CN108765344A - A method of the single image rain line removal based on depth convolutional neural networks - Google Patents
A method of the single image rain line removal based on depth convolutional neural networks Download PDFInfo
- Publication number
- CN108765344A CN108765344A CN201810538160.XA CN201810538160A CN108765344A CN 108765344 A CN108765344 A CN 108765344A CN 201810538160 A CN201810538160 A CN 201810538160A CN 108765344 A CN108765344 A CN 108765344A
- Authority
- CN
- China
- Prior art keywords
- rain
- layer
- picture
- high frequency
- detail
- 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 43
- 238000013527 convolutional neural network Methods 0.000 title claims abstract description 23
- 230000006870 function Effects 0.000 claims abstract description 24
- 238000013507 mapping Methods 0.000 claims abstract description 9
- 238000001914 filtration Methods 0.000 claims abstract description 5
- 238000003475 lamination Methods 0.000 claims description 37
- 238000012549 training Methods 0.000 claims description 14
- 238000005516 engineering process Methods 0.000 claims description 12
- 230000004913 activation Effects 0.000 claims description 10
- 230000015556 catabolic process Effects 0.000 claims description 3
- 230000017105 transposition Effects 0.000 claims description 3
- 238000013461 design Methods 0.000 abstract description 8
- 238000012545 processing Methods 0.000 abstract description 6
- 238000013135 deep learning Methods 0.000 abstract description 3
- 230000000694 effects Effects 0.000 description 7
- 238000011084 recovery Methods 0.000 description 5
- 230000015572 biosynthetic process Effects 0.000 description 3
- 238000003786 synthesis reaction Methods 0.000 description 3
- 230000000007 visual effect Effects 0.000 description 3
- 241000283973 Oryctolagus cuniculus Species 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 230000003321 amplification Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000009792 diffusion process Methods 0.000 description 1
- 230000008030 elimination Effects 0.000 description 1
- 238000003379 elimination reaction Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000003199 nucleic acid amplification method Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000002194 synthesizing effect Effects 0.000 description 1
- 238000012876 topography Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/77—Retouching; Inpainting; Scratch removal
-
- 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
-
- 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
- G06N3/08—Learning methods
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (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)
- Image Processing (AREA)
Abstract
The method for the single image rain line removal based on depth convolutional neural networks that the invention discloses a kind of, this method will be decomposed into lower frequency reference layer and high frequency detail layer using guiding filtering with rain figure picture first, then according to image processing field knowledge modified objective function, and by the deep learning network architecture of the high frequency detail layer In-put design with rain figure picture, to learn it and with the mapping between the clear image high frequency detail layer corresponding to rain figure picture.Finally by network exported remove rain after high frequency detail layer be added with the lower frequency reference layer with rain figure picture, obtain removal rain line after clear image.The detail section of image after the present invention remains rain while removing single image moderate rain line so that image definition greatly improves.
Description
Technical field
The present invention relates to a kind of single image rain line minimizing technology, and in particular to a kind of based on depth convolutional neural networks
Single image rain line minimizing technology.
Background technology
Removing rain based on single image research is one of the important directions in image restoration field, is widely used in object identification, mesh
The fields such as mark tracking.However, a large amount of rain lines quickly moved are dispersed in rainy day environment by random so that reflection and refraction
Phenomenon is present in target object with background light, causes the contrast of image to reduce, image blur, detailed information is lost, and
Show that clearly image is very difficult, it is therefore desirable to which recovery processing is carried out to the single image with rain line.
Existing removing rain based on single image line method is broadly divided into two classes.Problem is considered as image layer resolution problem by one kind.It is main
To include that structural similarity constrains, the methods of broad sense low-rank model.Another kind of is the method based on diffusion or based on filter, such as
Non-local mean is the methods of smooth.In recent years, acquired in learning of nonlinear functions ability due to convolutional neural networks (CNN)
Superiority, the problem of some methods based on CNN are also made to solve removing rain based on single image.Although Existing methods have achieved
Some successes, but there are the limitations of following two aspects：(1) for existing many methods, basic operation is small
The processing of rain line is carried out in acceptance region or topography's block, it will usually ignore influence and go between the acceptance region or receptive field of rain effect
Spatial context information.(2) since background texture structure and rain line are internal superpositions, existing most methods are to image
In rainless region domain also carried out that rain is gone to handle, cause restore image there are excess smoothness phenomenons.
Invention content
Goal of the invention：For overcome the deficiencies in the prior art, the present invention provides a kind of based on depth convolutional neural networks
Single image rain line minimizing technology, this method, which can solve existing picture contrast when removing rain based on single image, to be reduced, is imaged mould
The problem of paste, detailed information are lost.
Technical solution：Single image rain line minimizing technology of the present invention based on depth convolutional neural networks, the party
Method includes the following steps：
(1) use guiding filtering method that will be decomposed into lower frequency reference layer and high frequency detail layer with rain figure picture；
(2) mesh is constructed according to 2 norms between the high frequency detail layer with rain figure picture and the high frequency detail layer of clean image
Scalar functions, and L2 regularization terms are added in object function；
(3) network architecture for building a removing rain based on single image based on depth convolutional neural networks, including 4 convolution
Layer is denoted as the 1st convolutional layer, the 2nd convolutional layer, the 3rd convolutional layer, the 4th convolutional layer, and is swashed using network after each convolutional layer respectively
Function living, 4 warp laminations are denoted as the 1st warp lamination respectively, the 2nd warp lamination, the 3rd warp lamination, the 4th warp lamination, and
Network activation function, 3 jump connections, respectively by the 1st convolutional layer and the 3rd deconvolution are used after preceding 3 warp laminations
Layer composition jump connection, by the 2nd convolutional layer and the 2nd warp lamination composition jump connection, by the 3rd convolutional layer and the 1st
The composition jump connection of warp lamination；
(4) using the data set with rain figure picture as training data, it is input to the list based on depth convolutional neural networks
Width image goes in the network architecture of rain to be trained iteration, and is directed to each iteration, updates institute using stochastic gradient descent algorithm
State network parameter；
(5) training iteration after, by the datum layer with rain figure picture with remove rain line after high frequency detail layer be added revert to it is dry
Net image.
Preferably, in step (2), the object function is expressed as：
Wherein, N is the number of the image block after the picture breakdown with rain, and n is thumbnail, and W is network parameter,
IdetailAnd JdetailThe high frequency detail layer with the clear image corresponding with its of rain figure picture, fw () function stand network are indicated respectively
Body function,Regularization is punished for L_2, and λ is coefficient of balance.
Preferably, in step (3), the network activation function uses Tanh activation primitives.
Preferably, in step (3), the 1st convolutional layer, the feature of the 2nd convolutional layer, the 3rd convolutional layer, the 4th convolutional layer is reflected
It is 128 to penetrate number, and convolution kernel size is respectively set to 9*9,3*3,3*3,3*3.
Preferably, in step (3), the 1st warp lamination, the 2nd warp lamination, the 3rd warp lamination, the 4th warp lamination
Feature Mapping number is respectively 128,128,128,3, and the size of convolution kernel is respectively set to 3*3,3*3,3*3,1*1.
Preferably, the network parameter more new formula is：
Wherein, b is the bias term in network parameter, and W is network parameter, and s indicates that an iteration, α indicate that learning rate, T are
Transposition operator, IdetailAnd JdetailThe high frequency detail layer with the clear image corresponding with its of rain figure picture is indicated respectively.
Preferably, in step (5), it is described by the datum layer with rain figure picture with removal rain line after high frequency detail layer be added it is extensive
Again at clean image, it is formulaically expressed as：
E=fW(Idetail)+Ibase
Wherein, IdetailIndicate the high frequency detail layer with rain figure picture, IbaseIndicate the datum layer with rain figure picture.
Advantageous effect：Compared with prior art, the present invention its remarkable advantage is：1, the network architecture level used is deeper,
Convolution filter it is smaller, Feature Mapping reduce, so that the parameter of network is greatly decreased, and contribute to excavate more details letter
Breath and elimination rain line；2, to true rain figure and synthesis rain figure go rain effect all very significantly, improve picture quality and calculate imitate
It is better than other advanced methods in terms of rate；3, according to image processing field knowledge modified objective function, certain constraint is added,
Reduce parameter amount.
Description of the drawings
Fig. 1 is method flow diagram of the present invention；
Fig. 2 is the network architecture schematic diagram of the removing rain based on single image based on depth convolutional neural networks；
Fig. 3 is the method for the invention and other rain removing method design sketch under composograph environment, wherein Fig. 3 a, figure
3b, Fig. 3 c and Fig. 3 d are respectively harbour, umbrella, rabbit and the rain removing method of bird design sketch；
Fig. 4 is rain removing method of the present invention and other rain removing method design sketch under true picture environment, Fig. 4 a, figure
4b and Fig. 4 c are the design sketch under different images；
Fig. 5 is under the true picture environment described in Fig. 4, and rain removing method of the present invention is put with other rain removing method regions
Big design sketch, Fig. 5 a, Fig. 5 b and Fig. 5 c correspond to the region amplification effect figure at Fig. 4 a, Fig. 4 b and Fig. 4 c environment respectively.
Specific implementation mode
Embodiment
As shown in Figure 1, method shown in the present invention will be decomposed into lower frequency reference layer using guiding filtering with rain figure picture first
With high frequency detail layer.Then according to image processing field knowledge modified objective function, and the high frequency detail layer with rain figure picture is defeated
In the deep learning network architecture for entering design, to learn it and between the clear image high frequency detail layer corresponding to rain figure picture
Mapping.Finally by network exported remove rain after high frequency detail layer be added with the lower frequency reference layer with rain figure picture, obtain removal rain
Clear image after line.The detail section of image, makes after the present invention remains rain while removing single image moderate rain line
Image definition is obtained to greatly improve.
(1) image scaling down processing is carried out
Datum layer and levels of detail will be decomposed into rain figure picture using guiding filtering method first, wherein datum layer remains low
Frequency essential information, levels of detail include then high frequency detail part, such as rain line and other marginal informations.When rain figure removes datum layer
Afterwards, remaining details layer segment is very sparse.Sparse training set can allow convolutional neural networks to be easier, receive faster
It holds back.Therefore the present invention using the method be it is effective and rational, i.e.,：I=Ibase+Idetail, wherein IbaseIndicate band rain figure picture pair
The lower frequency reference layer answered, IdetailIndicate band rain figure as corresponding high frequency detail layer.
(2) object function is constructed
Object function is constructed according to 2 norms between the high frequency detail layer of rain figure and the high frequency detail layer of clean figure, this
Outside, in order to reduce over-fitting, we are added to L2 regularization terms in object function.
Wherein, N is the number of the image block after the picture breakdown with rain, and n is thumbnail, and W is network parameter,
IdetailAnd JdetailThe high frequency detail layer with the clear image corresponding with its of rain figure picture, fw () function stand network are indicated respectively
Body function,Regularization is punished for L_2, and λ is coefficient of balance.
(3) network architecture of the removing rain based on single image based on depth CNN is designed
As shown in Fig. 2, the network architecture of design is made of convolutional layer, warp lamination and jump connection etc..Using 4 convolution
Layer serves as feature extractor, retains the main details part of input picture and eliminates rain line, wherein the Feature Mapping of 4 convolutional layers
Number is all 128, is denoted as the 1st convolutional layer, the 2nd convolutional layer, the 3rd convolutional layer, the 4th convolutional layer, the size difference of convolution kernel respectively
It is set as 9*9,3*3,3*3,3*3, and uses Tanh activation primitives after each convolutional layer.
Since convolution operation focuses on the details of original image in smaller szie so that the detail section of original image can
Lost in capable of having, the resolution ratio of original image decreases, therefore 4 warp laminations are added after convolutional layer, is denoted as respectively
1st warp lamination, the 2nd warp lamination, the 3rd warp lamination, the 4th warp lamination, wherein the Feature Mapping number of 4 warp laminations
Respectively 128,128,128,3, the size of convolution kernel is respectively set to 3*3,3*3,3*3,1*1, and in the 1st, 2,3 deconvolution
Tanh activation primitives are used after layer.In view of the characteristic pattern generated by convolutional layer includes many image details, by these characteristic patterns
Their restoring image details can be helped by being integrated into uncoiling lamination, therefore the jump being added between 3 convolutional layers and warp lamination
Jump connection.Wherein respectively by the 1st convolutional layer and the 3rd warp lamination composition jump connection, by the 2nd convolutional layer and the 2nd
Warp lamination composition jump connection, by the 3rd convolutional layer and the 1st warp lamination composition jump connection.The introducing of jump connection
Contribute to gradient back-propagation to bottom, to make network more stablize in the training stage.The detail parameters of the network architecture are such as
Shown in table 1.
Table 1 removes rain network architecture Rain-removal Net (R2N detail parameters)
(4) Training strategy
The present invention realizes network using Tensorflow frames.On NVIDIA GTX Taitan-xp GPU, training
Network needs 2-3 hour convergence.The batch size and learning rate that network is trained every time are respectively set to 10 and 0.002, training
Iterations are 10.Compared with other image rain removing methods based on deep learning, network needs the less time to be restrained,
And the time spent is less.Speculate this is because：(a) smaller convolution kernel size and less Feature Mapping make network parameter
It reduces and calculation amount reduces.(b) less training sample.
The present invention uses Tanh functions as the activation primitive of network, and the synthesis rain figure data set created using forefathers
As training data, band rain/clean image of 200,000 64*64 sizes is randomly choosed to as training data, these training numbers
It is trained according to being inputted in network in batches.Using the gradient of stochastic gradient descent algorithm (SGD) undated parameter, for changing every time
For s, the parameter update of network is as follows：
Wherein, b is the bias term in network parameter, and W is network parameter, and s indicates that an iteration, α indicate that learning rate, T are
Transposition operator, IdetailAnd JdetailThe high frequency detail layer with the clear image corresponding with its of rain figure picture is indicated respectively.
(5) clear image after removal rain line is recovered
After model training, the I after removal rain line can be obtained by the output layer of networkdetailImage.Then by rain figure
Datum layer IbaseWith the I after removal rain linedetailIt is added the clean image that can be recovered.This process can be expressed as：
E=fW(Idetail)+Ibase
Wherein, IdetailIndicate the high frequency detail layer with rain figure picture, IbaseIndicate the datum layer with rain figure picture.
Fig. 3 illustrates the recovery example of 4 width anamorphic zone rain figure pictures, Fig. 3 a, 3b, 3c and 3d be respectively harbour, umbrella, rabbit and
Example of the bird with rain figure picture and recovery image.Wherein Input table shows that input picture, Ground truth indicate that input rain figure corresponds to
Clear image, DOC, DSC, GMM-LP, DrainNet are 4 kinds of state-of-the-art sides being compared during the experiment of the present invention
Method, Ours are that the obtained image of the present invention goes rain result.It is known due to synthesizing the clear image corresponding to rain figure in Fig. 3
, therefore the present invention weighs the recovery effects of synthesis rain figure using structural similarity index (SSIM).SSIM values are higher, closer
Clear image.DOR and DSC can remove part rain line and reduce the dense degree of rain line it can be seen from Fig. 3 experimental results,
But they cannot completely remove rain line.GMM-LP can remove rain line, but its result tends to excess smoothness, and cannot protect
Stay the details of original image.Compared with other methods, the method applied in the present invention can remove most of rain line and protect simultaneously
Stay the detail section of image after rain.The visual effect of DrainNet is approximate with the method applied in the present invention effect, but from
The composograph of table 2 go in the comparison of the structural similarity index after rain it can be seen that, the method applied in the present invention has reached most
High SSIM values, this demonstrate that the validity of proposed method.
Fig. 4 a, Fig. 4 b and Fig. 4 c illustrate 3 pairs and really remove rain instance graph with rain figure picture.It can be seen that GMM-LP, DSC
Rain line cannot be all completely removed with DOR.From the point of view of visual angle, in the recovery process to true rain figure, DrainNet's goes
Rain effect has a long way to go with method proposed by the invention, and method proposed by the invention can retain more image details.
In order to preferably compare, Fig. 5 a illustrate DrainNet and method proposed by the invention goes to a given zone in rain result figure
Domain, it can be seen that compared with DrainNet method proposed by the invention remain with the relevant more details of input picture and
Feature.In Fig. 5 b and Fig. 5 c, it is shown that a methodical specific region scaling figure.After rain being removed such as 2 composograph of table
Structural similarity index comparison.
2 composograph of table goes the structural similarity index comparison after rain
By observing these regions, it can be seen that method proposed by the invention obtains best visual effect, is going
Details is remained while except rain line, further demonstrates the validity of proposed method.
Claims (7)
1. a kind of single image rain line minimizing technology based on depth convolutional neural networks, which is characterized in that this method include with
Lower step：
(1) use guiding filtering method that will be decomposed into lower frequency reference layer and high frequency detail layer with rain figure picture；
(2) target letter is constructed according to 2 norms between the high frequency detail layer with rain figure picture and the high frequency detail layer of clean image
Number, and L2 regularization terms are added in object function；
(3) network architecture of a removing rain based on single image based on depth convolutional neural networks is built, including 4 convolutional layers, point
It is not denoted as the 1st convolutional layer, the 2nd convolutional layer, the 3rd convolutional layer, the 4th convolutional layer, and uses network activation letter after each convolutional layer
Number, 4 warp laminations are denoted as the 1st warp lamination respectively, the 2nd warp lamination, the 3rd warp lamination, the 4th warp lamination, and preceding
Network activation function, 3 jump connections, respectively by the 1st convolutional layer and the 3rd warp lamination group are used after 3 warp laminations
It is connected at jump, by the 2nd convolutional layer and the 2nd warp lamination composition jump connection, by the 3rd convolutional layer and the 1st warp
Lamination composition jump connection；
(4) using the data set with rain figure picture as training data, it is input to the single width figure based on depth convolutional neural networks
It is trained iteration in the network architecture as removing rain, and is directed to each iteration, the net is updated using stochastic gradient descent algorithm
Network parameter；
(5) after training iteration, the datum layer with rain figure picture is added with the high frequency detail layer after removal rain line and reverts to clean figure
Picture.
2. the single image rain line minimizing technology according to claim 1 based on depth convolutional neural networks, feature exist
In in step (2), the object function is expressed as：
Wherein, N is the number of the image block after the picture breakdown with rain, and n is thumbnail, and W is network parameter, IdetailWith
JdetailIndicate the high frequency detail layer with the clear image corresponding with its of rain figure picture respectively, fw () function stand network body function,Regularization is punished for L_2, and λ is coefficient of balance.
3. the single image rain line minimizing technology according to claim 1 based on depth convolutional neural networks, feature exist
In in step (3), the network activation function uses Tanh activation primitives.
4. the single image rain line minimizing technology according to claim 1 based on depth convolutional neural networks, feature exist
In in step (3), the 1st convolutional layer, the Feature Mapping number of the 2nd convolutional layer, the 3rd convolutional layer, the 4th convolutional layer is
128, convolution kernel size is respectively set to 9*9,3*3,3*3,3*3.
5. the single image rain line minimizing technology according to claim 1 based on depth convolutional neural networks, feature exist
In, in step (3), the 1st warp lamination, the 2nd warp lamination, the 3rd warp lamination, the 4th warp lamination Feature Mapping number
Respectively 128,128,128,3, the size of convolution kernel is respectively set to 3*3,3*3,3*3,1*1.
6. the single image rain line minimizing technology according to claim 1 based on depth convolutional neural networks, feature exist
In the network parameter more new formula is：
Wherein, b is the bias term in network parameter, and W is network parameter, and s indicates that an iteration, α indicate learning rate, and T is transposition
Operator, IdetailAnd JdetailThe high frequency detail layer with the clear image corresponding with its of rain figure picture is indicated respectively.
7. the single image rain line minimizing technology according to claim 1 based on depth convolutional neural networks, feature exist
In in step (5), described be added the datum layer with rain figure picture with the high frequency detail layer after removal rain line reverts to clean figure
Picture is formulaically expressed as：
E=fW(Idetail)+Ibase
Wherein, IdetailIndicate the high frequency detail layer with rain figure picture, IbaseIndicate the datum layer with rain figure picture.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810538160.XA CN108765344A (en) | 2018-05-30 | 2018-05-30 | A method of the single image rain line removal based on depth convolutional neural networks |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810538160.XA CN108765344A (en) | 2018-05-30 | 2018-05-30 | A method of the single image rain line removal based on depth convolutional neural networks |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108765344A true CN108765344A (en) | 2018-11-06 |
Family
ID=64004275
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810538160.XA Pending CN108765344A (en) | 2018-05-30 | 2018-05-30 | A method of the single image rain line removal based on depth convolutional neural networks |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108765344A (en) |
Cited By (21)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109636747A (en) * | 2018-12-04 | 2019-04-16 | 上海理工大学 | Depth high frequency network correcting algorithm applied to single width Non Uniformity Correction of Infrared Image |
| CN109685863A (en) * | 2018-12-11 | 2019-04-26 | 帝工(杭州)科技产业有限公司 | A method of rebuilding medicine breast image |
| CN110163813A (en) * | 2019-04-16 | 2019-08-23 | 中国科学院深圳先进技术研究院 | A kind of image rain removing method, device, readable storage medium storing program for executing and terminal device |
| 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 |
| CN110455747A (en) * | 2019-07-19 | 2019-11-15 | 浙江师范大学 | It is a kind of based on deep learning without halo effect white light phase imaging method and system |
| CN110517199A (en) * | 2019-08-26 | 2019-11-29 | 电子科技大学 | A kind of image rain removing method driven convenient for intelligent vehicle |
| CN110675330A (en) * | 2019-08-12 | 2020-01-10 | 广东石油化工学院 | Image rain removing method of encoding-decoding network based on channel level attention mechanism |
| CN110728640A (en) * | 2019-10-12 | 2020-01-24 | 合肥工业大学 | Double-channel single-image fine rain removing method |
| CN110751612A (en) * | 2019-11-05 | 2020-02-04 | 哈尔滨理工大学 | Single image rain removing method of multi-channel multi-scale convolution neural network |
| CN110866879A (en) * | 2019-11-13 | 2020-03-06 | 江西师范大学 | Image rain removing method based on multi-density rain print perception |
| CN111062892A (en) * | 2019-12-26 | 2020-04-24 | 华南理工大学 | Single image rain removing method based on composite residual error network and deep supervision |
| CN111127354A (en) * | 2019-12-17 | 2020-05-08 | 武汉大学 | Single-image rain removing method based on multi-scale dictionary learning |
| CN111462014A (en) * | 2020-04-03 | 2020-07-28 | 西安交通大学 | Single-image rain removing method based on deep learning and model driving |
| CN111861925A (en) * | 2020-07-24 | 2020-10-30 | 南京信息工程大学滨江学院 | Image rain removing method based on attention mechanism and gate control circulation unit |
| CN112070701A (en) * | 2020-09-08 | 2020-12-11 | 北京字节跳动网络技术有限公司 | Image generation method, device, equipment and computer readable medium |
| CN112087556A (en) * | 2019-06-12 | 2020-12-15 | 武汉Tcl集团工业研究院有限公司 | Dark light imaging method and device, readable storage medium and terminal equipment |
| CN112215789A (en) * | 2020-10-12 | 2021-01-12 | 北京字节跳动网络技术有限公司 | Image defogging method, device, equipment and computer readable medium |
| CN112259075A (en) * | 2020-10-10 | 2021-01-22 | 腾讯科技（深圳）有限公司 | Voice signal processing method, device, electronic equipment and storage medium |
| CN112488943A (en) * | 2020-12-02 | 2021-03-12 | 北京字跳网络技术有限公司 | Model training and image defogging method, device and equipment |
| CN113379641A (en) * | 2021-06-25 | 2021-09-10 | 南昌航空大学 | Single image rain removing method and system based on self-coding convolutional neural network |
| CN114677306A (en) * | 2022-03-29 | 2022-06-28 | 中国矿业大学 | Context aggregation image rain removing method based on edge information guidance |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107133935A (en) * | 2017-05-25 | 2017-09-05 | 华南农业大学 | A kind of fine rain removing method of single image based on depth convolutional neural networks |
| CN107909556A (en) * | 2017-11-27 | 2018-04-13 | 天津大学 | Video image rain removing method based on convolutional neural networks |
-
2018
- 2018-05-30 CN CN201810538160.XA patent/CN108765344A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107133935A (en) * | 2017-05-25 | 2017-09-05 | 华南农业大学 | A kind of fine rain removing method of single image based on depth convolutional neural networks |
| CN107909556A (en) * | 2017-11-27 | 2018-04-13 | 天津大学 | Video image rain removing method based on convolutional neural networks |
Non-Patent Citations (2)
| Title |
|---|
| XIAO-JIAO MAO等: "Image Denoising Using Very Deep Fully Convolutional Encoder-Decoder Networks with Symmetric Skip Connections", 《HTTPS://ARXIV.ORG/PDF/1603.09056V1.PDF》 * |
| XUEYANG FU等: "Clearing the Skies: A Deep Network Architecture for Single-Image Rain Removal", 《IEEE TRANSACTIONS ON IMAGE PROCESSING》 * |
Cited By (31)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109636747A (en) * | 2018-12-04 | 2019-04-16 | 上海理工大学 | Depth high frequency network correcting algorithm applied to single width Non Uniformity Correction of Infrared Image |
| CN109685863A (en) * | 2018-12-11 | 2019-04-26 | 帝工(杭州)科技产业有限公司 | A method of rebuilding medicine breast image |
| CN110163813A (en) * | 2019-04-16 | 2019-08-23 | 中国科学院深圳先进技术研究院 | A kind of image rain removing method, device, readable storage medium storing program for executing and terminal device |
| CN110163813B (en) * | 2019-04-16 | 2022-02-01 | 中国科学院深圳先进技术研究院 | Image rain removing method and device, readable storage medium and terminal equipment |
| CN112087556A (en) * | 2019-06-12 | 2020-12-15 | 武汉Tcl集团工业研究院有限公司 | Dark light imaging method and device, readable storage medium and terminal equipment |
| CN112087556B (en) * | 2019-06-12 | 2023-04-07 | 武汉Tcl集团工业研究院有限公司 | Dark light imaging method and device, readable storage medium and terminal equipment |
| 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 |
| CN110455747A (en) * | 2019-07-19 | 2019-11-15 | 浙江师范大学 | It is a kind of based on deep learning without halo effect white light phase imaging method and system |
| CN110455747B (en) * | 2019-07-19 | 2021-09-28 | 浙江师范大学 | Deep learning-based white light phase imaging method and system without halo effect |
| CN110675330A (en) * | 2019-08-12 | 2020-01-10 | 广东石油化工学院 | Image rain removing method of encoding-decoding network based on channel level attention mechanism |
| CN110517199A (en) * | 2019-08-26 | 2019-11-29 | 电子科技大学 | A kind of image rain removing method driven convenient for intelligent vehicle |
| CN110517199B (en) * | 2019-08-26 | 2022-03-08 | 电子科技大学 | Image rain removing method convenient for intelligent vehicle driving |
| CN110728640A (en) * | 2019-10-12 | 2020-01-24 | 合肥工业大学 | Double-channel single-image fine rain removing method |
| CN110728640B (en) * | 2019-10-12 | 2023-07-18 | 合肥工业大学 | Fine rain removing method for double-channel single image |
| CN110751612A (en) * | 2019-11-05 | 2020-02-04 | 哈尔滨理工大学 | Single image rain removing method of multi-channel multi-scale convolution neural network |
| CN110866879B (en) * | 2019-11-13 | 2022-08-05 | 江西师范大学 | Image rain removing method based on multi-density rain print perception |
| CN110866879A (en) * | 2019-11-13 | 2020-03-06 | 江西师范大学 | Image rain removing method based on multi-density rain print perception |
| CN111127354A (en) * | 2019-12-17 | 2020-05-08 | 武汉大学 | Single-image rain removing method based on multi-scale dictionary learning |
| CN111127354B (en) * | 2019-12-17 | 2022-07-26 | 武汉大学 | Single-image rain removing method based on multi-scale dictionary learning |
| CN111062892A (en) * | 2019-12-26 | 2020-04-24 | 华南理工大学 | Single image rain removing method based on composite residual error network and deep supervision |
| CN111062892B (en) * | 2019-12-26 | 2023-06-16 | 华南理工大学 | Single image rain removing method based on composite residual error network and deep supervision |
| CN111462014A (en) * | 2020-04-03 | 2020-07-28 | 西安交通大学 | Single-image rain removing method based on deep learning and model driving |
| CN111861925A (en) * | 2020-07-24 | 2020-10-30 | 南京信息工程大学滨江学院 | Image rain removing method based on attention mechanism and gate control circulation unit |
| CN111861925B (en) * | 2020-07-24 | 2023-09-29 | 南京信息工程大学滨江学院 | Image rain removing method based on attention mechanism and door control circulation unit |
| CN112070701A (en) * | 2020-09-08 | 2020-12-11 | 北京字节跳动网络技术有限公司 | Image generation method, device, equipment and computer readable medium |
| CN112259075A (en) * | 2020-10-10 | 2021-01-22 | 腾讯科技（深圳）有限公司 | Voice signal processing method, device, electronic equipment and storage medium |
| CN112215789A (en) * | 2020-10-12 | 2021-01-12 | 北京字节跳动网络技术有限公司 | Image defogging method, device, equipment and computer readable medium |
| CN112488943A (en) * | 2020-12-02 | 2021-03-12 | 北京字跳网络技术有限公司 | Model training and image defogging method, device and equipment |
| CN112488943B (en) * | 2020-12-02 | 2024-02-02 | 北京字跳网络技术有限公司 | Model training and image defogging method, device and equipment |
| CN113379641A (en) * | 2021-06-25 | 2021-09-10 | 南昌航空大学 | Single image rain removing method and system based on self-coding convolutional neural network |
| CN114677306A (en) * | 2022-03-29 | 2022-06-28 | 中国矿业大学 | Context aggregation image rain removing method based on edge information guidance |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108765344A (en) | A method of the single image rain line removal based on depth convolutional neural networks | |
| CN109389556B (en) | Multi-scale cavity convolutional neural network super-resolution reconstruction method and device | |
| Gurrola-Ramos et al. | A residual dense u-net neural network for image denoising | |
| CN108510504B (en) | Image segmentation method and device | |
| CN107767413B (en) | Image depth estimation method based on convolutional neural network | |
| CN109740652B (en) | Pathological image classification method and computer equipment | |
| Zhang et al. | Adaptive residual networks for high-quality image restoration | |
| CN110675411B (en) | Cervical squamous intraepithelial lesion recognition algorithm based on deep learning | |
| CN106845529A (en) | Image feature recognition methods based on many visual field convolutional neural networks | |
| CN110378844A (en) | Motion blur method is gone based on the multiple dimensioned Image Blind for generating confrontation network is recycled | |
| CN109949224B (en) | Deep learning-based cascade super-resolution reconstruction method and device | |
| CN111476249B (en) | Construction method of multi-scale large-receptive-field convolutional neural network | |
| CN107730536B (en) | High-speed correlation filtering object tracking method based on depth features | |
| CN108765425A (en) | Image partition method, device, computer equipment and storage medium | |
| Ram et al. | Image denoising using nl-means via smooth patch ordering | |
| CN107944459A (en) | A kind of RGB D object identification methods | |
| CN109740451A (en) | Road scene image semantic segmentation method based on importance weighting | |
| CN110503610A (en) | A kind of image sleet trace minimizing technology based on GAN network | |
| CN109840483A (en) | A kind of method and device of landslide fissure detection and identification | |
| CN109447897B (en) | Real scene image synthesis method and system | |
| CN106855996A (en) | A kind of gray scale image color method and its device based on convolutional neural networks | |
| CN112580662A (en) | Method and system for recognizing fish body direction based on image features | |
| CN109190666B (en) | Flower image classification method based on improved deep neural network | |
| CN114897728A (en) | Image enhancement method and device, terminal equipment and storage medium | |
| CN111126185B (en) | Deep learning vehicle target recognition method for road gate scene |
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 |