Summary of the invention
Technical problem to be solved by this invention is to provide a kind of infrared blind element backoff algorithm based on mixed self-adapting regression model,From multiscale analysis angle, combining adaptive regression model, has solved blind element has been detected to the problem too relying on, simultaneously for toolThere are a large amount of images random or fixing blind element all to there is good adaptability, and verified by experiment validity and the reality of new algorithmThe property used.
The present invention is for solving the problems of the technologies described above by the following technical solutions:
The infrared blind element compensation method returning based on mixed self-adapting, comprises following concrete steps:
Step 1, by Curvelet transfer pair infrared image, I carries out multiple dimensioned decomposition, obtains the each scalogram picture after decomposingI1,…,In-1,In, n is natural number;
Step 2, returns compensation to the each yardstick image configuration nonparametric model after decomposing, and recovers to realize blind element, obtainsRecover the image I of blind element1r,…,In-1r,Inr; Wherein, described structure nonparametric model returns compensation, is specially:
First, utilize blind element to detect and obtain blind element index matrix, learn the particular location of blind element;
Then, in the each scalogram picture from decomposing, obtain the neighborhood territory pixel collection of its blind element, carry out nonparametric by nonparametric modelEstimate, obtain the estimated value of blind element pixel, thus the image of the blind element that is restored;
Step 3, builds auto-regressive parameter model to recovering in step 2 between the image after blind element, carries out the linking between each yardstickWith study, by the image I of parametric technique demosaicing blind elementnrObtain interpolation image In-1i, and set it as n-1 layer yardstickPrior image; Wherein, described structure auto-regressive parameter model, is specially:
First, from the recovery blind element image of step 2, obtain known blind element set of pixels and neighborhood territory pixel collection thereof, composition linear equationAnd try to achieve total least square solution;
Then, utilize the linearity of blind element set of pixels in required total least square solution, interpolation image and neighborhood territory pixel collection composition thereofEquation, tries to achieve blind element set of pixels and the concentrated unknown pixel of neighborhood territory pixel thereof in interpolation image by least square method, therebyTo interpolation image;
Step 4, step-by-step recursion step 2-step 3, finally exports Recovery image.
As further prioritization scheme of the present invention, in step 2 to repeatedly iteration of nonparametric model, to improve non-parametric estmationPrecision; Described iterations is three times.
The present invention adopts above technical scheme compared with prior art, has overcome existing algorithm and has too relied on blind element detection and self adaptationThe deficiency of ability, looks for another way, and from multiscale analysis angle combining adaptive regression model, has solved blind element was detectedIn the problem relying on, all there is good adaptability for thering are a large amount of images random or fixing blind element simultaneously, and by experimentValidity and the practicality of new algorithm are verified. Blind element compensation strong adaptability of the present invention, for having a large amount of isolated and blind elements bunchImage all can effective compensation.
Detailed description of the invention
Below in conjunction with accompanying drawing, technical scheme of the present invention is described in further detail:
The infrared blind element compensation method returning based on mixed self-adapting, as shown in Figure 1, comprises following concrete steps:
Step 1, by Curvelet transfer pair infrared image, I carries out multiple dimensioned decomposition, obtains the each scalogram picture after decomposingI1,…,In-1,In, n is natural number;
Step 2, returns compensation to the each yardstick image configuration nonparametric model after decomposing, and recovers to realize blind element, obtainsRecover the image I of blind element1r,…,In-1r,Inr; Wherein, described structure nonparametric model returns compensation, is specially:
First, utilize blind element to detect and obtain blind element index matrix, learn the particular location of blind element;
Then, in the each scalogram picture from decomposing, obtain the neighborhood territory pixel collection of its blind element, carry out nonparametric by nonparametric modelEstimate, obtain the estimated value of blind element pixel, thus the image of the blind element that is restored;
Step 3, builds auto-regressive parameter model to recovering in step 2 between the image after blind element, carries out the linking between each yardstickWith study, by the image I of parametric technique demosaicing blind elementnrObtain interpolation image In-1i, and set it as n-1 layer yardstickPrior image; Wherein, described structure auto-regressive parameter model, is specially:
First, from the recovery blind element image of step 2, obtain known blind element set of pixels and neighborhood territory pixel collection thereof, composition linear equationAnd try to achieve total least square solution;
Then, utilize the linearity of blind element set of pixels in required total least square solution, interpolation image and neighborhood territory pixel collection composition thereofEquation, tries to achieve blind element set of pixels and the concentrated unknown pixel of neighborhood territory pixel thereof in interpolation image by least square method, therebyTo interpolation image;
Step 4, step-by-step recursion step 2-step 3, finally exports Recovery image.
By following specific embodiment, method of the present invention is further elaborated, as shown in Figure 2.
1, input blind element imageBy imageConvert respectively and form by CurveletThe embodiment decomposition that haves three layers altogether,Be specially: first willBe decomposed into low frequency sub-bandAnd high-frequency sub-bandThen high-frequency sub-band is decomposed into 2 through direction bank of filters againIndividual directional subband, wherein low frequency sub-band is
2, estimate by nonparametric modelMiddle blind element pixel value, obtains low-resolution imageSpecific as follows:
1) utilize blind element to detect the blind element index matrix obtaining, the particular location of known blind element; Use y1,…,ynRepresent that n the unknown is blindUnit, n is natural number, y0For the general name of unknown blind element;
2) fromIn obtain its set of pixels { y1,…,ynAnd { x0,x1,…,xn};x0,x1,…,xnFor blind element y0Neighborhood territory pixel;
If may there is blind element pixel in neighborhood in compensation for the first time, remove the impact of blind element by mask vector, mask toIn amount, " 1 " represents normal pixel, and " 0 " represents blind element, the following formula of substitution:
In formula,For blind element y0Pixel estimated value; M is mask vector; KhFor the standard deviation gaussian kernel function that is 4;
If not compensation for the first time, the direct following formula of substitution:
3) by continuous iterationCan carry out non-parametric estmation repeatedly; The present embodiment iterations is 3 times;
3, by parameter model interpolationObtain high-definition pictureSpecific as follows:
1) according to fromIn the known pixels collection { y that obtains1,…,ynAnd { x0,x1,…,xnForm X and Y, linear equation composed as follows,And adopt total least square method to solve to draw total least square solution
2) high-definition pictureSet of pixelsWithComprise known low-resolution pixel and unknown highResolution ratio pixel, equally this two groups of pixel set and above-mentioned trying to achieveCan form equation
3) solve and draw by least square methodWithThe high-resolution pixel value of middle the unknown, therebyArrive
4, utilize imageAs prior estimate, estimate to obtain by nonparametric modelIts detailed process is with step 2, only notCross and collecting { x0,x1,…,xnTime utilize
5, by parameter model interpolationObtainIts detailed process is with step 3;
6, utilize imageAs prior estimate, estimate to obtain by nonparametric modelIts detailed process is with step 4, finalObtain exporting Recovery image
The above, be only the detailed description of the invention in the present invention, but protection scope of the present invention is not limited to this, any ripeKnow the people of this technology in the disclosed technical scope of the present invention, can understand conversion or the replacement expected, all should be encompassed in the present inventionComprise scope within, therefore, protection scope of the present invention should be as the criterion with the protection domain of claims.