CN109086807A - A kind of semi-supervised light stream learning method stacking network based on empty convolution - Google Patents
A kind of semi-supervised light stream learning method stacking network based on empty convolution Download PDFInfo
- Publication number
- CN109086807A CN109086807A CN201810779483.8A CN201810779483A CN109086807A CN 109086807 A CN109086807 A CN 109086807A CN 201810779483 A CN201810779483 A CN 201810779483A CN 109086807 A CN109086807 A CN 109086807A
- Authority
- CN
- China
- Prior art keywords
- network
- light stream
- semi
- stacking
- supervised
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
- G06F18/2155—Generating training patterns; Bootstrap methods, e.g. bagging or boosting characterised by the incorporation of unlabelled data, e.g. multiple instance learning [MIL], semi-supervised techniques using expectation-maximisation [EM] or naïve labelling
-
- 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
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
Abstract
The semi-supervised learning optical flow approach based on convolutional neural networks that the present invention provides a kind of, belongs to network design field.Method provided by the invention can be trained for tape label and the blended data without label, and it designs one kind and blocks perception loss function, the end point error cost function for being used for supervised learning is combined with the data item and smooth item for being used for unsupervised learning, construct a kind of semi-supervised learning optical flow estimation, using stacking network structure in the network architecture, empty convolution is introduced in convolutional layer to increase receptive field, and design and block sensing layer to estimate occlusion area, which can carry out semi-supervised light stream study end-to-endly.Method provided by the invention can be improved light stream estimated accuracy, and also propose that one kind blocks the perception semi-supervised trained network of loss function, and a kind of stacking network structure is designed in the network architecture to further promote network performance.
Description
Technical field
The present invention provides a kind of light stream estimation methods, and in particular to the semi-supervised light stream of network is stacked based on empty convolution
Learning method belongs to network design field.
Background technique
Light stream estimation can be used as a supervised learning problem, and the supervised learning method based on convolutional neural networks is solving
Preferable effect is achieved on light stream estimation problem, but there are still many problems for supervised learning optical flow approach.Generation true first
The data true value on boundary is difficult to obtain, and lacks the limitation for largely becoming supervised learning optical flow approach with labeled data, is secondly to avoid
Motion information is lost, and is operated in existing many full convolutional network frameworks without pondization, however convolution operation can still lose
The detailed information of image, these problems remain fatal for Pixel-level task.Meanwhile occlusion issue is also that light stream is estimated
Urgent problem in meter.
In view of the above problems, the invention proposes a kind of semi-supervised learning optical flow estimation based on convolutional neural networks,
Empty convolution is introduced in network convolutional layer to increase receptive field, and is designed one layer and blocked sensing layer for occlusion area involvement network instruction
Practice process to improve light stream estimated accuracy, and then propose that one kind blocks the perception semi-supervised trained network of loss function, finally exists
A kind of stacking network structure is designed in the network architecture further promotes network performance.
Summary of the invention
The present invention provides a kind of semi-supervised light stream learning methods that network is stacked based on empty convolution, it is therefore intended that is directed to
Tape label and blended data without label are trained, and are designed one kind and blocked perception loss function, and supervised learning will be used for
End point error cost function is combined with the data item and smooth item for being used for unsupervised learning, constructs a kind of semi-supervised learning light stream
Model is named as SA-Net, and a kind of stacking network structure is used in the network architecture, introduces empty convolution in convolutional layer to increase
The receptive field of convolution kernel, and design one layer and block sensing layer to estimate occlusion area, which being capable of semi-supervised end-to-end
Practise light stream.
The object of the present invention is achieved like this:
Step 1: the 1st light stream of building learns sub-network, and it is named as SA-Net_1, SA-Net_1 light stream learning network
It using full convolution framework, is constituted by shrinking and expanding 2 parts, 4 layers of standard volume are respectively adopted to 2 width images first in constriction
Characteristic pattern is extracted in product operation, 2 width characteristic patterns input relevant layers is carried out characteristic matching and merging later, then pass through 4 layers of cavity
Convolutional layer extracts Optical-flow Feature, and expansion includes 4 layers of warp lamination, and the light stream that constriction is extracted is reverted to original image
Resolution ratio.
Step 2: the 2nd light stream of building learns sub-network, and it is named as SA-Net_2, SA-Net_2 light stream learning network
It using full convolution framework, is formed by shrinking and expanding 2 parts, input layer will input network after 2 width image stacks, network passes through 4
Layer Standard convolution layer and 4 layers of empty convolutional layer extract the light stream between image pair, and constriction is made of 4 layers of warp lamination, will
The light stream that constriction is extracted reverts to original image resolution.
Step 3: 2 stackings networks of building, after SA-Net_1 sub-network connect 2 SA-Net_2 sub-networks thus
The 1st stacking network is constituted, each sub-network junction uses deformation technology by the 2nd width image to the 1st width anamorphose, later
Using deformed image and the 1st width image as the input of next sub-network, the light stream increment of 2 width images is calculated.2nd heap
Folded network and the 1st stacking network share network architecture and parameter, the input terminal input t moment and t+1 of the 1st stacking network
The 2 width images at moment extract the positive light stream between image pair, while the image at t and t+1 moment being exchanged and is sequentially input to the
In 2 stacking networks, the reversed light stream between image pair is extracted.
Step 4: 2 stackings networks of training, wherein only need trained 1st stacking network, the 2nd network share its more
Network weight after new, when synchronizing the sub-network of 2 stacking network corresponding positions of training, every 1 layer of difference of expansion is defeated
Each layer of positive light stream and reversed light stream are inputted simultaneously and block perception by the positive light stream and reversed light stream of different resolution out
Layer, goes out occlusion area by consistency desired result criteria function, stops positive and negative one after positive light stream restores to original resolution
The verification of cause property.
Step 5: designing one kind blocks perception loss function, it can be with semi-supervised learning network.It will be used for the end of supervised learning
Point tolerance cost function is combined with the data item and smooth item for being used for unsupervised learning, can both be trained to tape label data,
Can also be to the training of the data of no label, wherein data item is to be decomposed based on picture structure texture and Census convert to design
It is constant it is assumed that smooth item designs smooth item using the isotropic diffusion based on image-driven, loss function can be by anti-
To propagation trained network semi-supervised end-to-end.
Step 6: the training stage, first in network input input largely without label data, by loss weight summation
Total losses is obtained, later using back-propagation algorithm training network, initial network weight is obtained, later with a small amount of tape label number
According to training network, network model to the end is obtained.
Step 7: tested using trained model, inputs as image pair, export as corresponding light stream.
Compared with prior art, present invention has an advantage that
Method provided by the invention introduces empty convolution in network convolutional layer to increase receptive field, and designs one layer and block
Occlusion area is incorporated network training process to improve light stream estimated accuracy by sensing layer, and then proposes that one kind blocks perception loss letter
Number carrys out semi-supervised trained network, and a kind of stacking network structure is also designed in the network architecture to further promote internetworking
Energy.
Detailed description of the invention
Fig. 1 is flow chart of the present invention.
Fig. 2 is study sub-network SA-Net_1 architecture diagram of the invention.
Fig. 3 is study sub-network SA-Net_2 architecture diagram of the invention.
Fig. 4 is stacking learning network architecture diagram of the invention.
Fig. 5 is empty convolution schematic diagram of the invention.
Fig. 6 is occlusion area error schematic diagram of the invention.
Fig. 7 is semi-supervised learning light stream network SA-Net general frame figure of the invention.
Specific embodiment
The present invention is more fully described with reference to the accompanying drawing.
Step 1: as shown in Fig. 2, 1 light stream of building learns sub-network SA-Net_1, first in constriction by 4 layers
Standard convolution layer extracts the characteristic pattern of t moment Yu t+1 moment image respectively, helps net mate feature by a relevant layers
Figure, finds the corresponding relationship between characteristic pattern, the correlation function of relevant layers is defined as follows:
WhereinThe characteristic pattern of t moment and t+1 moment is respectively indicated, π is indicated centered on pixel x, size K*
The agglomerate of K.
2 agglomerates in 2 width images centered on x1 and x2 are taken respectively, and then corresponding position, which is multiplied, to be added, relevant layers pair
Entire image carries out relevant operation, while the feature of 2 width images being merged, and extracts higher by 4 layers of empty convolutional layer later
Secondary feature, the visible attached drawing 5 of empty convolution schematic diagram, Fig. 5 indicate that 1 size is 3*3, is divided into 1 empty convolution kernel, due to
Pond layer is not included in network, if increasing receptive field using larger-size convolution kernel in deep layer network, calculation amount will
Can be very huge, using the convolution kernel in such as Fig. 5, receptive field can be increased in the case where not increasing network parameter.It is got the bid
The convolution kernel size of quasi- convolutional layer be 3*3, step-length 2, empty convolution kernel size be 3*3, step-length 1, interval numerical value from it is small to
Exponentially form increases greatly, respectively 2,4,8,16, and then ReLU layers after every layer of convolutional layer and empty convolutional layer, other ginsengs
Fig. 2 visible with details is arranged in number.The expansion of network is made of 4 layers of warp lamination, convolution kernel size be 3*3, step-length 2,
ReLU layers are connected after every layer of deconvolution, are operated by a series of deconvolution, characteristic pattern reverts to original image resolution size, obtains
Light stream to the end.
Step 2: as shown in figure 3, the 2nd light stream of building learns sub-network SA-Net_2, by the figure of t moment and t+1 moment
As inputting network in a stacked fashion, constriction first is made of 4 layers of Standard convolution layer and 4 layers of empty convolutional layer, and network is logical
It crosses Standard convolution layer and empty convolutional layer extracts Optic flow information, expansion is made of 4 layers of warp lamination, light stream is reverted to
Original image resolution.Wherein the size of the convolution kernel of Standard convolution layer is 3*3, and the convolution kernel of step-length 2, empty convolutional layer is big
Small is 3*3, and step-length 1, interval is 5*5 by exponential form growth, respectively 2,4,8,12, the convolution kernel size of warp lamination,
Step-length is 2, all connected nonlinearity ReLU layers after each layer of convolution.Since SA-Net_1 and SA-Net_2 sub-network does not include entirely
The input of articulamentum, two networks can be the image of arbitrary size.
Step 3: the identical stacking learning network of 2 frameworks is trained simultaneously, for learning two images in a manner of semi-supervised
Between positive light stream and reversed light stream, each stacking network is by 1 SA-Net_1 sub-network and 2 SA-Net_2 sub-network heaps
It is folded to form, in order to which the result to a upper sub-network is assessed, while making the easier calculating incremental update of whole network, this
In deformation operation is added between each stacking sub-network, with the output light stream of a sub-network to t+1 moment image into
Row deformation operation, obtained image can be used following formula to indicate:
Wherein It+1,It respectively indicates with deformed image before deforming, u, v respectively indicate pixel x, the light stream value at y.
It will be deformedThe input of t moment image and their luminance errors as next sub-network, under
One sub- e-learningIncrement light stream between t moment image.Here deformation operation passes through guidable bilinear interpolation
Algorithm is realized, to guarantee that stacking network can train end-to-end.In the training stage, it is only necessary to the 1st stacking network of training,
2nd stacking network share its weight, to stacking the Training strategy that network uses for SA-Net_1 sub-network trained first, after being
Two sub-networks provide good initial value, then keep the weight of SA-Net_1 constant, the next stacking sub-network SA- of training
Net_2, next the weight of 2 sub-networks of fixed front is constant, and the 3rd stacking sub-network of training updates weight.SA-Net
The number of iterations when increasing the depth of network by way of stacking network, while increasing network training, to improve network
Overall performance.
Step 4: as shown in fig. 6, one pixel is by positive light stream mapping and reversed light stream in de-occlusion region
Original location of pixels will be returned to after mapping, and in occlusion area, pixel is by positive light stream mapping and backlight
Pixel position and original pixel point position after stream mapping can have deviation, and occlusion area is the concentration zones that such error occurs
Domain, therefore the larger screening of light stream evaluated error can be obtained by carrying out positive and negative consistency desired result to positive light stream and backlight stream
Keep off region.Positive and negative consistency desired result discriminant function is as follows:
WhereinWithThe positive light stream and reversed light stream at pixel x are respectively indicated, ε is the threshold of discriminant function
Value.
Labeling function O is blocked in definitionx, when discriminant function be greater than threshold value when indicate this region light stream solve there are larger mistakes
Difference, it will be determined as occlusion area, this season Ox=1.When discriminant function be less than threshold value when indicate this region light stream solve compared with
It is accurate, it will be determined as de-occlusion region, enable Ox=0.When training 2 are stacked with every 1 layer positive light of network expansion part
Stream and backlight stream carry out consistency desired result, estimate occlusion area to be used for network training process.Occlusion area participates in network instruction
Practice process, is conducive to improve light stream precision.
Step 5: design constraint occlusion area pixel blocks perception loss function, it is adapted to semi-supervised learning light stream net
Network, loss function are only applied to the training stage, train light stream learning network semi-supervisedly by backpropagation.Half prison of the invention
Educational inspector practises optical flow estimation compared with supervised learning optical flow estimation, is no longer influenced by light stream true value and obtains difficult limitation, can both have
The study light stream of supervision, study light stream that can also be unsupervised are more suitable for solving the problems, such as the extraction of motion information of real world.
Loss function ElossIt is as follows:
Eloss=α Eepe+(1-α)(Edata+γEsmooth), (4)
Wherein EepeFor end point error cost function, EdataFor data item constraint cost function, EsmoothMotion smoothing constraint,
α and γ is weight, and when input data, which is, label data, α value is 1, when input data is no label data, α value
It is 0.
End point error cost function EepeIt is as follows:
Wherein m and n is respectively the width and height of input picture, ui,jAnd vi,jIt Wei not predict obtained light stream value, ui′,j
And vi′,jFor corresponding light stream true value.
Data item cost function EdataIt is as follows:
Wherein κxFor the light stream at pixel x, N is pixel number, and T (x) indicates the value of amount of texture at pixel x, C
(x) value of the Census transformation at pixel x is indicated, φ is robust penalty (x2+δ2)α, δ=0.001, OxMark is blocked in expression
Remember function.
Picture breakdown is the part I comprising geological information by picture structure texture algorithmS(x) and include image texture information
Part IT(x), that is, there is following formula:
I (x)=IS(x)+IT(x), (7)
The wherein texture part I of imageT(x) it is hardly illuminated by the light, the influence of shade isocandela variation.
Census transformation is a kind of nonlinear transformation, has conservation property in the case where illumination is acutely monotonically changed, will be in image
Pixel in a certain rectangular transform window is indicated with a string of binary sequences, carries out being applied to number after simply improving to Census transformation
According in item constraint, specific implementation formula is as follows:
Wherein W (p) is indicated using p as the rectangular transform window of center pixel, and q is other points in rectangular window, I (p), I (q)
Gray value respectively at p, q pixel,For character string connector, σ is the threshold value of discriminate.
Motion smoothing cost function is as follows:
WhereinWithThe respectively gradient value of light stream horizontal direction and vertical direction.
Step 6: the input terminal in network inputs a small amount of tape label data and largely without label data, passes through different damages
It loses weight to sum to obtain total losses, and semi-supervised learning network is trained using back-propagation algorithm.
Step 7: the input tape label data and without label data in trained model, to semi-supervised learning light stream net
Network is tested, and is exported as corresponding dense optical flow.
Claims (6)
1. a kind of semi-supervised light stream learning method for stacking network based on empty convolution, it is characterised in that:
Step 1: the 1st light stream of building learns sub-network, and it is named as SA-Net_1, SA-Net_1 light stream learning network uses
Full convolution framework, is constituted by shrinking and expanding 2 parts, and 4 layers of Standard convolution behaviour are respectively adopted to 2 width images first in constriction
Make extraction characteristic pattern, 2 width characteristic patterns input relevant layers is subjected to characteristic matching and merging later, then pass through 4 layers of empty convolution
Layer extracts Optical-flow Feature, and expansion includes 4 layers of warp lamination, and the light stream that constriction is extracted is reverted to original image and is differentiated
Rate;
Step 2: the 2nd light stream of building learns sub-network, and it is named as SA-Net_2, SA-Net_2 light stream learning network uses
Full convolution framework, is formed by shrinking and expanding 2 parts, and input layer will input network after 2 width image stacks, and network passes through 4 layers of mark
Quasi- convolutional layer and 4 layers of empty convolutional layer extract the light stream between image pair, and constriction is made of 4 layers of warp lamination, will shrink
The light stream of extracting section reverts to original image resolution;
Step 3: 2 stacking networks of building, connect 2 SA-Net_2 sub-networks after SA-Net_1 sub-network to constitute
1st stacking network, each sub-network junction use deformation technology by the 2nd width image to the 1st width anamorphose, will become later
The input of image and the 1st width image as next sub-network after shape calculates the light stream increment of 2 width images;2nd stacking net
The input terminal of network and the 1st stacking network share network architecture and parameter, the 1st stacking network inputs t moment and t+1 moment
2 width images extract the positive light stream between image pair, while the image of t and t+1 moment being exchanged and are sequentially input to the 2nd heap
In folded network, the reversed light stream between image pair is extracted;
Step 4: 2 stackings networks of training, wherein only need the 1st stacking network of training, the 2nd network share its update after
Network weight, synchronize training 2 stacking network corresponding position sub-network when, every 1 layer of expansion exports not respectively
With the positive light streams and reversed light stream of resolution ratio, each layer of positive light stream and reversed light stream are inputted simultaneously and block sensing layer,
Go out occlusion area by consistency desired result criteria function, stops positive converse consistency1 after positive light stream restores to original resolution
Verification;
Step 5: designing one kind blocks perception loss function, the endpoint for being used for supervised learning can be missed with semi-supervised learning network
Poor cost function is combined with the data item and smooth item for being used for unsupervised learning, both can be to the training of tape label data, can also
With the data training to no label, wherein data item is constant to design based on the decomposition of picture structure texture and Census transformation
It is assumed that using smooth item is designed based on the isotropic diffusion of image-driven, loss function passes through backpropagation end-to-end
Semi-supervised trained network;
Step 6: the training stage, first in network input input largely without label data, by summing to obtain to loss weight
Total losses obtains initial network weight, is instructed later with a small amount of tape label data later using back-propagation algorithm training network
Practice network, obtains network model to the end;
Step 7: tested using trained model, inputs as image pair, export as corresponding light stream.
2. a kind of semi-supervised light stream learning method for stacking network based on empty convolution according to claim 1, feature
Be: the process described in step 1 for extracting characteristic pattern is as follows:
The characteristic pattern for extracting t moment Yu t+1 moment image respectively by 4 layers of Standard convolution layer is helped by a relevant layers
Net mate characteristic pattern finds the corresponding relationship between characteristic pattern, and the correlation function of relevant layers is defined as follows:
WhereinThe characteristic pattern of t moment and t+1 moment is respectively indicated, π is indicated centered on pixel x, size is K*K's
Agglomerate.
3. a kind of semi-supervised light stream learning method for stacking network based on empty convolution according to claim 1, feature
Be: deformation operation process described in step 3 is as follows: with the output light stream of a sub-network to t+1 moment image into
Row deformation operation, obtained image can be used following formula to indicate:
Wherein It+1,It respectively indicates with deformed image before deforming, u, v respectively indicate pixel x, the light stream value at y.
4. a kind of semi-supervised light stream learning method for stacking network based on empty convolution according to claim 1, feature
Be: consistency desired result criteria function occlusion area process described in step 4 is specific as follows:
Positive and negative consistency desired result function formula is as follows
WhereinWithThe positive light stream and reversed light stream at pixel x are respectively indicated, ε is the threshold value of discriminant function;
Labeling function O is blocked in definitionx, indicate that the light stream in this region is solved there are large error when discriminant function is greater than threshold value, it will
It can be determined as occlusion area, this season Ox=1;It is more smart to indicate that the light stream in this region is solved when discriminant function is less than threshold value
It is quasi-, it will to be determined as de-occlusion region, enable Ox=0;Every 1 layer positive light stream stacking network expansion part to 2 when training and
Backlight stream carries out consistency desired result, estimates occlusion area to be used for network training process.
5. a kind of semi-supervised light stream learning method for stacking network based on empty convolution according to claim 1, feature
It is: loss function E described in step 5lossIt is as follows:
Eloss=α Eepe+(1-α)(Edata+γEsmooth), (4)
Wherein EepeFor end point error cost function, EdataFor data item constraint cost function, EsmoothMotion smoothing constraint, α and γ
For weight, when input data, which is, label data, α value is 1, and when input data is no label data, α value is 0.
6. a kind of semi-supervised light stream learning method for stacking network based on empty convolution according to claim 5, feature
It is: the end point error cost function EepeIt is as follows:
Wherein m and n is respectively the width and height of input picture, ui,jAnd vi,jIt Wei not predict obtained light stream value, u 'i,jWith
v′i,jFor corresponding light stream true value;
The data item cost function EdataIt is as follows:
Wherein κxFor the light stream at pixel x, N is pixel number, and T (x) indicates the value of amount of texture at pixel x, and C (x) is indicated
The value of Census transformation at pixel x, φ are robust penalty (x2+δ2)α, δ=0.001, OxLabeling function is blocked in expression;
The motion smoothing cost function is as follows:
WhereinWithThe respectively gradient value of light stream horizontal direction and vertical direction.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810779483.8A CN109086807B (en) | 2018-07-16 | 2018-07-16 | Semi-supervised optical flow learning method based on void convolution stacking network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810779483.8A CN109086807B (en) | 2018-07-16 | 2018-07-16 | Semi-supervised optical flow learning method based on void convolution stacking network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109086807A true CN109086807A (en) | 2018-12-25 |
CN109086807B CN109086807B (en) | 2022-03-18 |
Family
ID=64838001
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810779483.8A Active CN109086807B (en) | 2018-07-16 | 2018-07-16 | Semi-supervised optical flow learning method based on void convolution stacking network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109086807B (en) |
Cited By (17)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109741383A (en) * | 2018-12-26 | 2019-05-10 | 西安电子科技大学 | Picture depth estimating system and method based on empty convolution sum semi-supervised learning |
CN110533051A (en) * | 2019-08-02 | 2019-12-03 | 中国民航大学 | Contraband automatic testing method in X-ray safety check image based on convolutional neural networks |
CN110599428A (en) * | 2019-09-24 | 2019-12-20 | 北京航空航天大学青岛研究院 | Optical flow estimation heterogeneous hybrid network and embedding method thereof |
CN111325774A (en) * | 2020-02-14 | 2020-06-23 | 上海交通大学 | Optical flow unsupervised loss calculation method based on geometric relation |
CN111462191A (en) * | 2020-04-23 | 2020-07-28 | 武汉大学 | Non-local filter unsupervised optical flow estimation method based on deep learning |
WO2020156009A1 (en) * | 2019-01-31 | 2020-08-06 | 北京市商汤科技开发有限公司 | Video repair method and device, electronic device and storage medium |
CN111507185A (en) * | 2020-03-11 | 2020-08-07 | 杭州电子科技大学 | Tumble detection method based on stack cavity convolution network |
CN111524166A (en) * | 2020-04-22 | 2020-08-11 | 北京百度网讯科技有限公司 | Video frame processing method and device |
CN111582483A (en) * | 2020-05-14 | 2020-08-25 | 哈尔滨工程大学 | Unsupervised learning optical flow estimation method based on space and channel combined attention mechanism |
CN112465872A (en) * | 2020-12-10 | 2021-03-09 | 南昌航空大学 | Image sequence optical flow estimation method based on learnable occlusion mask and secondary deformation optimization |
CN112634331A (en) * | 2020-12-04 | 2021-04-09 | 北京迈格威科技有限公司 | Optical flow prediction method and device |
CN112785523A (en) * | 2021-01-22 | 2021-05-11 | 北京大学 | Semi-supervised image rain removing method and device for sub-band network bridging |
CN112950639A (en) * | 2020-12-31 | 2021-06-11 | 山西三友和智慧信息技术股份有限公司 | MRI medical image segmentation method based on SA-Net |
CN113344066A (en) * | 2021-05-31 | 2021-09-03 | 中国工商银行股份有限公司 | Model training method, service distribution method, device and equipment |
CN114119445A (en) * | 2020-08-27 | 2022-03-01 | 北京晟易机器人科技有限公司 | Pad voidage calculation method based on automatic X-ray imaging |
CN114667544A (en) * | 2020-08-14 | 2022-06-24 | 腾讯美国有限责任公司 | Multi-rate neural image compression method and device with stackable nested model structure |
CN115759202A (en) * | 2022-11-22 | 2023-03-07 | 江苏济远医疗科技有限公司 | Dense image super-resolution network with variable cavity convolution |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107679477A (en) * | 2017-09-27 | 2018-02-09 | 深圳市未来媒体技术研究院 | Face depth and surface normal Forecasting Methodology based on empty convolutional neural networks |
CN107808389A (en) * | 2017-10-24 | 2018-03-16 | 上海交通大学 | Unsupervised methods of video segmentation based on deep learning |
US9953236B1 (en) * | 2017-03-10 | 2018-04-24 | TuSimple | System and method for semantic segmentation using dense upsampling convolution (DUC) |
CN107967515A (en) * | 2016-10-19 | 2018-04-27 | 三星电子株式会社 | The method and apparatus quantified for neutral net |
CN108256562A (en) * | 2018-01-09 | 2018-07-06 | 深圳大学 | Well-marked target detection method and system based on Weakly supervised space-time cascade neural network |
-
2018
- 2018-07-16 CN CN201810779483.8A patent/CN109086807B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107967515A (en) * | 2016-10-19 | 2018-04-27 | 三星电子株式会社 | The method and apparatus quantified for neutral net |
US9953236B1 (en) * | 2017-03-10 | 2018-04-24 | TuSimple | System and method for semantic segmentation using dense upsampling convolution (DUC) |
CN107679477A (en) * | 2017-09-27 | 2018-02-09 | 深圳市未来媒体技术研究院 | Face depth and surface normal Forecasting Methodology based on empty convolutional neural networks |
CN107808389A (en) * | 2017-10-24 | 2018-03-16 | 上海交通大学 | Unsupervised methods of video segmentation based on deep learning |
CN108256562A (en) * | 2018-01-09 | 2018-07-06 | 深圳大学 | Well-marked target detection method and system based on Weakly supervised space-time cascade neural network |
Non-Patent Citations (3)
Title |
---|
XUEZHI XIANG 等: "Scene Flow Estimation Based on Adaptive Anisotropic Total Variation Flow-Driven Method", 《MATHEMATICAL PROBLEMS IN ENGINEERING》 * |
YI ZHU 等: "Learning Optical Flow via Dilated Networks and Occlusion Reasoning", 《ARXIV:1805.02733V1》 * |
刘勇: "基于光流场分析与深度学习的视频监控系统", 《湘南学院学报》 * |
Cited By (28)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109741383A (en) * | 2018-12-26 | 2019-05-10 | 西安电子科技大学 | Picture depth estimating system and method based on empty convolution sum semi-supervised learning |
US11544820B2 (en) | 2019-01-31 | 2023-01-03 | Beijing Sensetime Technology Development Co., Ltd. | Video repair method and apparatus, and storage medium |
WO2020156009A1 (en) * | 2019-01-31 | 2020-08-06 | 北京市商汤科技开发有限公司 | Video repair method and device, electronic device and storage medium |
CN110533051A (en) * | 2019-08-02 | 2019-12-03 | 中国民航大学 | Contraband automatic testing method in X-ray safety check image based on convolutional neural networks |
CN110533051B (en) * | 2019-08-02 | 2023-01-17 | 中国民航大学 | Automatic detection method for contraband in X-ray security inspection image based on convolutional neural network |
CN110599428A (en) * | 2019-09-24 | 2019-12-20 | 北京航空航天大学青岛研究院 | Optical flow estimation heterogeneous hybrid network and embedding method thereof |
CN110599428B (en) * | 2019-09-24 | 2023-06-20 | 北京航空航天大学青岛研究院 | Heterogeneous hybrid network for optical flow estimation and embedding method thereof |
CN111325774A (en) * | 2020-02-14 | 2020-06-23 | 上海交通大学 | Optical flow unsupervised loss calculation method based on geometric relation |
CN111325774B (en) * | 2020-02-14 | 2023-04-18 | 上海交通大学 | Optical flow unsupervised loss calculation method based on geometric relation |
CN111507185A (en) * | 2020-03-11 | 2020-08-07 | 杭州电子科技大学 | Tumble detection method based on stack cavity convolution network |
CN111507185B (en) * | 2020-03-11 | 2020-11-24 | 杭州电子科技大学 | Tumble detection method based on stack cavity convolution network |
US11748895B2 (en) | 2020-04-22 | 2023-09-05 | Beijing Baidu Netcom Science And Technology Co., Ltd. | Method and apparatus for processing video frame |
CN111524166A (en) * | 2020-04-22 | 2020-08-11 | 北京百度网讯科技有限公司 | Video frame processing method and device |
CN111462191A (en) * | 2020-04-23 | 2020-07-28 | 武汉大学 | Non-local filter unsupervised optical flow estimation method based on deep learning |
CN111462191B (en) * | 2020-04-23 | 2022-07-19 | 武汉大学 | Non-local filter unsupervised optical flow estimation method based on deep learning |
CN111582483A (en) * | 2020-05-14 | 2020-08-25 | 哈尔滨工程大学 | Unsupervised learning optical flow estimation method based on space and channel combined attention mechanism |
CN114667544A (en) * | 2020-08-14 | 2022-06-24 | 腾讯美国有限责任公司 | Multi-rate neural image compression method and device with stackable nested model structure |
CN114119445A (en) * | 2020-08-27 | 2022-03-01 | 北京晟易机器人科技有限公司 | Pad voidage calculation method based on automatic X-ray imaging |
CN112634331A (en) * | 2020-12-04 | 2021-04-09 | 北京迈格威科技有限公司 | Optical flow prediction method and device |
CN112465872B (en) * | 2020-12-10 | 2022-08-26 | 南昌航空大学 | Image sequence optical flow estimation method based on learnable occlusion mask and secondary deformation optimization |
CN112465872A (en) * | 2020-12-10 | 2021-03-09 | 南昌航空大学 | Image sequence optical flow estimation method based on learnable occlusion mask and secondary deformation optimization |
CN112950639B (en) * | 2020-12-31 | 2024-05-10 | 山西三友和智慧信息技术股份有限公司 | SA-Net-based MRI medical image segmentation method |
CN112950639A (en) * | 2020-12-31 | 2021-06-11 | 山西三友和智慧信息技术股份有限公司 | MRI medical image segmentation method based on SA-Net |
CN112785523B (en) * | 2021-01-22 | 2023-10-17 | 北京大学 | Semi-supervised image rain removing method and device for sub-band network bridging |
CN112785523A (en) * | 2021-01-22 | 2021-05-11 | 北京大学 | Semi-supervised image rain removing method and device for sub-band network bridging |
CN113344066A (en) * | 2021-05-31 | 2021-09-03 | 中国工商银行股份有限公司 | Model training method, service distribution method, device and equipment |
CN115759202B (en) * | 2022-11-22 | 2023-11-28 | 江苏济远医疗科技有限公司 | Variable cavity convolution dense image super-resolution method |
CN115759202A (en) * | 2022-11-22 | 2023-03-07 | 江苏济远医疗科技有限公司 | Dense image super-resolution network with variable cavity convolution |
Also Published As
Publication number | Publication date |
---|---|
CN109086807B (en) | 2022-03-18 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109086807A (en) | A kind of semi-supervised light stream learning method stacking network based on empty convolution | |
CN106157319B (en) | The conspicuousness detection method in region and Pixel-level fusion based on convolutional neural networks | |
CN115601549A (en) | River and lake remote sensing image segmentation method based on deformable convolution and self-attention model | |
CN110223324A (en) | A kind of method for tracking target of the twin matching network indicated based on robust features | |
CN109271933A (en) | The method for carrying out 3 D human body Attitude estimation based on video flowing | |
CN110322495A (en) | A kind of scene text dividing method based on Weakly supervised deep learning | |
CN103020909B (en) | Single-image super-resolution method based on multi-scale structural self-similarity and compressive sensing | |
CN107292247A (en) | A kind of Human bodys' response method and device based on residual error network | |
CN102982544B (en) | Many foreground object image interactive segmentation method | |
CN110674742B (en) | Remote sensing image road extraction method based on DLinkNet | |
CN110399518A (en) | A kind of vision question and answer Enhancement Method based on picture scroll product | |
CN103578107B (en) | A kind of interactive image segmentation method | |
Hua et al. | Depth estimation with convolutional conditional random field network | |
CN116486489B (en) | Three-dimensional hand object posture estimation method and system based on semantic perception graph convolution | |
Bae et al. | Densely distilled flow-based knowledge transfer in teacher-student framework for image classification | |
CN110705340A (en) | Crowd counting method based on attention neural network field | |
CN116206133A (en) | RGB-D significance target detection method | |
CN114092824A (en) | Remote sensing image road segmentation method combining intensive attention and parallel up-sampling | |
Peng et al. | A proxy model to predict reservoir dynamic pressure profile of fracture network based on deep convolutional generative adversarial networks (DCGAN) | |
CN113313176A (en) | Point cloud analysis method based on dynamic graph convolution neural network | |
CN114120148B (en) | Method for detecting changing area of remote sensing image building | |
CN116822382A (en) | Sea surface temperature prediction method and network based on space-time multiple characteristic diagram convolution | |
Ren et al. | Co-saliency detection using collaborative feature extraction and high-to-low feature integration | |
CN117557856A (en) | Pathological full-slice feature learning method based on self-supervision learning | |
CN104537694A (en) | Online learning offline video tracking method based on key frames |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |