CN108924385A - A kind of video stabilization method based on width study - Google Patents
A kind of video stabilization method based on width study Download PDFInfo
- Publication number
- CN108924385A CN108924385A CN201810682319.5A CN201810682319A CN108924385A CN 108924385 A CN108924385 A CN 108924385A CN 201810682319 A CN201810682319 A CN 201810682319A CN 108924385 A CN108924385 A CN 108924385A
- Authority
- CN
- China
- Prior art keywords
- video
- frame
- output
- test set
- stabilization method
- 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
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N5/00—Details of television systems
- H04N5/14—Picture signal circuitry for video frequency region
- H04N5/21—Circuitry for suppressing or minimising disturbance, e.g. moiré or halo
Landscapes
- Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Signal Processing (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The present invention provides a kind of video stabilization method based on width study, according to the currently pending frame of original video, processed video correspond to frame, non-learning-oriented processing method output video correspondence frame previous frame, obtain the input data of training set and the input data of test set, then the successional primary features of video time are extracted using mapping function, feature enhancing is carried out to primary features followed by activation primitive, obtains Enhanced feature；By primary features and Enhanced feature simultaneous, obtain all features extracted in n-th of network, building is using video time continuity and video content fidelity as the energy function of constraint condition in training set, the weight met in above-mentioned energy function is solved by the minimum angle Return Law, and the target weight of characteristic layer and output layer is connected, finally obtain the output frame of the video stabilization of test set with all features extracted according to target weight in test set.
Description
Technical field
The present invention relates to computer vision and field of image processing more particularly to a kind of video debounces based on width study
Dynamic method.
Background technique
Video stabilization method shows as shake present in removal video, generally comprises tone shake and brightness jitter.
Video stabilization method algorithm removes existing shake between video frame by the time continuity between addition frame, exports one
The time continuity video of non-jitter.
In the prior art, for video stabilization, common implementation method is based on jitter compensation technology, it is intended to be passed through
Tone or brightness between aligned frame remove the flutter effect in video.Although this method can be reduced to a certain extent
Flutter effect present in video, still, this method must select first several frames as key frame, and from the quilt shaken
Several frames are chosen in the video of processing as key frame, whether these key frames itself have time consistency, it is difficult to guarantee；Again
Other frames are aligned, not ensuring that can by person if there are flutter effects for selected key frame itself with the key frame that there is shake
To remove the shake of processed video.In addition, another implementation method can also be excellent containing time consistency by minimizing
Change the energy function of item to maintain the time consistency between video frame, but such methods are specifically applied mainly for certain class,
Limit the generalization ability of method of video image processing.For example, such common video image processing algorithm includes：Intrinsic figure point
Solution, color classification, solid colour, white balance etc..In addition, the algorithm of the removal video jitter based on specific application is not particularly suited for
Most of other situations, limit the generalization ability of this kind of algorithms.
For the deficiency of above-mentioned existing method, a kind of novel video stabilization method how is designed, to improve or eliminate
Many defects remove shake present in processed video can to the maximum extent, be in computer vision development process
Urgent problem to be solved.
Summary of the invention
To solve deficiency present in existing video stabilization method, the present invention provides a kind of video based on width study
De-jittering method, debounce movable model learn based on width can be established according to the feature in input video and processed video thus
Remove video jitter.
According to one aspect of the present invention, a kind of video stabilization method based on width study, including following step are provided
Suddenly：
A) according to the currently pending frame I of original videon, with based on image processing method processed video frame by frame
Corresponding frame Pn, non-learning-oriented processing method output video correspondence frame previous frame On-1, obtain the input data X of training setn
And the input data F of test setn, wherein Xn=[In|Pn|On-1], Fn=[In|Pn]；
B) the input data X is extracted using mapping functionnFor realizing the successional primary features of video timeWherein, primary featuresIt is expressed as：
Wherein WeiAnd βeiIndicate the weight and deviation generated at random,For mapping function；
C) feature enhancing is carried out to the extracted primary features using activation primitive, obtains Enhanced featureWherein,
Enhanced featureIt is expressed as：
Wherein WhjAnd βhiIndicate the weight and deviation generated at random, ξjFor activation primitive,Indicate primary features all
M shared primary features in frame；
D) primary features for arriving said extractedAnd Enhanced feature Connection
It is vertical, obtain all feature A extracted in n-th of networkn；
WhereinIndicate p in all frames shared Enhanced feature of Enhanced feature；
E) it in the training set, constructs with video time continuity CtWith video content fidelity CfFor constraint condition
Energy function E, wherein energy function E is defined as expression formula：
The weights omega for meeting above-mentioned energy function E is solved by the minimum angle Return Lawn, and by weights omeganLearn as width
Network is used to the target weight of connection features layer and output layer；
F) in test set, according to target weight ωnWith all feature A extracted in n-th of networkn, obtain width
Practise the output Y of the test set of networkn：
Yn=An·ωn
Wherein, the output Y of test setnOutput frame for the video stabilization learnt based on width.
In an embodiment wherein, mapping functionFor sigmoid function or tangent function.
In an embodiment wherein, activation primitive ξjFor sigmoid function or tangent function.
In an embodiment wherein, weights omeganFor minimize the test set output frame and former frame difference from
And calculate the energy loss work factor of the time continuity between output video consecutive frame：
Ct=| | An·ωn-On-1||2。
In an embodiment wherein, weights omeganFor minimize the output video of the test set n-th video frame and
The difference between n-th of video frame in processed video is to calculate the energy loss work factor of video content fidelity:
Cf=| | An·ωn-Pn||2。
In an embodiment wherein, weights omeganIt is used to the target of connection features layer and output layer as width learning network
When weight, while meeting the constraint condition of video time continuity and video content fidelity.
In an embodiment wherein, the image processing method that processed video uses frame by frame include color classification processing,
The processing of space white balance, color harmonization processing and high dynamic range mapping processing.
Using the video stabilization method of the invention based on width study, first according to the currently pending of original video
Frame, with based on image processing method frame by frame the correspondence frame of processed video, non-learning-oriented processing method output video
The previous frame of corresponding frame, is obtained the input data of training set and the input data of test set, is then mentioned using mapping function
Take the input data of above-mentioned training set for realizing the successional primary features of video time, followed by activation primitive to first
Grade feature carries out feature enhancing, obtains Enhanced feature；Then the primary features and Enhanced feature simultaneous said extracted arrived, obtain
All features extracted into n-th of network, building is in training set with video time continuity and video content fidelity
For the energy function of constraint condition, the weight met in above-mentioned energy function is solved by the minimum angle Return Law, and as
Width learning network is used to the target weight of connection features layer and output layer, finally according to target weight and extraction in test set
All features arrived obtain the output frame of the video stabilization of the test set of width learning network.Compared with the prior art, this Shen
The output video please obtained using original input video, processed video and traditional de-jittering method as inputting, with by
Layer constantly extracts the width learning network that feature is established, and is constraint in video time continuity and video content fidelity
Under the conditions of, to obtain eliminating the output video of shake.
Detailed description of the invention
Reader is after having read a specific embodiment of the invention referring to attached drawing, it will more clearly understands of the invention
Various aspects.Wherein,
Fig. 1 shows the flow chart of the video stabilization method of the invention based on width study；
Fig. 2 shows the configuration diagrams of the width learning network of the video stabilization method for realizing Fig. 1；
Fig. 3 A shows the schematic diagram for a certain video frame that original video is Interview；
Fig. 3 B shows the schematic diagram for a certain video frame that original video is Cable；
Fig. 3 C shows the schematic diagram for a certain video frame that original video is Chicken；
Fig. 3 D shows the schematic diagram for a certain video frame that original video is CheckingEmail；
Fig. 3 E shows the schematic diagram for a certain video frame that original video is Travel；And
Fig. 4 is shown using the video stabilization method of Fig. 1 and two kinds of video stabilization methods of the prior art in original view
The comparison schematic diagram of video debounce effect when frequency is respectively Fig. 3 A~Fig. 3 E.
Specific embodiment
In order to keep techniques disclosed in this application content more detailed with it is complete, be referred to the embodiment of the present invention son in
Attached drawing, the technical solution and realization details implemented in the present invention will be further described in more detail in we.
Fig. 1 shows the flow chart of the video stabilization method of the invention based on width study, and Fig. 2 shows for realizing figure
The configuration diagram of the width learning network of 1 video stabilization method, Fig. 3 A~Fig. 3 E are shown respectively original video and are
The schematic diagram of a certain video frame of Interview, Cable, Chicken, CheckingEmail and Travel, Fig. 4, which is shown, to be adopted
It in original video is respectively Fig. 3 A~Fig. 3 E with two kinds of video stabilization methods of the video stabilization method of Fig. 1 and the prior art
When video debounce effect comparison schematic diagram.
Hardware condition of the invention is cpu frequency 2.40GHz, the computer of memory 8G, software tool Matlab
2014b。
Referring to Fig.1, in this embodiment, the video stabilization method based on width study of the application mainly passes through following
Step is realized.
Firstly, in step sl, according to the currently pending frame I of original videon, with based on image processing method frame by frame
The correspondence frame P of processed videon, non-learning-oriented processing method (that is, traditional treatment method) output video correspondence frame
Previous frame On-1, obtain the input data X of training setnAnd the input data F of test setn, wherein Xn=[In|Pn|On-1], Fn
=[In|Pn]。
It, be in view of corresponding output frame O in the test set data of training width learning networknAnd PnBetween video in
Hold fidelity and output frame OnWith its former frame On-1Between time continuity, we are first by original video, processed
Video and former output video in input X of the correspondence frame as primary features mapping functionn=[In|Pn|On-1], pass through mapping
Our obtained i-th of primary features of function WhereinIt can be arbitrary activation primitive,
It can be sigmoid or tangent function, WeiAnd βeiThe weight and deviation with suitable dimension being randomly generated respectively,
N-th for reconstructing OnNeural network in, if there is the primary mappings characteristics of m group, Wo MenlingTo indicate
The primary mappings characteristics of m group in the width learning network of n-th of video stabilization, as shown in Figure 2.
Secondly, in step s 2, to the m group primary features generated in step S1Feature enhancing is carried out, retraining obtains
Enhanced featureWherein ξj() can be arbitrary sigmoid or tangent function, WhjWith
βhiThe weight and deviation with suitable dimension being randomly generated respectively, at n-th for reconstructing OnNeural network in, if
There are p group Enhanced feature, Wo MenlingThe p in width learning network for indicating n-th of video stabilization
Group Enhanced feature, as shown in Figure 2.
M group primary features in the width learning network for obtaining n-th of video stabilizationWith p group Enhanced feature
Afterwards, Wo MenlingIndicate all features extracted in the width learning network of n-th of Key dithering.Then, I
Pass through target weight ω to be askednBy AnWith output layer OnIt connects.Solving target weight ωnWidth study afterwards
In network, the output Y of test setn=An·ωn.It may be noted that in training set, output frame OnIt is by known by traditional
What non-learning-oriented de-jittering method obtained, in the stage of training width learning network, unique unknown number is for connection to characteristic layer
With the target weight ω of output layern.In test set, output frame YnIt is unknown, utilization trained width learning network
It can solve, that is, Yn=An·ωn。
In step S31 and step S32, in the unknown weight for solving the width learning network for realizing video stabilization
ωnDuring, video time continuity and video content fidelity must be considered simultaneously.
Specifically, when considering the time continuity between video consecutive frame, we are enabled between output video consecutive frame
The energy loss cost of time continuity is Ct, wherein target weight ωnIt can be used for minimizing the output frame of test set and previous
The difference of frame, so as to calculate above-mentioned energy loss work factor：
Ct=| | An·ωn-On-1||2
Wherein, | | | |2Indicate L2Normal form (quadratic sum and then evolution of vector each element), On-1It indicates to use in training set
(n-1) frame that traditional video stabilization method obtains indicates to have solved target weight ω in test setnWidth
Practise (n-1) frame of network output.
Similarly, in order to guarantee that the content of the dynamic scene in processed video saves as much as possible in output video,
When considering video content fidelity, it would be desirable to minimize processed video and export the difference between video, and enable output
Energy loss cost between video and processed video is Cf.Wherein, target weight ωnIt can be used for minimizing test set
The difference between n-th of video frame in n-th of the video frame and processed video of video is exported, so as to calculate above-mentioned view
The energy loss work factor of frequency content fidelity：
Cf=| | An·ωn-Pn||2
Wherein, PnIndicate the n-th frame in processed video.
In step s 4, simultaneous video time continuity constraint and video content fidelity difference are constructed with video time
Continuity CtWith video content fidelity CfFor the energy function E of constraint condition, above-mentioned energy is met by the solution of the minimum angle Return Law
The weights omega of flow function En, and by weights omeganIt is used to the target weight of connection features layer and output layer as width learning network.
Energy function E is represented by：
Wherein, the first item of above-mentioned expression formula is the output frame A obtained for minimizing training setn·ωnWith use tradition
The output frame O that video stabilization method obtainsnDifference, improve width learning model accuracy, Section 2 λ1·‖ωn‖1With
Section 3 λ2·‖ωn‖2It is all the regular terms for preventing over-fitting, wherein λ1And λ2It is L respectively1Normal form and L2The canonical of normal form
Term coefficient.λtAnd λfIt is the coefficient of video time continuity and video content fidelity respectively.
To the unknown quantity weights omega in above formulan, we can be solved with the method that minimum angular convolution is returned, so that it is determined that being based on
The video stabilization model of width study.As shown in Fig. 3 A~3E, Fig. 4, using the video stabilization method and existing view of Fig. 1
When frequency de-jittering method is compared, it will therefore be readily appreciated that Interview video, Cable video, Chicken video,
On CheckingEmail video and Travel video, it is utilized respectively the video stabilization method of Lang in the prior art et al.
The video stabilization method (such as curve 3) of (such as curve 2), Bonneel in the prior art et al. and the video debounce of the application
Y-PSNR (Peak Signal to Noise Ratio, the PSNR) number for the output video that dynamic method (such as curve 1) obtains
Value, as shown in the vertical dotted line in Fig. 4.For example, when Interview video, Cable video, the Chicken view of Fig. 3 A~Fig. 3 E
Frequently, the shake in CheckingEmail video and Travel video, which is respectively derived from, uses based on figure respective original video
Color classification, space white balance, the intrinsic figure of picture decompose, high dynamic range mapping and defogging method are handled frame by frame, but not
In view of the video time consistency between consecutive frame.Since the value of PSNR can reflect the quality and Key dithering effect of output video
Fruit, therefore, PSNR value are higher, and quality and the Key dithering effect for exporting video are also better.The view of the application it can be seen from upper figure
Frequency de-jittering method (such as curve 1), based on traditional de-jittering method (such as curve 2 and curve 3), is measured compared to various in PSNR
Substandard debounce performance is intended to more excellent.
Using the video stabilization method of the invention based on width study, first according to the currently pending of original video
Frame, with based on image processing method frame by frame the correspondence frame of processed video, non-learning-oriented processing method output video
The previous frame of corresponding frame, is obtained the input data of training set and the input data of test set, is then mentioned using mapping function
Take the input data of above-mentioned training set for realizing the successional primary features of video time, followed by activation primitive to first
Grade feature carries out feature enhancing, obtains Enhanced feature；Then the primary features and Enhanced feature simultaneous said extracted arrived, obtain
All features extracted into n-th of network, building is in training set with video time continuity and video content fidelity
For the energy function of constraint condition, the weight met in above-mentioned energy function is solved by the minimum angle Return Law, and as
Width learning network is used to the target weight of connection features layer and output layer, finally according to target weight and extraction in test set
All features arrived obtain the output frame of the video stabilization of the test set of width learning network.Compared with the prior art, this Shen
The output video please obtained using original input video, processed video and traditional de-jittering method as inputting, with by
Layer constantly extracts the width learning network that feature is established, and is constraint in video time continuity and video content fidelity
Under the conditions of, to obtain eliminating the output video of shake.
Above, a specific embodiment of the invention is described with reference to the accompanying drawings.But those skilled in the art
It is understood that without departing from the spirit and scope of the present invention, a specific embodiment of the invention can also be made etc.
With replacement, without departing from essential core of the invention, these modifications and replacement should all fall in claims of the present invention and be limited
In fixed range.
Claims (7)
1. a kind of video stabilization method based on width study, which is characterized in that the video stabilization method includes following step
Suddenly：
A) according to the currently pending frame I of original videon, with the correspondence based on image processing method processed video frame by frame
Frame Pn, non-learning-oriented processing method output video correspondence frame previous frame On-1, obtain the input data X of training setnAnd
The input data F of test setn, wherein Xn=[In|Pn|On-1], Fn=[In|Pn]；
B) the input data X is extracted using mapping functionnFor realizing the successional primary features of video timeIts
In, primary featuresIt is expressed as：
Wherein WeiAnd βeiIndicate the weight and deviation generated at random,For mapping function；
C) feature enhancing is carried out to the extracted primary features using activation primitive, obtains Enhanced featureWherein, enhance
FeatureIt is expressed as：
Wherein WhiAnd βhiIndicate the weight and deviation generated at random, ξjFor activation primitive,Indicate primary features in all frames
M shared primary features；
D) primary features for arriving said extractedAnd Enhanced feature Simultaneous obtains
All feature A extracted in n-th of networkn；
WhereinIndicate p in all frames shared Enhanced feature of Enhanced feature；
E) it in the training set, constructs with video time continuity CtWith video content fidelity CfFor the energy of constraint condition
Function E, wherein energy function E is defined as expression formula：
The weights omega for meeting above-mentioned energy function E is solved by the minimum angle Return Lawn, and by weights omeganAs width learning network
For the target weight of connection features layer and output layer；
F) in test set, according to target weight ωnWith all feature A extracted in n-th of networkn, obtain width and learn net
The output Y of the test set of networkn：
Yn=An·ωn
Wherein, the output Y of test setnOutput frame for the video stabilization learnt based on width.
2. video stabilization method according to claim 1, which is characterized in that mapping functionFor sigmoid function or
Tangent function.
3. video stabilization method according to claim 1, which is characterized in that activation primitive ξjFor sigmoid function or
Tangent function.
4. video stabilization method according to claim 1, which is characterized in that weights omeganFor minimizing the test set
Output frame and former frame difference to calculate output video consecutive frame between time continuity energy loss cost because
Son
Ct=| | An·ωn-On-1||2。
5. video stabilization method according to claim 1, which is characterized in that weights omeganFor minimizing the test set
Output video n-th of video frame and processed video in n-th of video frame between difference to calculating video content
The energy loss work factor of fidelity
Cf=| | An·ωn-Pn||2。
6. video stabilization method according to claim 1, which is characterized in that weights omeganIt is used to as width learning network
When the target weight of connection features layer and output layer, while meeting the constraint item of video time continuity and video content fidelity
Part.
7. video stabilization method according to claim 1, which is characterized in that the image that processed video uses frame by frame
Processing method includes color classification processing, the processing of space white balance, color harmonization processing and high dynamic range mapping processing.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810682319.5A CN108924385B (en) | 2018-06-27 | 2018-06-27 | Video de-jittering method based on width learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810682319.5A CN108924385B (en) | 2018-06-27 | 2018-06-27 | Video de-jittering method based on width learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108924385A true CN108924385A (en) | 2018-11-30 |
| CN108924385B CN108924385B (en) | 2020-11-03 |
Family
ID=64421608
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810682319.5A Active CN108924385B (en) | 2018-06-27 | 2018-06-27 | Video de-jittering method based on width learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108924385B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109905565A (en) * | 2019-03-06 | 2019-06-18 | 南京理工大学 | Video stabilization method based on motor pattern separation |
| CN110222234A (en) * | 2019-06-14 | 2019-09-10 | 北京奇艺世纪科技有限公司 | A kind of video classification methods and device |
| CN110472741A (en) * | 2019-06-27 | 2019-11-19 | 广东工业大学 | A kind of small wave width study filtering system of three-domain fuzzy and method |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101616310A (en) * | 2009-07-17 | 2009-12-30 | 清华大学 | The target image stabilizing method of binocular vision system of variable visual angle and resolution |
| CN103929568A (en) * | 2013-01-11 | 2014-07-16 | 索尼公司 | Method For Stabilizing A First Sequence Of Digital Image Frames And Image Stabilization Unit |
| US20170278223A1 (en) * | 2016-03-22 | 2017-09-28 | Kabushiki Kaisha Toshiba | Image adjustment |
| CN107481185A (en) * | 2017-08-24 | 2017-12-15 | 深圳市唯特视科技有限公司 | A kind of style conversion method based on video image optimization |
| CN107808144A (en) * | 2017-11-10 | 2018-03-16 | 深圳市唯特视科技有限公司 | One kind carries out self-supervision insertion posture learning method based on video time-space relationship |
-
2018
- 2018-06-27 CN CN201810682319.5A patent/CN108924385B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101616310A (en) * | 2009-07-17 | 2009-12-30 | 清华大学 | The target image stabilizing method of binocular vision system of variable visual angle and resolution |
| CN103929568A (en) * | 2013-01-11 | 2014-07-16 | 索尼公司 | Method For Stabilizing A First Sequence Of Digital Image Frames And Image Stabilization Unit |
| US20170278223A1 (en) * | 2016-03-22 | 2017-09-28 | Kabushiki Kaisha Toshiba | Image adjustment |
| CN107481185A (en) * | 2017-08-24 | 2017-12-15 | 深圳市唯特视科技有限公司 | A kind of style conversion method based on video image optimization |
| CN107808144A (en) * | 2017-11-10 | 2018-03-16 | 深圳市唯特视科技有限公司 | One kind carries out self-supervision insertion posture learning method based on video time-space relationship |
Non-Patent Citations (3)
| Title |
|---|
| XIN LI ET AL: "Video Processing Via Implicit and Mixture Motion Models", 《IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY》 * |
| 孔悦等: "基于时空一致性的立体视频稳像改进方法", 《电视技术》 * |
| 王峰等: "利用平稳光流估计的海上视频去抖", 《中国图象图形学报》 * |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109905565A (en) * | 2019-03-06 | 2019-06-18 | 南京理工大学 | Video stabilization method based on motor pattern separation |
| CN109905565B (en) * | 2019-03-06 | 2021-04-27 | 南京理工大学 | Video de-jittering method based on motion mode separation |
| CN110222234A (en) * | 2019-06-14 | 2019-09-10 | 北京奇艺世纪科技有限公司 | A kind of video classification methods and device |
| CN110472741A (en) * | 2019-06-27 | 2019-11-19 | 广东工业大学 | A kind of small wave width study filtering system of three-domain fuzzy and method |
| CN110472741B (en) * | 2019-06-27 | 2022-06-03 | 广东工业大学 | Three-domain fuzzy wavelet width learning filtering system and method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108924385B (en) | 2020-11-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US11537873B2 (en) | Processing method and system for convolutional neural network, and storage medium | |
| Celik | Spatial mutual information and PageRank-based contrast enhancement and quality-aware relative contrast measure | |
| US9619749B2 (en) | Neural network and method of neural network training | |
| KR102166105B1 (en) | Neural network and method of neural network training | |
| CN108924385A (en) | A kind of video stabilization method based on width study | |
| CN109961396B (en) | Image super-resolution reconstruction method based on convolutional neural network | |
| CN110717953B (en) | Coloring method and system for black-and-white pictures based on CNN-LSTM (computer-aided three-dimensional network-link) combination model | |
| CN109510918A (en) | Image processing apparatus and image processing method | |
| TWI664853B (en) | Method and device for constructing the sensing of video compression | |
| CN112823379A (en) | Method and device for training machine learning model and device for video style transfer | |
| Hristova et al. | Style-aware robust color transfer. | |
| CN108288253A (en) | HDR image generation method and device | |
| CN112037144A (en) | Low-illumination image enhancement method based on local contrast stretching | |
| CN111047543A (en) | Image enhancement method, device and storage medium | |
| CN109829510A (en) | A kind of method, apparatus and equipment of product quality classification | |
| CN109064431B (en) | Picture brightness adjusting method, equipment and storage medium thereof | |
| Daga et al. | Image compression using harmony search algorithm | |
| He et al. | A night low‐illumination image enhancement model based on small probability area filtering and lossless mapping enhancement | |
| CN115908602A (en) | Style migration method for converting landscape photos into Chinese landscape paintings | |
| Tao et al. | The discretization of continuous attributes based on improved SOM clustering | |
| Vamsidhar et al. | Image Enhancement Using Chicken Swarm Optimization | |
| Celebi et al. | Histogram Equalization for Grayscale Images and Comparison with OpenCV Library | |
| CN108038828B (en) | Image denoising method based on self-adaptive weighted total variation | |
| Mcvey et al. | Towards a generic neural network architecture for approximating tone mapping algorithms | |
| US20220237412A1 (en) | Method for modelling synthetic data in generative adversarial networks |
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 |