CN108805839A - Combined estimator image defogging method based on convolutional neural networks - Google Patents
Combined estimator image defogging method based on convolutional neural networks Download PDFInfo
- Publication number
- CN108805839A CN108805839A CN201810583489.8A CN201810583489A CN108805839A CN 108805839 A CN108805839 A CN 108805839A CN 201810583489 A CN201810583489 A CN 201810583489A CN 108805839 A CN108805839 A CN 108805839A
- Authority
- CN
- China
- Prior art keywords
- image
- neural network
- convolution
- transmissivity
- 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 44
- 238000013527 convolutional neural network Methods 0.000 title claims abstract description 21
- 238000013528 artificial neural network Methods 0.000 claims abstract description 39
- 239000003595 mist Substances 0.000 claims abstract description 36
- 238000012549 training Methods 0.000 claims abstract description 23
- 238000012545 processing Methods 0.000 claims abstract description 22
- 238000010606 normalization Methods 0.000 claims description 8
- 238000003475 lamination Methods 0.000 claims description 7
- 230000004913 activation Effects 0.000 claims description 6
- 239000000203 mixture Substances 0.000 claims description 4
- 239000011159 matrix material Substances 0.000 claims description 2
- 230000000694 effects Effects 0.000 description 12
- 230000006870 function Effects 0.000 description 10
- 230000015572 biosynthetic process Effects 0.000 description 7
- 238000013135 deep learning Methods 0.000 description 6
- 238000003786 synthesis reaction Methods 0.000 description 6
- 238000012360 testing method Methods 0.000 description 5
- 238000007796 conventional method Methods 0.000 description 3
- 230000008030 elimination Effects 0.000 description 3
- 238000003379 elimination reaction Methods 0.000 description 3
- 230000005540 biological transmission Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000011084 recovery Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000009738 saturating Methods 0.000 description 2
- 238000004088 simulation Methods 0.000 description 2
- 241001269238 Data Species 0.000 description 1
- 101100136092 Drosophila melanogaster peng gene Proteins 0.000 description 1
- 238000009825 accumulation Methods 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000012512 characterization method Methods 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 230000010365 information processing Effects 0.000 description 1
- 210000005036 nerve Anatomy 0.000 description 1
- 210000004218 nerve net Anatomy 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 238000010998 test method Methods 0.000 description 1
- 238000012876 topography Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Classifications
-
- G06T5/73—
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Abstract
The invention discloses a kind of single image to the fog method, mainly solve the problems, such as that prior art nonlinear fitting ability is constrained and training is cumbersome and applicable scene is single.Its scheme is：The convolutional neural networks that structure is made of the shared part of feature and mist figure air light value estimation branch and transmissivity estimation branch under Caffe frames；One group of fog free images collection J is obtained, J is carried out manually plus mist obtains foggy image collection I；I, J be bisected into the image group of multiple pairings respectively by batch size, and cycle input 200000 times to neural network is trained successively；The image I for needing defogging is input to the neural network of training completion, exports air light value A and transmissivity T；Fog free images J is calculated according to air light value A and transmissivity Tc.The present invention can keep restoring the contrast and color saturation of image, Y-PSNR and structural similarity two indices to be superior to the prior art well, can be used for the sharpening processing of foggy image.
Description
Technical field
The invention belongs to technical field of image processing, more particularly to a kind of single image to the fog method can be used for being imaged system
The sharpening processing for the single width foggy image shot of uniting.
Background technology
Haze etc. is boisterous in by current conditions influences, and the picture quality of imaging device capture is generally relatively low, often
It is low that there are contrasts, and hue shift, information can the low phenomenons of identification.These images to degrade not only influence the subjective sense of human eye
By more having seriously affected the performance of all kinds of intelligent vision information processing systems.Therefore, sharpening processing tool is carried out to Misty Image
There is very important actual application value.
Currently, image and video defogging algorithm based on atmospherical scattering model are the hot spots of research concern, critical issue
It is how to estimate atmosphere light and transmissivity.
Traditional defogging method extracts the feature of mist figure by various a priori assumptions, is then respectively designed with using this feature
The atmosphere light and transmissivity method of estimation of effect, to realize image defogging.Typical defogging method such as He et al. propose based on dark
The method (Dark ChannelPrior, DCP) of channel prior is shown in HE K, SUN J, TANG X.Single image haze
removal using dark channel prior[J].//IEEE transactions on pattern analysis
and machine intelligence,2011,33(12):2341-2353. this method thoughts are simple and effective, but are actually answering
In, for including the image of large area sky areas or larger white object, the defog effect of this method is limited,
Restore image and will produce more serious color distortion.
With flourishing for depth learning technology, the defogging method based on deep learning is increasingly becoming present image defogging
The research hotspot in field.The defogging method of early stage is all to estimate transmissivity merely with convolutional neural networks CNN, and for air
Light then directly uses conventional method, is first to air light value and transmission such as the method based on dark channel prior that He et al. is proposed
Rate is estimated, mist elimination image is calculated further according to the atmospherical scattering model for generating mist figure.Earliest deep learning defogging net
Network is the DehazeNet networks that Cai et al. is proposed, sees CAI B, XU X, JIA K, et al.Dehazenet:An end-
to-end system for single image haze removal[J]//IEEE Transactions on Image
Processing,2016,25(11):5187-5198. Web vector graphic generated datas have the image of mist as training set, input
Block exports the transmissivity of corresponding blocks via trained CNN networks.Although this method obtains good defog effect, by
In training data be topography's block so that occur color distortion in the processing of some mist figures and defogging be halfway existing
As.Then, Ren et al. proposes MSCNN networks, sees REN W, LIU S, ZHANG H, et al.Single image
dehazing via multi-scale convolutional neural networks[C]//European
conference on computer vision.2016:154-169. networks using entire image as input, using more
Scale convolutional neural networks estimate the transmissivity of whole figure, improve the accuracy of transmissivity estimation, obtain higher picture quality,
But this method still has three aspect deficiencies：1) training data is the synthesis mist figure based on off-the-air picture, is unable to accurate characterization
The characteristic of outdoor mist figure；2) in the network architecture pooling layers lead to network overall feeling with unpooling layers of connection type
It is wild smaller, cause network end-point single pixel point relatively low to the global information learning ability of input picture；3) for air light value
Estimation this method still use conventional method, estimation is not accurate enough, so as to cause restoring image section region there are defoggings not
Thorough and cross-color.
In order to solve the problems, such as that conventional method is estimated to be not allowed to atmosphere light, and avoid individually estimating transmissivity and atmosphere light
The error accumulation that brings influences, the defogging method occurred recently be mostly used uniformly convolutional neural networks and meanwhile estimate atmosphere light with
Transmissivity directly exports the image after defogging, realizes the processing of " end-to-end " from mist figure to defogging figure.Typical method packet
It includes：
1) the Bilinear networks that Yang et al. is proposed, are shown in YANG H, PAN J, YAN Q, et al.Image
Dehazing using Bilinear Composition Loss Function[J]//arXiv preprint arXiv:
1710.00279,2017. this method estimate transmissivity and atmosphere light respectively first with parallel sub-network, then by estimated
Transmissivity and air light value collective fit to mist elimination image, although this method can reduce individually estimation transmissivity and atmosphere light
The error of value brought, but network that two groups of convolution stack is used alone come when estimating transmissivity and air light value in it, it is non-thread
Property capability of fitting it is limited, in addition the method needs first to be trained transmissivity and atmosphere light estimation network, then to utilizing air
The network training of light value and transmissivity defogging, training complexity are higher；
2) AOD-Net that Li et al. people proposes, is shown in LI B, PENG X, WANG Z, et al.Aod-net:All-in-one
dehazing network[C]//Proceedings of the IEEE International Conference on
Computer Vision.2017:4770-4778. this method to atmospherical scattering model by carrying out fortran, by transmissivity
It is uniformly brought into atmosphere light in a new variable K, estimates K using convolutional neural networks, and then restore corresponding defogging figure
Picture.Though this method improves the precision of atmosphere light and transmissivity estimation to a certain extent, mist elimination image quality is improved,
Since network only only used the stacking of convolutional layer and bridge layer, depth affects the fitting of variable K to a certain extent, and
And the nonlinear fitting ability of network, also by certain constraint, the training data used when additionally, due to network training is only had family
Interior synthesis mist figure, it is undesirable for the recovery effects of outdoor mist figure.
In conclusion existing mist figure transmissivity and the neural network of air light value Combined estimator have the following disadvantages：One
It is that AOD-net networks only only used convolutional layer and bridging layer, the fitting of variation K；Second is that Bilinear networks individually make
Network is stacked with two-part convolution and estimates that transmissivity and atmosphere light, nonlinear fitting ability suffer restraints respectively, and training is multiple
It is miscellaneous and cumbersome；Third, only using indoor synthesis data carries out network training, it is not suitable for outdoor mist figure.
Invention content
It is an object of the invention to propose a kind of Combined estimator image defogging method based on convolutional neural networks, to solve
Prior art nonlinear fitting ability is constrained and training is cumbersome and is applicable in the too single problem of scene, and volume is fully utilized
Product neural network carries out defogging to image.
To achieve the above object, technical scheme of the present invention includes as follows：
The technical scheme is that：A kind of Combined estimator image defogging algorithm based on convolutional neural networks, feature
It is, includes the following steps：
1) structure is estimated by the shared part of feature and mist figure air light value estimation branch and transmissivity under Caffe frames
The convolutional neural networks framework of branch's composition, wherein:
The shared part of feature includes three convolution units, and air light value estimates that branch includes three convolutional layers, and transmissivity is estimated
The Zhi Yici that scores includes a pyramid pond module, a convolutional layer and a warp lamination and two convolutional layers；
Each convolution unit includes convolutional layer, normalizes layer, and each convolutional layer includes convolution operation and ReLU activation letters
Several layers, each warp lamination includes deconvolution operation and ReLU activation primitive layers, and deconvolution operation all has corresponding weights
Dn, convolution operation is all with corresponding weights WnWith bias Bn, normalization operation is with corresponding zoom factor γnAnd offset
Factor alphan；
2) one group of clear fog free images collection J is obtainedt, and to fog free images JtAccording to the air light value A of artificial settings and transmission
Rate T is carried out manually plus mist obtains foggy image collection It, by fog free images collection JtWith foggy image collection ItAs training image collection；
3) by training image collection It、JtThe image group for being bisected into multiple pairings respectively by batch size carries out image procossing, i.e.,
Input I simultaneouslyt、JtThe first image group training, obtain the initial weight W of each convolution operation of neural networknWith bias BnWith
The weights D of deconvolution operationnAnd the zoom factor γ of normalization operationnWith deviation ratio αn；
4) by weights WnWith bias BnWith the weights D of deconvolution operationnAnd zoom factor γnWith deviation ratio αnUsing
In the convolutional neural networks framework built to step 1, the updated neural network of the first subparameter is obtained；Again by the second image group
It is input to the updated neural network of the first subparameter, obtains the updated neural network of the second subparameter；And so on until
Last group of image group primary updated neural network before being input to, obtains the neural network for completing once to train；
5) again by It、JtAll image groups sequentially input to trained neural network is completed once, obtain completing instructing twice
Experienced neural network；And so on until all image groups are both input into 200000 times, obtain completing the skilled nerve net of institute
Network；
6) width is needed into the foggy image I that defogging is handledcIt is input to the skilled neural network of completion institute, passes through network
Atmosphere light estimation branch output air light value Ac, estimate that branch exports transmissivity T by the transmissivity of networkc；
7) according to 6) as a result, high quality fog free images are calculated：Jc=(Ic-Ac)/Tc+Ac。
Beneficial effects of the present invention are：
1) present invention avoids the prior art and estimates due to estimating transmissivity and air light value simultaneously by network according to mist figure
Cross-color and defogging caused by meter transmissivity and atmosphere light are not thorough problem；
2) present invention by convolutional neural networks due to solving the problems, such as defogging completely, and more efficient, the image after recovery is more clear
It is clear.
Simulation result shows the present invention and can compare now under the premise of keeping restoring the contrast and color saturation of image
Two methods of some AOD-Net and Bilinear Network have apparent defog effect, can preferably restore image
Background information improves visual effect；And Y-PSNR PNSR and structural similarity SSIM two indices are superior to the prior art.
Description of the drawings
Fig. 1 is the realization general flow chart of the present invention；
Fig. 2 is the transmissivity built in the present invention and air light value Combined estimator network structure；
Fig. 3 is with the present invention and the existing defog effect comparison diagram based on deep learning defogging algorithm pairing mist formation figure；
Fig. 4 is to be compared to the defog effect of true mist figure with the present invention and the existing defogging algorithm based on deep learning
Figure.
Specific implementation mode
The specific implementation mode of the present invention is described further below in conjunction with the accompanying drawings：
Referring to Fig.1, of the invention to be implemented as follows：
Step 1：Convolutional neural networks framework is built under Caffe frames.
As shown in Fig. 2, the neural network that the present invention is built shares part by feature, mist figure air light value estimates branch and saturating
Rate estimation branch three parts composition is penetrated, wherein:
The shared part of feature includes three convolution units, and each convolution unit includes convolutional layer and normalization operation, normalizing
Changing operation has corresponding zoom factor γnWith deviation ratio αn, the weights W of convolution operation in convolutional layernSize is followed successively by 7*7,
5*5,3*3, convolution step-length are 1；
Air light value estimates that branch includes three convolutional layers, and each convolutional layer includes convolution operation and ReLU activation primitives
Layer, convolution operation all have corresponding weights WnWith bias Bn, convolution weights WnSize is followed successively by 3*3,3*3,3*3, convolution step
Length is 1；
Transmissivity estimate branch successively include a pyramid pond module, a convolutional layer and a warp lamination and
Two convolutional layers, wherein：
The convolution weights W of three convolutional layersnSize is followed successively by 3*3,3*3,3*3, and convolution step-length is 1；
The warp lamination includes deconvolution operation and ReLU activation primitive layers, and deconvolution operation has corresponding weights
Dn, size 2*2, step-length 2；
The pyramid module is as follows handled input data：
The data I for being input to module 1a) is subjected to 0.5 times of maximum pondization operation, then it is 3*3, step to carry out convolution kernel size
A length of 1 convolution operation, then it is made comparisons with 0, the maximum value in the two is taken, then operation is normalized and obtains at normalization
Data I after reason1；
The data I for being input to module 1b) is subjected to 0.25 times of maximum pondization operation, then it is 5*5, step to carry out convolution kernel size
A length of 1 convolution operation, then it is made comparisons with 0, the two is maximized, then operation is normalized and deconvolution core size is
2*2, the deconvolution that step-length is 2 operate to obtain a deconvolution treated data I2；
The data I for being input to module 1c) is subjected to 0.125 times of maximum pondization operation, then it is 3*3 to carry out convolution kernel size,
Step-length is 1 convolution operation, then it is made comparisons with 0, and the two is maximized, then operation and deconvolution core size is normalized
For 4*4, the secondary counter convolution operation that step-length is 4 obtains the data I after secondary counter process of convolution3；
The data I for being input to module 1d) is subjected to 0.5 times of maximum pondization operation, obtains the data after the operation processing of pond
I4；
1e) by above-mentioned I1、I2、I3、I4This four data are spliced on fourth dimension degree, obtain final output data Q.
Step 2：Make training image collection.
9914 fog free images collection J of different scenes 2a) are downloaded from networkt, using bilinear interpolation algorithm by JtRuler
It is very little uniformly to zoom to 160 × 120；
2b) to fog free images collection JtThe corresponding depth of view information d of every image is estimated respectively using depth of field estimation CNN models
(Jt)；
2c) transmissivity parameter alpha (r) is generated at random between 1.0-2.5 using random functions, the saturating of every figure is calculated
Penetrate rate
2d) air light value A is generated at random between 0.8-1.0 using random functions, 12000 mist figure I are calculatedt=
JtT+A (1-T), as foggy image collection；
2e) by fog free images collection JtWith foggy image collection ItAs training image collection.
Step 3：Training neural network.
3a) build loss function：
3a1) using Euclidean distance formula as loss function, since there are atmosphere light estimation branches and transmissivity estimation for network
Branch, therefore there are two branch penalty functions, respectively：
Wherein ‖ ‖2To be operated to two norm of Matrix Calculating, m is the pixel number of input picture, An(I) it is neural network atmosphere light
Estimate the output of branch,For it is corresponding manually add mist to image when A, Tn(I) it is that neural network transmissivity estimates branch
Output,For it is corresponding manually add mist to image when T.
3a2) total losses function is obtained according to two branch penalty functions：
3b) loss function of network is set to Loss, by training image collection It、JtIt is bisected into respectively by batch size more
The image group of a pairing carries out image procossing, i.e., inputs I simultaneouslyt、JtThe training of the first image group, network, which passes through, calculates following letter
Number obtains the initial weight W of each convolution operation of neural networknWith bias BnWith the weights D of deconvolution operationnAnd normalization
The zoom factor γ of operationnWith deviation ratio αn：
Wherein functionRefer to so that above-mentioned total losses function obtains all independent variable W of its minimum valuen,
Bn,Dn,γn,αnSet；
3c) by weights WnWith bias BnWith the weights D of deconvolution operationnAnd zoom factor γnWith deviation ratio αnIt answers
With in the convolutional neural networks framework built to step 1, the updated neural network of the first subparameter is obtained；
The second image group 3d) is input to the updated neural network of the first subparameter again, obtains the update of the second subparameter
Neural network afterwards；
3e) and so on until primary updated neural network before last group of image group is input to, is completed
Once trained neural network；
3f) again by It、JtAll image groups sequentially input to trained neural network is completed once, obtain completing twice
Trained neural network；And so on until all image groups are both input into 200000 times, obtain completing the skilled nerve of institute
Network；
Step 4：Carry out image defogging.
A width 4a) is needed into the foggy image I that defogging is handledcIt is input to the skilled neural network of completion institute, passes through net
The atmosphere light estimation branch output air light value A of networkc, estimate that branch exports transmissivity T by the transmissivity of networkc；
High quality fog free images 4b) are calculated：Jc=(Ic-Ac)/Tc+Ac。
The effect of the present invention is further illustrated by following emulation：
1. test pictures：The three width synthesis mist figure and the true mist figure of four width and Make3D and NYU numbers downloaded from network
According to collection；
2. test method：Use existing Bilinear Net algorithms, AOD-Net algorithms, DehazeNet algorithms, MSCNN
Algorithm and totally five kinds of methods of the invention；
3. emulation testing content：
Emulation testing 1：Defogging processing is carried out using three width of above-mentioned five kinds of methods pair synthesis mist figure, the results are shown in Figure 3,
In：
Fig. 3 a are three width foggy images of synthesis,
Fig. 3 b be using DehazeNet algorithms to Fig. 3 a foggy images carry out defogging processing as a result,
Fig. 3 c be using MSCNN algorithms to Fig. 3 a foggy images carry out defogging processing as a result,
Fig. 3 d be using Bilinear Net algorithms to Fig. 3 a foggy images carry out defogging processing as a result,
Fig. 3 e be using AOD-Net algorithms to Fig. 3 a foggy images carry out defogging processing as a result,
Fig. 3 f be using the method for the present invention to Fig. 3 a foggy images carry out defogging processing as a result,
Fig. 3 g are three width fog free images；
From figure 3, it can be seen that being still had using the image that existing Bilinear Net algorithms and AOD-Net algorithms restore
A large amount of fog is still remained in the image restored using existing DehazeNet algorithms and MSCNN obvious thin
Mist.The image effect restored with the method for the present invention is superior to other four kinds and is based on deep learning defogging algorithm, closer to fogless
Image 3g.
Emulation testing 2：Carry out defogging processing using the true mist figure of four width of above-mentioned five kinds of methods pair, effect as shown in figure 4, its
In：
Fig. 4 a are the true foggy image of four width,
Fig. 4 b be using DehazeNet algorithms to Fig. 4 a foggy images carry out defogging processing as a result,
Fig. 4 c be using MSCNN algorithms to Fig. 4 a foggy images carry out defogging processing as a result,
Fig. 4 d be using Bilinear Net algorithms to Fig. 4 a foggy images carry out defogging processing as a result,
Fig. 4 e be using AOD-Net algorithms to Fig. 4 a foggy images carry out defogging processing as a result,
Fig. 4 f are the result for carrying out defogging processing to Fig. 4 a foggy images using the method for the present invention；
As can be seen from Figure 4, the image restored using existing Bilinear Net algorithms and AOD-Net algorithms is still had
A large amount of fog is still remained in the image restored using existing DehazeNet algorithms and MSCNN obvious thin
Mist.The image effect restored with the method for the present invention is superior to other four kinds and is based on deep learning defogging algorithm.
Emulation testing 3：Defogging processing is carried out to Make3D and NYU data sets using above-mentioned five kinds of methods, structure is similar
Property SSIM indexs and Y-PSNR PNSR indexs comparison, as shown in table 1
Table 1
It can be obtained by table 1, PSNR the and SSIM numerical value of the method for the present invention is all greater than or equal to other four kinds of algorithms, indicates
Mist treated picture and corresponding fog free images are closer.
In summary comparison of five kinds of algorithms on simulation result, defogging effect of the method for the present invention in all kinds of foggy images
Fruit is superior to other four kinds of algorithms.
Claims (7)
1. a kind of single image to the fog method based on convolutional neural networks, including：
1) structure estimates branch by the shared part of feature and mist figure air light value estimation branch and transmissivity under Caffe frames
The convolutional neural networks framework of composition, wherein:
The shared part of feature includes three convolution units, and air light value estimates that branch includes three convolutional layers, transmissivity estimation point
Zhi Yici includes a pyramid pond module, a convolutional layer and a warp lamination and two convolutional layers；
Each convolution unit includes convolutional layer, normalizes layer, and each convolutional layer includes convolution operation and ReLU activation primitive layers,
Each warp lamination includes deconvolution operation and ReLU activation primitive layers, and deconvolution operation all has corresponding weights Dn, volume
Product operation all has corresponding weights WnWith bias Bn, normalization operation is with corresponding zoom factor γnAnd deviation ratio
αn；
2) one group of clear fog free images collection J is obtainedt, and to fog free images JtAccording to the air light value A and transmissivity T of artificial settings
It carries out manually plus mist obtains foggy image collection It, by fog free images collection JtWith foggy image collection ItAs training image collection；
3) by training image collection It、JtThe image group for being bisected into multiple pairings respectively by batch size carries out image procossing, i.e., simultaneously
Input It、JtThe first image group training, obtain the initial weight W of each convolution operation of neural networknWith bias BnAnd warp
The weights D of product operationnAnd the zoom factor γ of normalization operationnWith deviation ratio αn；
4) by weights WnWith bias BnWith the weights D of deconvolution operationnAnd zoom factor γnWith deviation ratio αnIt is applied to step
In the convolutional neural networks framework of rapid 1 structure, the updated neural network of the first subparameter is obtained；The second image group is inputted again
To the updated neural network of the first subparameter, the updated neural network of the second subparameter is obtained；And so on until last
One group of image group primary updated neural network before being input to, obtains the neural network for completing once to train；
5) again by It、JtAll image groups sequentially input to trained neural network is completed once, obtain completing to train twice
Neural network；And so on until all image groups are both input into 200000 times, obtain completing the skilled neural network of institute；
6) width is needed into the foggy image I that defogging is handledcIt is input to the skilled neural network of completion institute, passes through the big of network
Gas light estimates branch output air light value Ac, estimate that branch exports transmissivity T by the transmissivity of networkc；
7) according to 6) as a result, high quality fog free images are calculated：Jc=(Ic-Ac)/Tc+Ac。
2. according to the method described in claim 1, it is characterized in that：Pyramid pond module in step 1), by following step
Suddenly data are handled：
The data I for being input to module 1a) is subjected to 0.5 times of maximum pondization operation, then it is 3*3, step-length 1 to carry out convolution kernel size
Convolution operation, then it is made comparisons with 0, takes the maximum value in the two, then be normalized after operation obtains normalized
Data I1；
The data I for being input to module 1b) is subjected to 0.25 times of maximum pondization operation, then it is 5*5 to carry out convolution kernel size, step-length is
1 convolution operation, then it is made comparisons with 0, the two is maximized, then it is 2*2 that operation and deconvolution core size, which is normalized,
The deconvolution that step-length is 2 operates to obtain a deconvolution treated data I2；
The data I for being input to module 1c) is subjected to 0.125 times of maximum pondization operation, then it is 3*3, step-length to carry out convolution kernel size
For 1 convolution operation, then it is made comparisons with 0, the two is maximized, then it is 4* that operation and deconvolution core size, which is normalized,
4, the secondary counter convolution operation that step-length is 4 obtains the data I after secondary counter process of convolution3；
The data I for being input to module 1d) is subjected to 0.5 times of maximum pondization operation, obtains the data I after the operation processing of pond4；
1e) by above-mentioned I1、I2、I3、I4This four data are spliced on fourth dimension degree, obtain final output data Q.
3. according to the method described in claim 1, it is characterized in that：The volume of three convolution units in sharing feature part in step 1)
Product weights Wn, size is followed successively by 7*7,5*5,3*3, and convolution step-length is 1.
4. according to the method described in claim 1, it is characterized in that：The volume of three convolutional layers of atmosphere light estimation branch in step 1)
Product weights Wn, size is followed successively by 3*3,3*3,3*3, and convolution step-length is 1.
5. according to the method described in claim 1, it is characterized in that：The volume of three convolutional layers of transmissivity estimation branch in step 1)
Product weights Wn, size is followed successively by 3*3,3*3,3*3, and convolution step-length is 1.
6. according to the method described in claim 1, it is characterized in that：The warp of transmissivity estimation branch warp lamination in step 1)
Product weights Dn, size 2*2, step-length 2.
7. according to the method described in claim 1, it is characterized in that：In step 3) I is inputted under Caffe frames simultaneouslyt、Jt's
First image group is trained, and the weights W of each convolution operation of neural network is obtainednWith bias BnWith the weights D of deconvolution operationn
And the zoom factor γ of normalization operationnWith deviation ratio αn, it is calculated by such as minor function：
Wherein, functionRefer to so that function obtains all independent variable W of its minimum valuen,Bn,Dn,γn,αnCollection
It closes, ‖ ‖2To be operated to two norm of Matrix Calculating, m is the pixel number of input picture, An(I) it is that neural network atmosphere light estimates branch
Output,For it is corresponding manually add mist to image when A, Tn(I) output of branch is estimated for neural network transmissivity,
For it is corresponding manually add mist to image when T.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810583489.8A CN108805839A (en) | 2018-06-08 | 2018-06-08 | Combined estimator image defogging method based on convolutional neural networks |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810583489.8A CN108805839A (en) | 2018-06-08 | 2018-06-08 | Combined estimator image defogging method based on convolutional neural networks |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108805839A true CN108805839A (en) | 2018-11-13 |
Family
ID=64087664
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810583489.8A Pending CN108805839A (en) | 2018-06-08 | 2018-06-08 | Combined estimator image defogging method based on convolutional neural networks |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108805839A (en) |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109493300A (en) * | 2018-11-15 | 2019-03-19 | 湖南鲲鹏智汇无人机技术有限公司 | The real-time defogging method of Aerial Images and unmanned plane based on FPGA convolutional neural networks |
| CN109584188A (en) * | 2019-01-15 | 2019-04-05 | 东北大学 | A kind of image defogging method based on convolutional neural networks |
| CN110223251A (en) * | 2019-06-02 | 2019-09-10 | 西安电子科技大学 | Suitable for manually with the convolutional neural networks underwater image restoration method of lamp |
| CN110310238A (en) * | 2019-06-18 | 2019-10-08 | 华南农业大学 | A kind of removing rain based on single image method based on the compression rewards and punishments neural network for reusing raw information |
| CN110738622A (en) * | 2019-10-17 | 2020-01-31 | 温州大学 | Lightweight neural network single image defogging method based on multi-scale convolution |
| CN110738623A (en) * | 2019-10-18 | 2020-01-31 | 电子科技大学 | multistage contrast stretching defogging method based on transmission spectrum guidance |
| CN111161161A (en) * | 2019-12-04 | 2020-05-15 | 天津大学 | Feature fusion defogging method for color preservation |
| CN111539250A (en) * | 2020-03-12 | 2020-08-14 | 上海交通大学 | Image fog concentration estimation method, system and terminal based on neural network |
| CN111553856A (en) * | 2020-04-24 | 2020-08-18 | 西安电子科技大学 | Image defogging method based on depth estimation assistance |
| CN111681180A (en) * | 2020-05-25 | 2020-09-18 | 厦门大学 | Priori-driven deep learning image defogging method |
| CN113066028A (en) * | 2021-03-31 | 2021-07-02 | 山东师范大学 | Image defogging method based on Transformer deep neural network |
| CN113706395A (en) * | 2020-05-21 | 2021-11-26 | 无锡科美达医疗科技有限公司 | Image defogging method based on antagonistic neural network |
| CN117237677A (en) * | 2023-11-15 | 2023-12-15 | 南京信息工程大学 | Precipitation prediction correction method for overall similarity of strong precipitation space based on deep learning |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106600560A (en) * | 2016-12-22 | 2017-04-26 | 福州大学 | Image defogging method for automobile data recorder |
| CN107749052A (en) * | 2017-10-24 | 2018-03-02 | 中国科学院长春光学精密机械与物理研究所 | Image defogging method and system based on deep learning neutral net |
| CN107945185A (en) * | 2017-11-29 | 2018-04-20 | 北京工商大学 | Image partition method and system based on wide residual pyramid pond network |
| CN107958465A (en) * | 2017-10-23 | 2018-04-24 | 华南农业大学 | A kind of single image to the fog method based on depth convolutional neural networks |
| KR20180050832A (en) * | 2016-11-07 | 2018-05-16 | 한국과학기술원 | Method and system for dehazing image using convolutional neural network |
-
2018
- 2018-06-08 CN CN201810583489.8A patent/CN108805839A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20180050832A (en) * | 2016-11-07 | 2018-05-16 | 한국과학기술원 | Method and system for dehazing image using convolutional neural network |
| CN106600560A (en) * | 2016-12-22 | 2017-04-26 | 福州大学 | Image defogging method for automobile data recorder |
| CN107958465A (en) * | 2017-10-23 | 2018-04-24 | 华南农业大学 | A kind of single image to the fog method based on depth convolutional neural networks |
| CN107749052A (en) * | 2017-10-24 | 2018-03-02 | 中国科学院长春光学精密机械与物理研究所 | Image defogging method and system based on deep learning neutral net |
| CN107945185A (en) * | 2017-11-29 | 2018-04-20 | 北京工商大学 | Image partition method and system based on wide residual pyramid pond network |
Non-Patent Citations (5)
| Title |
|---|
| CHONGYI LI 等: "A Cascaded Convolutional Neural Network for Single Image Dehazing", 《IEEE ACCESS》 * |
| WENQI REN 等: "Single Image Dehazing via Multi-Scale Convolutional Neural Networks", 《ECCV 2016》 * |
| XI ZHAO 等: "Deep fully convolutional regression networks for single image haze removal", 《2017 IEEE VISUAL COMMUNICATIONS AND IMAGE PROCESSING (VCIP)》 * |
| 袁保平: "基于卷积神经网络的图像去雾方法研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 邱东芳 等: "透射率和大气光自适应估计的暗通道去雾", 《计算机应用》 * |
Cited By (20)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109493300A (en) * | 2018-11-15 | 2019-03-19 | 湖南鲲鹏智汇无人机技术有限公司 | The real-time defogging method of Aerial Images and unmanned plane based on FPGA convolutional neural networks |
| CN109584188B (en) * | 2019-01-15 | 2022-11-11 | 东北大学 | Image defogging method based on convolutional neural network |
| CN109584188A (en) * | 2019-01-15 | 2019-04-05 | 东北大学 | A kind of image defogging method based on convolutional neural networks |
| CN110223251A (en) * | 2019-06-02 | 2019-09-10 | 西安电子科技大学 | Suitable for manually with the convolutional neural networks underwater image restoration method of lamp |
| CN110223251B (en) * | 2019-06-02 | 2023-03-24 | 西安电子科技大学 | Convolution neural network underwater image restoration method suitable for artificial and natural light sources |
| CN110310238A (en) * | 2019-06-18 | 2019-10-08 | 华南农业大学 | A kind of removing rain based on single image method based on the compression rewards and punishments neural network for reusing raw information |
| CN110738622A (en) * | 2019-10-17 | 2020-01-31 | 温州大学 | Lightweight neural network single image defogging method based on multi-scale convolution |
| CN110738623A (en) * | 2019-10-18 | 2020-01-31 | 电子科技大学 | multistage contrast stretching defogging method based on transmission spectrum guidance |
| CN111161161A (en) * | 2019-12-04 | 2020-05-15 | 天津大学 | Feature fusion defogging method for color preservation |
| CN111161161B (en) * | 2019-12-04 | 2023-05-16 | 天津大学 | Feature fusion defogging method for maintaining color |
| CN111539250A (en) * | 2020-03-12 | 2020-08-14 | 上海交通大学 | Image fog concentration estimation method, system and terminal based on neural network |
| CN111539250B (en) * | 2020-03-12 | 2024-02-27 | 上海交通大学 | Image fog concentration estimation method, system and terminal based on neural network |
| CN111553856A (en) * | 2020-04-24 | 2020-08-18 | 西安电子科技大学 | Image defogging method based on depth estimation assistance |
| CN111553856B (en) * | 2020-04-24 | 2023-03-24 | 西安电子科技大学 | Image defogging method based on depth estimation assistance |
| CN113706395A (en) * | 2020-05-21 | 2021-11-26 | 无锡科美达医疗科技有限公司 | Image defogging method based on antagonistic neural network |
| CN111681180A (en) * | 2020-05-25 | 2020-09-18 | 厦门大学 | Priori-driven deep learning image defogging method |
| CN111681180B (en) * | 2020-05-25 | 2022-04-26 | 厦门大学 | Priori-driven deep learning image defogging method |
| CN113066028A (en) * | 2021-03-31 | 2021-07-02 | 山东师范大学 | Image defogging method based on Transformer deep neural network |
| CN117237677A (en) * | 2023-11-15 | 2023-12-15 | 南京信息工程大学 | Precipitation prediction correction method for overall similarity of strong precipitation space based on deep learning |
| CN117237677B (en) * | 2023-11-15 | 2024-03-26 | 南京信息工程大学 | Precipitation prediction correction method for overall similarity of strong precipitation space based on deep learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108805839A (en) | Combined estimator image defogging method based on convolutional neural networks | |
| CN109785258B (en) | Face image restoration method based on multi-discriminator generated countermeasure network | |
| CN107492070B (en) | A kind of single image super-resolution calculation method of binary channels convolutional neural networks | |
| CN111861961B (en) | Single image super-resolution multi-scale residual error fusion model and restoration method thereof | |
| CN109584170B (en) | Underwater image restoration method based on convolutional neural network | |
| CN106920221B (en) | Take into account the exposure fusion method that Luminance Distribution and details are presented | |
| CN108921786A (en) | Image super-resolution reconstructing method based on residual error convolutional neural networks | |
| CN109300090A (en) | A kind of single image to the fog method generating network based on sub-pix and condition confrontation | |
| CN108269244A (en) | It is a kind of based on deep learning and prior-constrained image defogging system | |
| CN110223251B (en) | Convolution neural network underwater image restoration method suitable for artificial and natural light sources | |
| CN110544213A (en) | Image defogging method based on global and local feature fusion | |
| CN103886565A (en) | Nighttime color image enhancement method based on purpose optimization and histogram equalization | |
| CN108961227A (en) | A kind of image quality evaluating method based on airspace and transform domain multiple features fusion | |
| CN112308803A (en) | Self-supervision low-illumination image enhancement and denoising method based on deep learning | |
| CN112288630A (en) | Super-resolution image reconstruction method and system based on improved wide-depth neural network | |
| CN111462002B (en) | Underwater image enhancement and restoration method based on convolutional neural network | |
| CN109829925A (en) | A kind of method and model training method for extracting clean prospect in scratching figure task | |
| CN109598695B (en) | No-reference image fuzzy degree estimation method based on deep learning network | |
| CN112419163B (en) | Single image weak supervision defogging method based on priori knowledge and deep learning | |
| CN109685735B (en) | Single picture defogging method based on fog layer smoothing prior | |
| CN109671044B (en) | A kind of more exposure image fusion methods decomposed based on variable image | |
| CN108765337B (en) | Single color image defogging processing method based on dark channel prior and non-local MTV model | |
| CN112070691B (en) | Image defogging method based on U-Net | |
| CN111553856B (en) | Image defogging method based on depth estimation assistance | |
| CN111353964B (en) | Structure-consistent stereo image style migration method based on convolutional neural network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20181113 |