CN110363794A - Light stream prediction technique between video successive frame - Google Patents
Light stream prediction technique between video successive frame Download PDFInfo
- Publication number
- CN110363794A CN110363794A CN201910645583.6A CN201910645583A CN110363794A CN 110363794 A CN110363794 A CN 110363794A CN 201910645583 A CN201910645583 A CN 201910645583A CN 110363794 A CN110363794 A CN 110363794A
- Authority
- CN
- China
- Prior art keywords
- network
- light stream
- feature
- layer
- convolution
- 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 28
- 238000012549 training Methods 0.000 claims abstract description 25
- 230000004927 fusion Effects 0.000 claims abstract description 20
- 239000013598 vector Substances 0.000 claims description 13
- 238000013507 mapping Methods 0.000 claims description 12
- 230000004913 activation Effects 0.000 claims description 7
- 230000015572 biosynthetic process Effects 0.000 claims description 6
- 238000003786 synthesis reaction Methods 0.000 claims description 6
- 238000005070 sampling Methods 0.000 claims description 5
- 230000003321 amplification Effects 0.000 claims description 4
- 238000003199 nucleic acid amplification method Methods 0.000 claims description 4
- 235000013399 edible fruits Nutrition 0.000 claims 1
- 238000004364 calculation method Methods 0.000 abstract description 3
- 230000009286 beneficial effect Effects 0.000 abstract description 2
- 238000010586 diagram Methods 0.000 description 5
- 230000006870 function Effects 0.000 description 5
- 238000013527 convolutional neural network Methods 0.000 description 3
- 239000000463 material Substances 0.000 description 3
- 230000008569 process Effects 0.000 description 3
- 230000003044 adaptive effect Effects 0.000 description 2
- 238000006073 displacement reaction Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 230000005284 excitation Effects 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 230000002301 combined effect Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000007499 fusion processing Methods 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000000452 restraining effect Effects 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
- 239000002699 waste material Substances 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/269—Analysis of motion using gradient-based 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
- G06T2207/10—Image acquisition modality
- G06T2207/10016—Video; Image sequence
-
- 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]
-
- 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
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30241—Trajectory
Landscapes
- Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses the light stream prediction techniques between a kind of video successive frame, are related to computer machine vision technique field, comprising the following steps: step a: by can deformation convolution unit extract consecutive frame space characteristics;Step b: fusion reconstruct is carried out to consecutive frame space characteristics;Step c: deconvolution operation is carried out to the feature of fusion reconstruct, constructs network stack;Step d: loss function training network stack is utilized;Step e: output result.The beneficial effects of the present invention are: convolution kernel is optimized from basic structure type, by fixed rectangular convolution be changed to can deformation convolution, improve precision of prediction and saved calculation resources;By training by the feature reconstruction of fusion and allocation of parameters and channel weight again, thus it is a small amount of increase calculating cost in the case where retain the feature correlation of consecutive frame to greatest extent.
Description
Technical field
Light stream prediction side the present invention relates to computer machine vision technique field, between especially a kind of video successive frame
Method.
Background technique
Light stream prediction can be applied to trajectory track, and the fields such as preceding background segment and human behavior identification are computer views
Feel the common method and major issue in research.Light stream includes space motion object pixel motion on observation imaging plane
The information such as instantaneous velocity and displacement vector, light stream image indicate image object movement by way of color gamut and spatial domain combine
State.Relative to other image studies methods, the emphasis of light stream is essentially consisted in " movement ", and light stream not only includes observed object
Motion information also includes the abundant information in relation to scenery three-dimensional structure.Light stream prediction referred to using pixel in image sequence in the time
The correlation between variation and consecutive frame on domain finds previous frame with corresponding relationship existing between present frame, to count
The method for calculating the motion information (light stream figure) of object between consecutive frame.
The previous algorithm that light stream prediction is carried out using convolutional neural networks, in convolution characteristic extraction part usually using fixed
The rectangular convolution kernel of size shape;This convolution form limits network in the ability adjusted to different images body self-adaptation,
The universality being displaced to pixel size between consecutive frame is poor, also will cause the waste of computing resource.
The key of light stream prediction algorithm is to calculate the sports ground of adjacent interframe, therefore the association fusion of adjacent two frames feature
It is the most important thing solved the problems, such as, and existing method is directly superimposed or is based on window mobile computing correlation in precision and time
There are obvious shortcoming, part edge details easy to be lost in fusion process in complexity, precision is affected.
Summary of the invention
The purpose of the present invention is to solve the light stream forecasting problems between video image successive frame, retain to greatest extent empty
Between structure and the feature correlation information of consecutive frame, promote precision of prediction, devise the light stream between a kind of video successive frame
Prediction technique.
To achieve the goals above, the technical scheme is that, the light stream prediction technique between a kind of video successive frame,
The following steps are included: step a: by can deformation convolution unit extract consecutive frame space characteristics;Step b: to consecutive frame space spy
Sign carries out fusion reconstruct;Step c: deconvolution operation is carried out to the feature of fusion reconstruct, constructs network stack;Step d: damage is utilized
Lose function training network stack;Step e: output.
Further, in the step a, it is described can deformation convolution unit first layer be 7*7 can deformation convolutional layer,
The second layer is traditional convolutional layer of 5*5, and traditional convolutional layer that third layer and the 4th layer are 3*3, the step-length of each convolutional layer is 2;
It is described can convolution operation used in deformation convolutional layer are as follows:
In formula, P0For can the exported Feature Mapping u of deformation convolutional layer a bit, the input feature vector that wherein x represents this layer reflects
It penetrates or original image, R is covered on the region of u by convolution kernel, w is weighted value, PnIt is R enumerating in x institute overlay area.
Further, in the step b, the consecutive frame space characteristics by the overall situation by channel be averaged pondization convert
The scalar of port number is characterized for a length;It includes full connection, ReLU activation primitive, full connection that the scalar, which is sent into one,
With the full articulamentum of Sigmoid activation primitive, the weight vectors as fusion feature are obtained by subsequent training operation;
Using the weight vectors as the selection parameter in feature channel, then by multiplication by channel weighting to consecutive frame space characteristics
On, complete the recalibration to primitive character on channel dimension.
Further, in the step c, by deconvolution operation by the feature amplification of the fusion reconstruct and close to light
Then resolution ratio is restored to original size by up-sampling by stream.
Further, in the step c, the building network stack includes network one, network two and network three, described
Network one using can deformation convolution sum SE-net, network two simultaneously removes SE-net using traditional convolution kernel, and the network three uses
Traditional convolution kernel adds SE-net, and the network one is connected in parallel with network two and being connected in series through for network three.
Further, the network two transmits prediction light stream and loss amount to network three.
Further, in the step d, the ground truch and output light stream mean end-point mistake in data set are used
Difference is used as loss function, calculates loss after output light stream is carried out primary interpolation.
Further, the strategy of the trained network is first the first hierarchical network of stand-alone training, then again by the guidance
The fixed training undernet of parameters within network weight, until all-ones subnet network training is completed, in the last layer addition synthesis mould
Block finely tunes synthesis module under fixed upper layer network inner parameter.
Further, the step e includes output result light stream image corresponding with output video consecutive frame.
The beneficial effects of the present invention are: convolution kernel is optimized from basic structure type, by fixed rectangular convolution
Be changed to can deformation convolution, improve precision of prediction and saved calculation resources;By training the feature reconstruction of fusion and again
Allocation of parameters and channel weight, so that the feature for retaining consecutive frame to greatest extent in the case where a small amount of increase calculates cost is related
Property.
Detailed description of the invention
Fig. 1 is application scheme flow chart;
Fig. 2 is traditional convolutional layer operation chart;
Fig. 3 is can deformation convolution kernel schematic diagram;
Fig. 4 be can deformation convolution schematic diagram;
Fig. 5 be added offset can deformation pond schematic diagram;
Fig. 6 is the operation principle schematic diagram of SE-net
Fig. 7 is complete sub-network structure figure;
Fig. 8 is complete network stack diagram.
Specific embodiment
It is of the invention to reach the technical means and efficacy that predetermined goal of the invention is taken further to illustrate, below in conjunction with
Attached drawing and preferred embodiment, to specific embodiment, structure, feature and its effect according to the present invention, detailed description are as follows:
The application is for the light stream forecasting problem solved between video image successive frame.Firstly, the application use can deformation
The self-adaptive features of adjacent two frame of convolution feature extracting method, to greatest extent retaining space structural information.Secondly, by two parts sky
Between feature superposition be input to SE-net feature reconstruction module and calculate correlation and restructuring allocation channel weight, obtain fusion feature;
Fusion feature is subjected to up-sampling again and deconvolution operates, is recycled after obtaining rough light stream image by network stack and warp
The method of optimization is corrected, and parameters within network weight learns to adjust by sample training.
A kind of light stream prediction technique between video successive frame, the process with reference to shown in Fig. 1, comprising the following steps:
Step a: by can deformation convolution unit extract consecutive frame space characteristics.
Core of the convolution operation as convolutional neural networks, be usually seen as on local receptor field, will spatially
The condensate that information in information and characteristic dimension is polymerize, the concrete operations of traditional convolutional layer are as shown in Fig. 2, traditional convolution
Layer one image of input or characteristic pattern, convolution kernel is mobile in input and is calculated with overlay area and is mapped to next layer
Grade, this calculating is subchannel.Traditional square convolution kernel operating principle can be formulated are as follows:
In formula, P0For can the exported Feature Mapping u of deformation convolutional layer a bit, the input feature vector that wherein x represents this layer reflects
It penetrates or original image, R is covered on the region of u by convolution kernel, w is weighted value, PnIt is R enumerating in x institute overlay area.
This layer output has just been obtained after multiple channels as above operate.
In the application, can deformation convolution unit first layer be 7*7 can deformation convolutional layer, the second layer be 5*5 tradition
Convolutional layer, third layer and the 4th layer of traditional convolutional layer for 3*3, the step-length of each convolutional layer are 2;
It can convolution operation used in deformation convolutional layer are as follows:
In formula, P0For can the exported Feature Mapping u of deformation convolutional layer a bit, the input feature vector that wherein x represents this layer reflects
It penetrates or original image, R is covered on the region of u by convolution kernel, w is weighted value, PnIt is R enumerating in x institute overlay area,
△PnFor offset.
The previous convolution in light stream prediction field is square convolution kernel, and the form of this square convolution kernel limits net
Network is in the space that image adaptive optimizes, and the universality being displaced to pixel size between frame is poor, and calculated performance is low.In Fig. 3, figure
It (a) is traditional convolution kernel, figure (b), figure (c) and figure (d) are can deformation convolution kernel.The application improves convolution kernel,
Be added have adaptivity can deformation convolution, first layer using can deformation convolution kernel, be with the difference of traditional convolution kernel
It both increases the variable △ Pn of an offset in the position of each sampled point.By these variables, convolution kernel can work as
Front position nearby arbitrarily samples, and the regular lattice point before being no longer limited to.In fact, increased inclined in deformable convolutional layer
Shifting amount is a part of network structure, is calculated by another parallel Standard convolution layer, and then can also pass through ladder
Degree backpropagation is learnt end to end.
In addition after the study of the offset, as shown in figure 4, the size and location of deformable convolution kernel can be according to current
The picture material for needing to identify carries out dynamic adjustment, and visual effect is exactly the convolution kernel sampling point position of different location can basis
Adaptive variation occurs for picture material, to adapt to the geometric deformations such as the shape of different objects, size.
Step b: fusion reconstruct is carried out to consecutive frame space characteristics.
In this step, consecutive frame space characteristics are converted into a length and are characterized by the pond that is averaged of the overall situation by channel
The scalar of port number.Convolutional neural networks for light stream prediction would generally establish two relatively independent nets for consecutive frame image
Then the two streams are combined carry out subsequent processing in some stage by network stream.The movement and three of light stream expression image object
Tie up information, therefore find both relevance and reservation be the key that the step.The SE-net that the application uses is to pass through foundation
Relation of interdependence between fusion feature mapping channel, and feature recalibration (reconstruct) is carried out as next step input.
With reference to Fig. 6, if giving an input X, feature port number is C ', and wide and high respectively W ', H ' pass through a series of convolution transforms
It is C, the Feature Mapping U of wide and a height of W, H that a feature port number is obtained after Ftr.Pass through three operations realizations pair after this
The reconstruct of Feature Mapping, the parameters weighting of each operation are all obtained simultaneously finally obtaining the important of C feature channel by learning training
Degree, and promote important feature on this basis and inhibit secondary feature.
Squeeze operation, i.e., Feature Mapping is compressed on Spatial Dimension, and each two-dimensional feature channel is become one
A real number, the feature port number phase of dimension and input that this real number has global receptive field in a way, and exports
Matching, essence of this step operation are that an overall situation is averaged pond.The two dimensional character uc that channel is c in U adds up on entire area to be taken
Average, final each channel obtain a scalar zc, the one-dimensional variable z that C combination of channels is C at a length, expression spy
The global numeric distribution situation that sign mapping U is responded in channel C, and make the layer close to input that can also obtain the overall situation
Receptive field, concrete operations can be formulated as:
It include the full connection of full connection, ReLU activation primitive, full connection and Sigmoid activation primitive by scalar feeding one
Layer obtains the weight vectors as fusion feature by subsequent training operation.It is operated using Excitation, it is one
Similar to the mechanism of door in Recognition with Recurrent Neural Network.Weight is generated for each feature channel by parameter W, wherein parameter W is learned
Commonly use the correlation for carrying out explicitly Modelling feature interchannel.The mark between C 0 to 1 is obtained by FC-ReLU-FC-Sigmoid
Amount, as each channel weight, then each channel of original output channel is weighted (corresponding channel with corresponding weight
Each element is multiplied respectively with weight), the feature after obtaining new weighting.Following formula illustrates the step operation, first uses W1 here
Multiplied by z, i.e., full articulamentum FC (Fully Connected Layer) operation, the dimension of W1 is C/r*C, this r is a contracting
Parameter is put, this programme takes 16, and the purpose of this parameter is to reduce by channel number to reduce calculation amount.Again because of z's
Dimension is 1/C, so W1z result dimension is 1/C/r;Then using a ReLU (Rectified Linear Unit) line
Property rectification layer, it is still 1/C/r that it is constant, which to export dimension, here training when it is adjustable;Be multiplied again with W2 be multiplied with W2 be also one
The process of a full articulamentum, the dimension of W2 are C*C/r, therefore output dimension here is exactly 1/C;Finally using sigmoid
Activation primitive obtains the weight vectors s that length is C.
S=Fex(z, W)=σ (g (z, W))=σ (W2δ(W1z))
Using weight vectors as the selection parameter in feature channel, then by multiplication by channel weighting to consecutive frame space spy
In sign, the recalibration to primitive character on channel dimension is completed.It is the operation of a Reweight, by Excitation
Output weight regard as into cross feature selecting after each feature channel importance, then by multiplication by channel weighting
Onto previous feature, the recalibration to primitive character on channel dimension is completed.
In formula, UcFor c-th of channel characteristics of U, the weight that the c that Sc is weight vectors S is tieed up,For output.
Step c: deconvolution operation is carried out to the feature of fusion reconstruct, constructs network stack;
It is then by up-sampling that resolution ratio is extensive by deconvolution operation by the feature amplification of fusion reconstruct and close to light stream
Original size is arrived again.Deconvolution operation is the inverse operation of convolution operation, and major function is amplification characteristic mapping, improves image resolution
Rate keeps picture material abundant.The output F (rough light stream) of upper one layer of deconvolution is added in deconvolution step as reference, leads to
This mode is crossed, the high-level information that thicker characteristic pattern is transmitted both has been remained, the essence provided in low-level image feature figure is also provided
Thin local message.As shown in fig. 7, having carried out four deconvolution operation in this step, the light stream figure finally obtained is by above adopting
Sample (bilinear interpolation) is restored to the clarity of original image rank, and to here, whole processes of a sub-network terminate.
In order to optimize final result, network performance is promoted, this patent is in the subsequent part that joined network stack and warp.
Building network stack includes network one, network two and network three, is three kinds of structures son different with internal module
Network.Network one using can deformation convolution sum SE-net, network two simultaneously removes SE-net using traditional convolution kernel, and network three uses
Traditional convolution kernel adds SE-net, and network one is connected in parallel with network two and being connected in series through for network three.Experiment shows using can shape
Becoming network when convolution does not add SE-net module can not restrain, therefore not do the network settings of the combination herein.Experimental result card
It is bright optimal according to the structure progress sub-network storehouse combined effect of Fig. 8, show that prediction light stream is smooth.
Preferably, network two transmits prediction light stream and loss amount to network three, and undernet is made to be absorbed in study video phase
The lost motion of adjacent frame, loss are operated to obtain, be shown below by warp:
I′2=(I1,F)
C=| | I2-I′2||
In formula, I1、I2Respectively video consecutive frame, F are that sub-network exports light stream, I '2For first frame and output light stream mapping
Image (being similar to the second frame) out, C is loss amount.It can effectively prevent the network over-fitting of heap poststack by Warp operation
And optimize precision simultaneously.
Step d: loss function training network stack is utilized
The mode of learning of the application network is supervised learning, uses the ground truch and output light levelling in data set
Equal end point error constantly adjusts network parameter as loss function, calculates loss after output light stream is carried out primary interpolation.Cause
Data set and strategy used by this is trained will largely influence network performance, although and network stack can have
Effect improves precision of prediction, but has following disadvantage: 1, the complicated network structure is huge, and training speed is slow and is easy to appear quasi-
The case where closing or not restraining;2, the sub-network downlink shared information in the more tributaries of multi-layer, does not only result in the transmitting of error, can also draw
Play the problem of costing bio disturbance confusion;3, heap stack network needs a large amount of calculating cost, is easy to cause in the lesser equipment of memory
Insufficient space.Therefore herein using the strategy of substep training.Specific network training strategy is: first the first level of stand-alone training net
Then guidance parameters within network weight is fixed training undernet again by network, until all-ones subnet network training is completed, last
One layer of addition synthesis module finely tunes synthesis module under fixed upper layer network inner parameter.Wherein, original sample used by training
It should work as and have ground truth, and suitable sample data is chosen according to model application.And it should contain in sample
It blocks, obscure and the case where big displacement, so that e-learning is suitable for handling these situations.
Step e: output.
Step e includes output result light stream image corresponding with output video consecutive frame.
Above with reference to preferred embodiment, invention has been described, but protection scope of the present invention is not restricted to
This can carry out various improvement to it and can be replaced wherein with equivalent without departing from the scope of the invention
Component, as long as be not present structural conflict, it is mentioned in the various embodiments items technical characteristic can combine in any way
Get up, and any reference signs in the claims should not be construed as limiting the involved claims, no matter comes from which point
It sees, the present embodiments are to be considered as illustrative and not restrictive.Therefore, any to fall within the scope of the appended claims
All technical solutions be within the scope of the invention.
Claims (9)
1. the light stream prediction technique between a kind of video successive frame, which comprises the following steps:
Step a: by can deformation convolution unit extract consecutive frame space characteristics;
Step b: fusion reconstruct is carried out to consecutive frame space characteristics;
Step c: deconvolution operation is carried out to the feature of fusion reconstruct, constructs network stack;
Step d: loss function training network stack is utilized;
Step e: output.
2. the light stream prediction technique between video successive frame according to claim 1, which is characterized in that in the step a
In, it is described can deformation convolution unit first layer be 7*7 can deformation convolutional layer, the second layer be 5*5 traditional convolutional layer, third
Layer and the 4th layer of traditional convolutional layer for 3*3, the step-length of each convolutional layer are 2;It is described can convolution operation used in deformation convolutional layer
Are as follows:
In formula, P0For can the exported Feature Mapping u of deformation convolutional layer a bit, wherein x represent this layer input feature vector mapping or
Original image, R are covered on the region of u by convolution kernel, and w is weighted value, PnIt is R enumerating in x institute overlay area.
3. the light stream prediction technique between video successive frame according to claim 1, which is characterized in that in the step b
In, the consecutive frame space characteristics are converted into the mark that a length is characterized port number by the pond that is averaged of the overall situation by channel
Amount;It include the full connection of full connection, ReLU activation primitive, full connection and Sigmoid activation primitive by scalar feeding one
Layer obtains the weight vectors as fusion feature by subsequent training operation;Using the weight vectors as feature channel
Selection parameter, then by multiplication by channel weighting to consecutive frame space characteristics, complete on channel dimension to original
The recalibration of feature.
4. the light stream prediction technique between video successive frame according to claim 1, which is characterized in that in the step c
In, it is then by up-sampling that resolution ratio is extensive by deconvolution operation by the feature amplification of the fusion reconstruct and close to light stream
Original size is arrived again.
5. the light stream prediction technique between video successive frame according to claim 1, which is characterized in that in the step c
In, the building network stack includes network one, network two and network three, and the network one uses can deformation convolution sum SE-
Net, network two is using traditional convolution kernel and removes SE-net, and the network three adds SE-net, the net using traditional convolution kernel
Network one is connected in parallel with network two and being connected in series through for network three.
6. the light stream prediction technique between video successive frame according to claim 5, which is characterized in that the network two to
The transmitting of network three prediction light stream and loss amount.
7. the light stream prediction technique between video successive frame according to claim 1, which is characterized in that in the step d
In, use the groundtruch in data set, as loss function, output light stream to be carried out with the equal end point error of output light levelling
Loss is calculated after interpolation.
8. the light stream prediction technique between video successive frame according to claim 1, which is characterized in that the trained network
Strategy be first the first hierarchical network of stand-alone training, then again by the fixed training junior's net of the guidance parameters within network weight
Network is added synthesis module in the last layer, finely tunes under fixed upper layer network inner parameter until all-ones subnet network training completion
Synthesis module.
9. the light stream prediction technique between a kind of video successive frame according to claim 1, the step e includes output knot
Fruit light stream image corresponding with output video consecutive frame.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910645583.6A CN110363794A (en) | 2019-07-17 | 2019-07-17 | Light stream prediction technique between video successive frame |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910645583.6A CN110363794A (en) | 2019-07-17 | 2019-07-17 | Light stream prediction technique between video successive frame |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110363794A true CN110363794A (en) | 2019-10-22 |
Family
ID=68219906
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910645583.6A Pending CN110363794A (en) | 2019-07-17 | 2019-07-17 | Light stream prediction technique between video successive frame |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110363794A (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110830808A (en) * | 2019-11-29 | 2020-02-21 | 合肥图鸭信息科技有限公司 | Video frame reconstruction method and device and terminal equipment |
CN110944212A (en) * | 2019-11-29 | 2020-03-31 | 合肥图鸭信息科技有限公司 | Video frame reconstruction method and device and terminal equipment |
CN111683256A (en) * | 2020-08-11 | 2020-09-18 | 蔻斯科技(上海)有限公司 | Video frame prediction method, video frame prediction device, computer equipment and storage medium |
CN112085717A (en) * | 2020-09-04 | 2020-12-15 | 厦门大学 | Video prediction method and system for laparoscopic surgery |
CN113838102A (en) * | 2021-09-26 | 2021-12-24 | 南昌航空大学 | Optical flow determination method and system based on anisotropic dense convolution |
CN114511485A (en) * | 2022-01-29 | 2022-05-17 | 电子科技大学 | Compressed video quality enhancement method based on cyclic deformable fusion |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108710826A (en) * | 2018-04-13 | 2018-10-26 | 燕山大学 | A kind of traffic sign deep learning mode identification method |
CN109784150A (en) * | 2018-12-06 | 2019-05-21 | 东南大学 | Video driving behavior recognition methods based on multitask space-time convolutional neural networks |
-
2019
- 2019-07-17 CN CN201910645583.6A patent/CN110363794A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108710826A (en) * | 2018-04-13 | 2018-10-26 | 燕山大学 | A kind of traffic sign deep learning mode identification method |
CN109784150A (en) * | 2018-12-06 | 2019-05-21 | 东南大学 | Video driving behavior recognition methods based on multitask space-time convolutional neural networks |
Non-Patent Citations (3)
Title |
---|
ALEXEY DOSOVITSKIY: "FlowNet: Learning Optical Flow with Convolutional Networks", 《IEEE》 * |
EDDY ILG: "FlowNet 2.0: Evolution of Optical Flow Estimation with Deep Networks", 《IEEE》 * |
JIFENG DAI: "Deformable Convolutional Networks", 《ARXIV:1703.06211V3》 * |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110830808A (en) * | 2019-11-29 | 2020-02-21 | 合肥图鸭信息科技有限公司 | Video frame reconstruction method and device and terminal equipment |
CN110944212A (en) * | 2019-11-29 | 2020-03-31 | 合肥图鸭信息科技有限公司 | Video frame reconstruction method and device and terminal equipment |
CN111683256A (en) * | 2020-08-11 | 2020-09-18 | 蔻斯科技(上海)有限公司 | Video frame prediction method, video frame prediction device, computer equipment and storage medium |
CN112085717A (en) * | 2020-09-04 | 2020-12-15 | 厦门大学 | Video prediction method and system for laparoscopic surgery |
CN112085717B (en) * | 2020-09-04 | 2024-03-19 | 厦门大学 | Video prediction method and system for laparoscopic surgery |
CN113838102A (en) * | 2021-09-26 | 2021-12-24 | 南昌航空大学 | Optical flow determination method and system based on anisotropic dense convolution |
CN113838102B (en) * | 2021-09-26 | 2023-06-06 | 南昌航空大学 | Optical flow determining method and system based on anisotropic dense convolution |
CN114511485A (en) * | 2022-01-29 | 2022-05-17 | 电子科技大学 | Compressed video quality enhancement method based on cyclic deformable fusion |
CN114511485B (en) * | 2022-01-29 | 2023-05-26 | 电子科技大学 | Compressed video quality enhancement method adopting cyclic deformable fusion |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110363794A (en) | Light stream prediction technique between video successive frame | |
CN113874883A (en) | Hand pose estimation | |
CN110188795A (en) | Image classification method, data processing method and device | |
CN106250931A (en) | A kind of high-definition picture scene classification method based on random convolutional neural networks | |
CN105069752B (en) | The optics ocean clutter cancellation method of chaos during based on sky | |
CN112489164B (en) | Image coloring method based on improved depth separable convolutional neural network | |
CN104599290B (en) | Video sensing node-oriented target detection method | |
CN110570363A (en) | Image defogging method based on Cycle-GAN with pyramid pooling and multi-scale discriminator | |
CN110717532A (en) | Real-time detection method for robot target grabbing area based on SE-RetinaGrasp model | |
TWI226193B (en) | Image segmentation method, image segmentation apparatus, image processing method, and image processing apparatus | |
CN107680116A (en) | A kind of method for monitoring moving object in video sequences | |
CN110705344A (en) | Crowd counting model based on deep learning and implementation method thereof | |
CN110222760A (en) | A kind of fast image processing method based on winograd algorithm | |
CN112288776B (en) | Target tracking method based on multi-time step pyramid codec | |
CN113449691A (en) | Human shape recognition system and method based on non-local attention mechanism | |
CN110909591A (en) | Self-adaptive non-maximum value inhibition processing method for pedestrian image detection by using coding vector | |
CN115019302A (en) | Improved YOLOX target detection model construction method and application thereof | |
CN113688765A (en) | Attention mechanism-based action recognition method for adaptive graph convolution network | |
CN111583340A (en) | Method for reducing monocular camera pose estimation error rate based on convolutional neural network | |
CN104408697A (en) | Image super-resolution reconstruction method based on genetic algorithm and regular prior model | |
CN116258757A (en) | Monocular image depth estimation method based on multi-scale cross attention | |
CN116740121A (en) | Straw image segmentation method based on special neural network and image preprocessing | |
CN115546654A (en) | Grouping mixed attention-based remote sensing scene image classification method | |
CN117809200A (en) | Multi-scale remote sensing image target detection method based on enhanced small target feature extraction | |
WO2021057091A1 (en) | Viewpoint image processing method and related device |
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: 20191022 |