[go: up one dir, main page]
More Web Proxy on the site http://driver.im/

CN105791980A - Resolution improvement based film and TV works renovation method - Google Patents

Resolution improvement based film and TV works renovation method Download PDF

Info

Publication number
CN105791980A
CN105791980A CN201610109909.XA CN201610109909A CN105791980A CN 105791980 A CN105791980 A CN 105791980A CN 201610109909 A CN201610109909 A CN 201610109909A CN 105791980 A CN105791980 A CN 105791980A
Authority
CN
China
Prior art keywords
expert
resolution
parameter
resolution video
video frame
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
Application number
CN201610109909.XA
Other languages
Chinese (zh)
Other versions
CN105791980B (en
Inventor
张宏志
赵秋实
左旺孟
石坚
张垒磊
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Harbin Super-Resolution Fx Technology Co Ltd
Original Assignee
Harbin Super-Resolution Fx Technology Co Ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Harbin Super-Resolution Fx Technology Co Ltd filed Critical Harbin Super-Resolution Fx Technology Co Ltd
Priority to CN201610109909.XA priority Critical patent/CN105791980B/en
Publication of CN105791980A publication Critical patent/CN105791980A/en
Application granted granted Critical
Publication of CN105791980B publication Critical patent/CN105791980B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04NPICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N21/00Selective content distribution, e.g. interactive television or video on demand [VOD]
    • H04N21/40Client devices specifically adapted for the reception of or interaction with content, e.g. set-top-box [STB]; Operations thereof
    • H04N21/43Processing of content or additional data, e.g. demultiplexing additional data from a digital video stream; Elementary client operations, e.g. monitoring of home network or synchronising decoder's clock; Client middleware
    • H04N21/44Processing of video elementary streams, e.g. splicing a video clip retrieved from local storage with an incoming video stream or rendering scenes according to encoded video stream scene graphs
    • H04N21/4402Processing of video elementary streams, e.g. splicing a video clip retrieved from local storage with an incoming video stream or rendering scenes according to encoded video stream scene graphs involving reformatting operations of video signals for household redistribution, storage or real-time display
    • H04N21/440263Processing of video elementary streams, e.g. splicing a video clip retrieved from local storage with an incoming video stream or rendering scenes according to encoded video stream scene graphs involving reformatting operations of video signals for household redistribution, storage or real-time display by altering the spatial resolution, e.g. for displaying on a connected PDA
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/70Information retrieval; Database structures therefor; File system structures therefor of video data
    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04NPICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N21/00Selective content distribution, e.g. interactive television or video on demand [VOD]
    • H04N21/40Client devices specifically adapted for the reception of or interaction with content, e.g. set-top-box [STB]; Operations thereof
    • H04N21/43Processing of content or additional data, e.g. demultiplexing additional data from a digital video stream; Elementary client operations, e.g. monitoring of home network or synchronising decoder's clock; Client middleware
    • H04N21/44Processing of video elementary streams, e.g. splicing a video clip retrieved from local storage with an incoming video stream or rendering scenes according to encoded video stream scene graphs
    • H04N21/4402Processing of video elementary streams, e.g. splicing a video clip retrieved from local storage with an incoming video stream or rendering scenes according to encoded video stream scene graphs involving reformatting operations of video signals for household redistribution, storage or real-time display
    • H04N21/440236Processing of video elementary streams, e.g. splicing a video clip retrieved from local storage with an incoming video stream or rendering scenes according to encoded video stream scene graphs involving reformatting operations of video signals for household redistribution, storage or real-time display by media transcoding, e.g. video is transformed into a slideshow of still pictures, audio is converted into text

Landscapes

  • Engineering & Computer Science (AREA)
  • Multimedia (AREA)
  • Signal Processing (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Databases & Information Systems (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Television Systems (AREA)
  • Image Processing (AREA)

Abstract

The present invention provides a resolution improvement based film and TV works renovation method for film and TV works with low resolution and low definition. The method includes the specific steps of firstly, acquiring a resolution of an original video and a target resolution, and calculating a zoom ratio; then, segmenting an input video into an image frame set according to a certain segmentation manner; then, performing conversion according to a prestored mapping relation, so as to obtain high resolution video frames; and finally, combing the high resolution video frames to obtain a high resolution video. The prestored mapping relation is obtained by learning a hybrid expert model, and the process is finished offline in a computer. The method has the advantages of good adaptability, fast speed, good effect, expandability and so on.

Description

Films and television programs based on increase resolution renovate method
Technical field
The invention belongs to computer vision and image processing field, relate to films and television programs renovation method, be specifically related to a kind of films and television programs based on increase resolution and renovate method and system.
Background technology
Along with video acquisition, transmission, storage, Display Technique development, films and television programs constantly towards high-resolution develop.The taste that people appreciate video is also more and more higher, constantly pursues the films and television programs of high-resolution, fine definition.Meanwhile, the appearance (such as 4K, 5K TV and display) of high-resolution display device, make again the universal of high-resolution films and television programs be possibly realized.
But on the other hand, many classical films and television programs of the remote past limit due to technological means, still have relatively low resolution, relatively low definition and poor visual effect.Simultaneously because of the remote past, the film holding time is longer, adds the destruction of the extraneous factor such as natural disaster, war, various quality degradation occurs, such as breakage, flicker, noise, shake etc..People want to review classical films and television programs on the one hand, on the one hand the quality of film have been had again new requirement.Reviewing classical films and television programs to meet people simultaneously and pursue the demand of high-quality video, films and television programs reconditioning technology arises at the historic moment.It is application image/video processing technique that films and television programs renovate its essence, and original video is processed, and eliminates various quality degradation, to improve the visual effect of original video.
Films and television programs are the same with books, are the important culture carriers of human society, some classical films and television programs, although of the remote past, but its cultureal value cannot replace.Therefore, the films and television programs that these are of the remote past are renovated, replay has just shown very important meaning.Specifically, the meaning of films and television programs renovation includes the following aspects:
1. some classical films and television programs, such as documentary film etc., are precious historical summaries.These films and television programs are renovated, can better preserve, propagate these historical summaries.
2. pair classical films and television programs renovate, and allow more modern appreciate, and are by the important form of culture and arts succession.
3. pair classical films and television programs renovate, and allow classical artistic work taking on a new look again, are to artistical maximum respect and souvenir.
The method of existing raising films and television programs visual effect is concentrated mainly in video source modeling means, and as removed noise, removal is fuzzy, remove staggered, enhancing contrast ratio, enhancing color etc..Former video can be played the effect that visual effect strengthens by these methods, but video is not carried out increase resolution, so there is no the demand inherently meeting people to classical films and television programs renovation.
So-called increase resolution, refers to the video (or frame of video) of low resolution, by certain method, generates a high-resolution video fast and effectively.Its difficult point is in that how to break through the restriction of original low-resolution video pixel quantity, fill original non-existent pixel, should keep the structure of former low-resolution video, texture, more naturally reasonable at human eye again.
Traditional increase resolution method, mainly includes based on interpolation, based on rebuilding and based on the method learnt.Method based on interpolation is by existing pixel is carried out linear combination, as the pixel of disappearance.Interpolation algorithm is simply rapid, but " mosaic " effect easily occurs or cross smooth phenomenon;Utilizing the similarity of multiple image to carry out registration reconstruction based on the algorithm rebuild, but this kind of algorithm is often simple combination multiple image, effect is unsatisfactory;Based on the algorithm learnt mainly by a number of training data, obtaining the low-resolution video mapping relations to high-resolution video according to special algorithm training, this type of algorithm is higher to the requirement of model, easy over-fitting or poor fitting, and operand is big, speed slow, practicality is not high.It can be said that above-mentioned video resolution Upgrade Problem annoyings users always.
Summary of the invention
The technology of the present invention solves problem: the present invention provides a kind of films and television programs renovation method and system promoted based on video resolution, by the films and television programs (being usually less than 720P) of low resolution through resolution enhancement technology, be converted to high-resolution video (such as 1080P, 4K etc.) and realize films and television programs renovation.
The technical solution of the present invention is: the concrete scheme of the present invention is: first, obtains resolution and the target resolution of original video, calculates scaling;Secondly, input video is divided into picture frame by certain partitioning scheme;Then, the mapping relations according to prestoring convert, and obtain high-resolution video frame;Finally, high-resolution video frame is combined into high-resolution video.Wherein, the mapping relations model prestored obtains based on Mixture of expert model learning, and this model training process off-line in a computer completes.Concrete steps bag is as follows:
Study mapping relations model:
(1) pretreatment training video
(1.1) choose high-resolution video as training sample, and be split as high-resolution video frame;
(1.2) the high-resolution video frame of step (1.1) gained is used gaussian kernel convolution
(1.3) calculate amplification according to original low-resolution video and target high-resolution video, carry out partiting row sampling according to the multiple of gained, obtain the low resolution video frame of correspondence;
(1.4) low resolution video frame of high-resolution video frame and sampling gained is split in bulk respectively, as training data.
(2) mapping relations model is obtained based on Mixture of expert model
(2.1) a Mixture of expert model is initialized.Mixture of expert model includes expert and gate function two parts, and its structure is tree-like, as shown in Figure 2.In figure, the leafy node in tree structure is called expert, is responsible for data are carried out mapping transformation;Root node is called gate function, is responsible for data and selects suitable expert.The present invention uses linear function as expert's function:
Y=Wx
Wherein W is expert's function parameter, and x and y represents low resolution video frame block and corresponding high-resolution video frame block respectively.
Gate function is responsible for determining to select which expert that data are converted, and in the present invention, i-th gate function is expressed as:
g i ( x , v i ) = exp ( x - v i ) 2 ( Σ j = 1 K ( x - v j ) 2 ) - 1
Wherein, x and y represents low resolution video frame block and corresponding high-resolution video frame block, v respectivelyiRepresent i-th gate function parameter, vjRepresenting jth gate function parameter, K is the number of expert in Mixture of expert model, i.e. the number of leaf node in tree structure.The initialization of Mixture of expert model specifically includes following steps:
(2.1.1) the quantity K of expert is specified;
(2.1.2) the probability distribution Gaussian distributed of each expert is supposed: p (y | x, Wi)=N (y (x, Wi), σ), wherein WiRepresenting the parameter of i-th expert, σ is the standard deviation of Gauss distribution.Assuming that parameter WiDistribution also Gaussian distributed: p (Wi)=N (0, μ), wherein μ represents the average of Gauss distribution.
(2.1.3) k-mean algorithm is adopted to be clustered according to the quantity K of expert by training data, the initial value W of the parameter of each experti (0)It is appointed as in class slope, the initial value v of each gate function parameteri (0)It is appointed as cluster centre;
(2.1.4) initial value of each gate function is calculated:
g i ( 0 ) ( x , v i ( 0 ) ) = exp ( x - v i ( 0 ) ) 2 ( Σ j = 1 K ( x - v j ( 0 ) ) 2 ) - 1
Wherein x represents low resolution video frame block, vi (0)Representing the initial value of i-th gate function parameter, K is the number of expert in Mixture of expert model, i.e. the number of leaf node in tree structure.
(2.2) training data using step (1.4) to obtain, is iterated Mixture of expert model optimizing, until iterative process convergence, the model parameter finally given is mapping relations model.Mapping relations model includes gate function parameter and expert parameter.
(2.2.1) allowable error ε during iteration ends is specified;
(2.2.2) posterior probability of each gate function in epicycle iteration is calculated:
h i ( k + 1 ) ( x ) = g i ( k ) ( x , v i ( k ) ) p i ( y | x , W i ( k ) ) Σ j = 1 K g i ( k ) ( x , v i ( k ) ) p j ( y | x , W j ( k ) )
Wherein k is iterative steps, pi(y | x, Wi (k)) and pj(y | x, Wj (k)) represent expert probability distribution, gi (k)(x, vi (k)) represent that the kth of i-th gate function walks iterative value.
(2.2.3) each expert parameter is updated:
W i ( k + 1 ) = YH i ( k + 1 ) X T μ 2 XH i ( k + 1 ) X T μ 2 + σ 2 I
Wherein k is iterative steps, and X is the vector of all low resolution block x composition in training data, and Y is the vector of all high-resolution block y composition, X in training dataTRepresent the transposition of X, I representation unit matrix, Hi (k+1)Represent in kth+1 step the vector of the posterior probability composition of all low resolution block x corresponding to i-th expert.
(2.2.4) each gate function parameter is updated:
v i ( k + 1 ) = v i ( k ) - ( x ( t ) - v i ( k ) ) Σ t = 1 N ( h i ( k ) ( x ( t ) ) - g i ( x ( t ) , v i ( k ) ) )
WhereinRepresent i-th gate function parameter in kth step iteration,The posterior probability of i-th gate function, x in iteration is walked for kth(t)Represent the t low resolution block.
(2.2.5) output of each gate function in epicycle iteration is calculated:
g i ( k + 1 ) ( x , v i ( k + 1 ) ) = exp ( x - v i ( k + 1 ) ) 2 ( Σ j = 1 K ( x - v j ( k + 1 ) ) 2 ) - 1
(2.2.6) likelihood probability in epicycle iteration is calculated:
Q ( k + 1 ) = Σ t = 1 N Σ i = 1 K h i ( k + 1 ) ( x ) ( log g i ( k + 1 ) ( x , v i ( k + 1 ) ) + log p i ( y | x , W i ( k + 1 ) ) p ( W i ( k + 1 ) ) )
Wherein, pi(y | x, Wi (k+1)) represent expert probability distribution, p (Wi (k+1)) represent expert parameter probability distribution.
(2.2.7) judge whether iteration restrains.During allowable error ε when the absolute value of difference of likelihood probability of likelihood probability and last round of iteration of epicycle iteration is less than iteration ends, finishing iteration.Otherwise repeated execution of steps (2.2.2)~(2.2.7).
The gate function parameter v obtained when iteration terminatesi, together with expert quantity K, expert parameter Wi, the standard deviation sigma of probability distribution of expert and expert parameter mean of a probability distribution μ, be stored in disk together as final mapping relations model.
After having learnt mapping relations model and having stored, use the mapping relations model of storage that video is carried out increase resolution:
(3) low-resolution video that pretreatment is pending
(3.1) low-resolution video is split as low resolution video frame;
(3.2) low resolution video frame of step (3.1) gained is divided into low resolution video frame block;
(4) the mapping relations model that low-resolution video obtains according to step (2) is promoted to high-resolution video, including:
(4.1) low resolution video frame block step (3) obtained is as the input of gate function, and the gate function parameter in the mapping relations model that use step (2) obtains calculates the output of each gate function:
g i ( x , v i ) = exp ( x - v i ) 2 ( Σ j = 1 K ( x - v j ) 2 ) - 1
Wherein, x is the low resolution video frame block of input.K is the number of expert, v in Mixture of expert modeliRepresent i-th gate function parameter, obtain each through step (2.2).
(4.2) parameter of the expert's function corresponding to gate function that use output valve is maximum calculates corresponding high-resolution block, and wherein the parameter of expert's function is obtained by step (2);
(4.2.1) calculating obtains the sequence number of maximum output valve gate function: i=argmax (gi)
Wherein, giFor the output of i-th gate function, obtained by step (4.1).
(4.2.2) i-th expert's function is used to calculate high-resolution video frame block: y=Wix
Wherein, WiFor the parameter of i-th expert's function, y is the high-resolution video frame block corresponding to low resolution video frame block x of input.
(4.3) each low resolution block is carried out increase resolution according to the step of (4.1) and (4.2), obtain the high-resolution block of correspondence, all of high-resolution block is spliced into according to position in low resolution video frame of the low resolution block of its correspondence the high-resolution video frame of correspondence;
(4.4), after obtaining the high-resolution video frame that all low resolution video frame are corresponding, it is combined into high-resolution video.
In described step (4), all without dependence between frame of video block, between frame of video, therefore can using GPU processor, parallel acceleration processes this step.
The described films and television programs based on increase resolution renovate method, it is possible to use with the form of computer software player, it is also possible to be integrated in hardware platform (such as Set Top Box, intelligent television etc.) and use.
The described films and television programs based on increase resolution renovate method, it is possible to coordinate other screen Enhancement Method as pretreatment or post processing means, it is possible to promote visual effect further.
Present invention advantage compared with prior art is in that: visual effect, and the high-resolution video details obtained by enforcement the present invention program is complete, edge clear, Acacia crassicarpaA are good, and fast and stable.Specifically, the feature of the present invention includes:
1. self adaptation.The present invention program adaptive polo placement scaling multiple, is suitable for different increase resolution demands.
2. speed is fast.Owing to being absent from dependence between sequence of frames of video, it is possible to improve processing speed by parallel processing.It addition, the processing procedure of algorithm is that sequence of frames of video is carried out Linear Mapping conversion, mapping parameters used can pre-save in memory, it is possible to improves processing speed further.
3. effective.The mapping parameters that the present invention carries out using in increase resolution process is based on Mixture of expert model learning and obtains, and has broken and has divided the drawback learning to separate with submodel in the tradition increase resolution algorithm based on study.Simultaneously, combine statistical robustness advantage and the accuracy advantage based on learning algorithm, avoid the drawback that in the past can not utilize mass data information based on learning algorithm, simultaneously also achieve the precision higher than simple statistical method, also can obtain good effect and processing speed faster even for the video that resolution own is higher.
4. expansible.Owing to being absent from dependence between sequence of frames of video, it is possible to realize parallel processing raising processing speed by applying the technological means such as GPU acceleration.It addition, algorithm proposed by the invention may be directly applied to image resolution ratio promotes field.
Accompanying drawing explanation
Fig. 1 is the flow chart that the films and television programs based on increase resolution of the present invention renovate method.
Fig. 2 is Mixture of expert model structure schematic diagram of the present invention.
Frame of video of the present invention is divided into frame of video block schematic diagram by Fig. 3.
Detailed description of the invention
The method of the invention is illustrated by detailed description below with an example.
Renovate method according to the films and television programs based on increase resolution of the present invention, the process that the video that original resolution is 768*432 is promoted to 3072*1728 comprised the following steps:
(1) pretreatment training video
(1.1) high-resolution films and television programs are selected, Video processing software is utilized to read in the video flowing of films and television programs, each frame in video flowing is saved as frame of video, in the present embodiment, films and television programs length is 1200 seconds, frame rate is 25 frames/second, and gained frame of video adds up to: 1200*25=15000;
(1.2) frame of video of step (1.1) gained being used average is 0, and standard deviation is the gaussian kernel convolution of 1;
(1.3) obtain original low-resolution video resolution and target resolution, calculate amplification according to both.Original resolution is 768*432, and target resolution is 3072*1728, and amplification is: 3072/768=4.Accordingly, the frame of video after convolution is down sampled to the 1/4 of original size, obtains the low resolution video frame of correspondence.
(1.4) each width low resolution video frame of gained by existing segmentation standard, it is divided into being sized to the not overlapping fritter of 10 × 10 pixels, as it is shown on figure 3, and therefrom choose 1,000,000 piece as training data.
(2) mapping relations model is obtained based on Mixture of expert model
(2.1) Mixture of expert model is initialized
(2.1.1) the quantity K of expert is specified.In the present embodiment, take K=100;
(2.1.2) specify parameter σ and the μ of the probability distribution of expert and the probability distribution of expert parameter, in the present embodiment, take σ=0.32, μ=0.58;
(2.1.3) k-mean algorithm is adopted to be clustered according to the quantity K of expert by training data, the W of each experti (0)Parameter initialization is slope in class, gate function parameter vi (0)It is initialized as cluster centre;
(2.1.4) initial value of each gate function is calculated according to following formula:
g i ( 0 ) ( x , v i ( 0 ) ) = exp ( x - v i ( 0 ) ) 2 ( Σ j = 1 K ( x - v j ( 0 ) ) 2 ) - 1
Wherein x represents low resolution video frame block, vi (0)Representing the initial value of i-th gate function parameter, K is the number of expert in Mixture of expert model, i.e. the number of leaf node in tree structure.
(2.2) using the training data that step (1.4) obtains, the Mixture of expert model that step (2.1) is obtained is iterated optimizing:
(2.2.1) allowable error ε during iteration ends is specified.In the present embodiment, error ε=0.005 that delivery type iteration ends allows.
(2.2.2) posterior probability of each gate function in epicycle iteration is calculated:
h i ( k + 1 ) ( x ) = g i ( k ) ( x , v i ( k ) ) p i ( y | x , W i ( k ) ) Σ j = 1 K g i ( k ) ( x , v i ( k ) ) p j ( y | x , W j ( k ) )
Wherein k is iterative steps, pi(y | x, Wi (k)) and pj(y | x, Wj (k)) represent expert probability distribution, gi (k)(x, vi (k)) represent that the kth of i-th gate function walks iterative value.
(2.2.3) each expert parameter is updated:
W i ( k + 1 ) = YH i ( k + 1 ) X T μ 2 XH i ( k + 1 ) X T μ 2 + σ 2 I
Wherein k is iterative steps, and X is the vector of all low resolution block x composition in training data, and Y is the vector of all high-resolution block y composition, X in training dataTRepresent the transposition of X, I representation unit matrix, Hi (k+1)Represent in kth+1 step the vector of the posterior probability composition of all low resolution block x corresponding to i-th expert.
(2.2.4) each gate function parameter is updated:
v i ( k + 1 ) = v i ( k ) - ( x ( t ) - v i ( k ) ) Σ t = 1 N ( h i ( k ) ( x ( t ) ) - g i ( x ( t ) , v i ( k ) ) )
WhereinRepresent i-th gate function parameter in kth step iteration,The posterior probability of i-th gate function, x in iteration is walked for kth(t)Represent the t low resolution block.
(2.2.5) output of each gate function in epicycle iteration is calculated:
g i ( k + 1 ) ( x , v i ( k + 1 ) ) = exp ( x - v i ( k + 1 ) ) 2 ( Σ j = 1 K ( x - v j ( k + 1 ) ) 2 ) - 1
(2.2.6) likelihood probability in epicycle iteration is calculated:
Q ( k + 1 ) = Σ t = 1 N Σ i = 1 K h i ( k + 1 ) ( x ) ( log g i ( k + 1 ) ( x , v i ( k + 1 ) ) + log p i ( y | x , W i ( k + 1 ) ) p ( W i ( k + 1 ) ) )
Wherein, pi(y | x, Wi (k+1)) represent expert probability distribution, p (Wi (k+1)) represent expert parameter probability distribution.
(2.2.7) judge whether iteration restrains.During allowable error ε when the absolute value of difference of likelihood probability of likelihood probability and last round of iteration of epicycle iteration is less than iteration ends, finishing iteration.Otherwise repeated execution of steps (2.2.2)~(2.2.7).
The gate function parameter v obtained when iteration terminatesi, together with expert quantity K, expert parameter Wi, the standard deviation sigma of probability distribution of expert and expert parameter mean of a probability distribution μ, be stored in disk together as final mapping relations model.Wherein, vi, K, σ, μ be called the gate function parameter of mapping relations model, WiIt is called the expert parameter of mapping relations model.
(3) low-resolution video that pretreatment is pending
(3.1) pending low-resolution video being split as low resolution video frame, in the present embodiment, films and television programs length is 2000 seconds, and frame rate is 25 frames/second, and gained frame of video adds up to: 2000*25=50000;
(3.2) low resolution video frame of step (3.1) gained is divided into the frame of video block of 10 × 10, as shown in Figure 3;
(4) low-resolution video is mapped as high-resolution video, including:
(4.1) low resolution video frame block step (3) obtained is as the input of gate function, and the gate function parameter in the mapping relations model that use step (2) obtains calculates the output of each gate function:
g i ( x , v i ) = exp ( x - v i ) 2 ( Σ j = 1 K ( x - v j ) 2 ) - 1
Wherein, x is the low resolution video frame block of input.K is the number of expert, v in Mixture of expert modeliRepresent i-th gate function parameter, obtain each through step (2.2).
(4.2) the expert's function parameter corresponding to gate function using output valve maximum calculates corresponding high-resolution block;
(4.2.1) calculating obtains the sequence number of maximum output valve gate function: i=argmax (gi).Wherein, giFor the output of i-th gate function, obtained by step (4.1).
(4.2.2) i-th expert's function is used to calculate high-resolution video frame block:
Y=Wix
Wherein, WiFor i-th expert's function parameter, obtained by step (2).X is the low resolution video frame block of input, and it is sized to 10 × 10, and y is the high-resolution video frame block after increase resolution, is sized to 40 × 40.
(4.3) each low resolution block is carried out increase resolution according to the step of (4.1) and (4.2), obtain the high-resolution block of correspondence, all of high-resolution block is spliced into according to position in low resolution video frame of the low resolution block of its correspondence the high-resolution video frame of correspondence;
(4.4), after obtaining the high-resolution video frame that all low resolution video frame are corresponding, it is combined into high-resolution video.

Claims (5)

1. the films and television programs based on increase resolution renovate method, it is characterised in that: include study mapping relations model and carry out increase resolution two parts according to mapping relations model;
Wherein, study mapping relations model includes following two step:
(1) pretreatment training video, including:
(1.1) choose high-resolution video as training sample, and be split as high-resolution video frame;
(1.2) the high-resolution video frame of step (1.1) gained is used gaussian kernel convolution
(1.3) calculate amplification according to original low-resolution video and target high-resolution video, carry out partiting row sampling according to the multiple of gained, obtain the low resolution video frame of correspondence;
(1.4) low resolution video frame of high-resolution video frame and sampling gained is split in bulk respectively, as training data.
(2) mapping relations model is obtained based on Mixture of expert model, including:
(2.1) a Mixture of expert model is initialized;
(2.2) training data using step (1.4) to obtain, is iterated Mixture of expert model optimizing, until iterative process convergence, the model parameter finally given is mapping relations model.Mapping relations model includes gate function parameter and expert parameter.
Carry out increase resolution according to mapping relations model and include following two step:
(3) low-resolution video that pretreatment is pending, including:
(3.1) pending low-resolution video is split as low resolution video frame;
(3.2) low resolution video frame is split in bulk.
(4) the mapping relations model that low-resolution video obtains according to step (2) is promoted to high-resolution video, including:
(4.1) low resolution video frame block step (3) obtained is as the input of Mixture of expert model gate function, and the gate function parameter in the mapping relations model that use step (2) obtains calculates the output of each gate function;
(4.2) parameter of the expert's function corresponding to gate function that use output valve is maximum calculates corresponding high-resolution block, and wherein the parameter of expert's function is obtained by step (2);
(4.3) each low resolution block is carried out increase resolution according to the step of (4.1) and (4.2), obtain the high-resolution block of correspondence, all of high-resolution block is spliced into according to position in low resolution video frame of the low resolution block of its correspondence the high-resolution video frame of correspondence;
(4.4), after obtaining the high-resolution video frame that all low resolution video frame are corresponding, it is combined into high-resolution video.
2. the films and television programs based on increase resolution according to claim 1 renovate method, it is characterised in that the Mixture of expert model described in step (2.1) includes expert and gate function two parts.
Expert is responsible for data are carried out mapping transformation, and the mapping transformation in the present invention uses linear function as expert's function:
Y=Wx
Wherein W is expert parameter, and x and y represents low resolution video frame block and corresponding high-resolution video frame block respectively;
Gate function is responsible for determining to select which expert that data are converted, and in the present invention, i-th gate function is expressed as:
Wherein, x and y represents low resolution video frame block and corresponding high-resolution video frame block, v respectivelyiRepresent i-th gate function parameter, vjRepresenting jth gate function parameter, K is the number of expert in Mixture of expert model.
3. the films and television programs based on increase resolution according to claim 1 renovate method, it is characterised in that: the described initialization Mixture of expert model of step (2.1) comprises the following steps:
1. the quantity K of expert is specified;
2. the probability distribution Gaussian distributed of each expert is supposed: p (y | x, Wi)=N (y (x, Wi), σ), wherein WiRepresenting the parameter of i-th expert, σ is the standard deviation of Gauss distribution.Assuming that parameter WiDistribution also Gaussian distributed: p (Wi)=N (0, μ), wherein μ represents the average of Gauss distribution.
3. k-mean algorithm is adopted to be clustered according to the quantity K of expert by training data, the initial value W of the parameter of each experti (0)It is appointed as in class slope, the initial value v of each gate function parameteri (0)It is appointed as cluster centre;
4. the initial value of each gate function is calculated:
Wherein x represents low resolution video frame block, vi (0)Representing the initial value of i-th gate function parameter, K is the number of expert in Mixture of expert model, i.e. the number of leaf node in tree structure.
4. the films and television programs based on increase resolution according to claim 1 renovate method, it is characterised in that: step (2.2) is described to be iterated optimization to model and comprises the following steps:
1. allowable error ε during iteration ends is specified;
2. the posterior probability of each gate function in epicycle iteration is calculated:
Wherein k is iterative steps, pi(y | x, Wi (k)) and pj(y | x, Wj (k)) represent expert probability distribution, gi (k)(x, vi (k)) represent that the kth of i-th gate function walks iterative value.
3. each expert parameter is updated:
Wherein k is iterative steps, and X is the vector of all low resolution block x composition in training data, and Y is the vector of all high-resolution block y composition, X in training dataTRepresent the transposition of X, I representation unit matrix, Hi (k+1)Represent in kth+1 step the vector of the posterior probability composition of all low resolution block x corresponding to i-th expert.
4. each gate function parameter is updated:
WhereinRepresent i-th gate function parameter in kth step iteration,The posterior probability of i-th gate function, x in iteration is walked for kth(t)Represent the t low resolution block.
5. the output of each gate function in epicycle iteration is calculated:
6. the likelihood probability in epicycle iteration is calculated:
Wherein, pi(y | x, Wi (k+1)) represent expert probability distribution, p (Wi (k+1)) represent expert parameter probability distribution.
7. judge whether iteration restrains.During allowable error ε when the absolute value of difference of likelihood probability of likelihood probability and last round of iteration of epicycle iteration is less than iteration ends, finishing iteration.Otherwise repeated execution of steps 2.~7..
The gate function parameter v obtained when iteration terminatesi, together with expert quantity K, expert parameter Wi, the standard deviation sigma of probability distribution of expert and expert parameter mean of a probability distribution μ, be stored in disk together as final mapping relations model.
5. the films and television programs based on increase resolution according to claim 1 renovate method, it is characterised in that: the parameter of the expert's function corresponding to gate function that step (4.2) described use output valve is maximum calculates corresponding high-resolution block and comprises the following steps:
1. calculating obtains the sequence number of maximum output valve gate function: i=argmax (gi)
Wherein, giFor the output of i-th gate function, obtained by step (4.1).
2. i-th expert's function is used to calculate high-resolution video frame block: y=Wix
Wherein, WiFor the parameter of i-th expert's function, y is the high-resolution video frame block corresponding to low resolution video frame block x of input.
CN201610109909.XA 2016-02-29 2016-02-29 Films and television programs renovation method based on increase resolution Active CN105791980B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201610109909.XA CN105791980B (en) 2016-02-29 2016-02-29 Films and television programs renovation method based on increase resolution

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201610109909.XA CN105791980B (en) 2016-02-29 2016-02-29 Films and television programs renovation method based on increase resolution

Publications (2)

Publication Number Publication Date
CN105791980A true CN105791980A (en) 2016-07-20
CN105791980B CN105791980B (en) 2018-09-14

Family

ID=56403789

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201610109909.XA Active CN105791980B (en) 2016-02-29 2016-02-29 Films and television programs renovation method based on increase resolution

Country Status (1)

Country Link
CN (1) CN105791980B (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109151339A (en) * 2018-08-27 2019-01-04 深圳艺达文化传媒有限公司 Character synthesizing method and Related product in promotion video
CN109981991A (en) * 2019-04-17 2019-07-05 北京旷视科技有限公司 Model training method, image processing method, device, medium and electronic equipment
CN114207640A (en) * 2018-11-06 2022-03-18 同步摄录公司 High resolution movie creation and management system

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101639937A (en) * 2009-09-03 2010-02-03 复旦大学 Super-resolution method based on artificial neural network
US20100290529A1 (en) * 2009-04-14 2010-11-18 Pankaj Topiwala Real-time superresolution and video transmission
CN104778671A (en) * 2015-04-21 2015-07-15 重庆大学 Image super-resolution method based on SAE and sparse representation

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20100290529A1 (en) * 2009-04-14 2010-11-18 Pankaj Topiwala Real-time superresolution and video transmission
CN101639937A (en) * 2009-09-03 2010-02-03 复旦大学 Super-resolution method based on artificial neural network
CN104778671A (en) * 2015-04-21 2015-07-15 重庆大学 Image super-resolution method based on SAE and sparse representation

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
KAI ZHANG ET AL: "Joint Learning of Multiple Regressors for Single Image Super-Resolution", 《IEEE SIGNAL PROCESSING LETTERS》 *

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109151339A (en) * 2018-08-27 2019-01-04 深圳艺达文化传媒有限公司 Character synthesizing method and Related product in promotion video
CN114207640A (en) * 2018-11-06 2022-03-18 同步摄录公司 High resolution movie creation and management system
CN109981991A (en) * 2019-04-17 2019-07-05 北京旷视科技有限公司 Model training method, image processing method, device, medium and electronic equipment

Also Published As

Publication number Publication date
CN105791980B (en) 2018-09-14

Similar Documents

Publication Publication Date Title
Golts et al. Unsupervised single image dehazing using dark channel prior loss
US11055828B2 (en) Video inpainting with deep internal learning
Zhang et al. Semantic image inpainting with progressive generative networks
CN105744357B (en) A kind of reduction network video bandwidth occupancy method based on online increase resolution
CN107507134B (en) Super-resolution method based on convolutional neural network
Wang et al. Robust object removal with an exemplar-based image inpainting approach
CN105243670B (en) A kind of sparse and accurate extracting method of video foreground object of low-rank Combined expression
DE102017009049A1 (en) Enable transformations sketch-to-painting
CN107590791B (en) Image enhancement method and image processing apparatus
CN110570377A (en) group normalization-based rapid image style migration method
CN103489174B (en) A kind of face super-resolution method kept based on residual error
CN103093444A (en) Image super-resolution reconstruction method based on self-similarity and structural information constraint
CN110717868B (en) Video high dynamic range inverse tone mapping model construction and mapping method and device
US20150071545A1 (en) Image Enhancement Using Self-Examples and External Examples
CN112561846A (en) Method and device for training image fusion model and electronic equipment
CN114049464B (en) Reconstruction method and device of three-dimensional model
CN104166967A (en) Method for improving definition of video image
Wang et al. Enhancement for dust-sand storm images
CN109191460A (en) A kind of quality evaluating method for tone mapping image
CN110533579B (en) Video style conversion method based on self-coding structure and gradient order preservation
CN110428382A (en) A kind of efficient video Enhancement Method, device and storage medium for mobile terminal
Zhang et al. Deep tone mapping network in HSV color space
Lv et al. Low-light image enhancement via deep Retinex decomposition and bilateral learning
CN105791980A (en) Resolution improvement based film and TV works renovation method
CN114339030A (en) Network live broadcast video image stabilization method based on self-adaptive separable convolution

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
C10 Entry into substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant