580285A (en) * | 2022-03-07 | 2022-06-03 | 哈尔滨理工大学 | Hyperbolic system model reduction method based on CAE network |
| CN116152054A (en) * | 2022-11-01 | 2023-05-23 | 海飞科(南京)信息技术有限公司 | Image super-resolution method for improving storage capacity and recall precision by using time iteration mode |
| CN116152054B (en) * | 2022-11-01 | 2024-03-01 | 海飞科(南京)信息技术有限公司 | Image super-resolution method for improving storage capacity and recall precision by using time iteration mode |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108765338A (en) | Spatial target images restored method based on convolution own coding convolutional neural networks | |
| Pathak et al. | Context encoders: Feature learning by inpainting | |
| CN113469356B (en) | Improved VGG16 network pig identity recognition method based on transfer learning | |
| Liu et al. | Learning discriminative representations from RGB-D video data | |
| CN110414498B (en) | Natural scene text recognition method based on cross attention mechanism | |
| CN110097519A (en) | Double supervision image defogging methods, system, medium and equipment based on deep learning | |
| CN110378208B (en) | Behavior identification method based on deep residual error network | |
| CN108765279A (en) | A kind of pedestrian's face super-resolution reconstruction method towards monitoring scene | |
| CN108520504A (en) | A kind of blurred picture blind restoration method based on generation confrontation network end-to-end | |
| CN109377459B (en) | Super-resolution deblurring method of generative confrontation network | |
| CN109410146A (en) | A kind of image deblurring algorithm based on Bi-Skip-Net | |
| CN113344806A (en) | Image defogging method and system based on global feature fusion attention network | |
| CN106204499A (en) | Single image rain removing method based on convolutional neural networks | |
| CN113392711A (en) | Smoke semantic segmentation method and system based on high-level semantics and noise suppression | |
| CN107563299A (en) | A kind of pedestrian detection method using ReCNN integrating context informations | |
| CN110503608A (en) | The image de-noising method of convolutional neural networks based on multi-angle of view | |
| Liu et al. | Low-quality license plate character recognition based on CNN | |
| Pires et al. | Image denoising using attention-residual convolutional neural networks | |
| CN113538258B (en) | Mask-based image deblurring model and method | |
| CN111242870A (en) | Low-light image enhancement method based on deep learning knowledge distillation technology | |
| CN111401209B (en) | Action recognition method based on deep learning | |
| CN116703750A (en) | Image defogging method and system based on edge attention and multi-order differential loss | |
| CN117078553A (en) | Image defogging method based on multi-scale deep learning | |
| CN106886819A (en) | A kind of improved method on restricted Boltzmann machine | |
| CN114155560B (en) | Light weight method of high-resolution human body posture estimation model based on space dimension reduction |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20181106 |
|
| RJ01 | Rejection of invention patent application after publication |