CN102013100A - Image quality discrimination method based on remote sensing image phase correlation - Google Patents

Abstract

The invention discloses an image quality discrimination method based on remote sensing image phase correlation. The method comprises the following steps: pretreating a remote sensing image to obtain a periodic component frequency spectrum of the remote sensing image; adding N random phases in the periodic component frequency spectrum of the remote sensing image to form N remote sensing images added with the random phases; carrying out discrete inverse Fourier transform to obtain N time domain remote sensing images; calculating total variation values and phase related values of the N time domain remote sensing images peru1, peru2,..., peruN; and utilizing the image phase related values to carry out quality judgment on multiple remote sensing images. The method is a reference-free digital image quality judgment method without any reference image data and is more suitable for satellite remote sensing images. Compared with other reference-free quality judgment methods, the method provided by the invention is adaptive to various quality-reducing factors, and is not affected by ring, noise and other quality reducing factors.

Description

A kind of method for judging quality of image based on the remote sensing images phase correlation

Technical field

The present invention relates to a kind of method for judging quality of image, be specifically related to a kind of method for judging quality of image, belong to the satellite remote control field of engineering technology based on the remote sensing images phase correlation.

Background technology

Image is the important information source of human perception and machine pattern-recognition, and picture quality is to the adequacy of obtaining information and accuracy decisive role all.The characteristics that the satellite remote sensing investigation has the viewpoint height, the ken is wide, data acquisition is fast and repeat, observe continuously, and the image that obtains can directly enter the user's computer image processing system, but all can produce the image deterioration problem in processes such as the collection of remote sensing images, compression, processing, transmission, this has influenced the development of satellite remote sensing technology greatly.How satellite remote sensing images being carried out effective picture quality is significant.

The method of image quality evaluation mainly contains two kinds: (1) subjective evaluation method: contrived experiment, by the observer picture quality is estimated; (2) method for objectively evaluating: adopt algorithm that picture quality is estimated.Wherein subjective evaluation method conforms to people's subjective feeling natively, but it is time-consuming, complicated, also can be subjected to the influence of subjective factors such as observer specialty background, psychology and motivation, and not can be incorporated into and use in other algorithms, this makes that subjective testing can not carry out under many circumstances.Whether the method for objectively evaluating basis possesses the original reference image, usually can divide and help reference, half reference and do not have reference three class evaluation methods, wherein full reference image quality appraisement method search time is the longest, development is comparatively ripe, is broadly divided into method of analyzing based on error-sensitivity and the method for analyzing based on structural similarity.

Fig. 1 is the judging quality of image framework of analyzing based on error-sensitivity, this evaluation method is on the basis of traditional evaluation algorithms such as PSNR, utilize the HVS feature that it is weighted correction, to improve the performance of evaluation algorithms, but the weighting weight of each characteristic factor, often obtain, lack the support of effective visible sensation sensor model by experience.Fig. 2 is the judging quality of image block diagram based on structural similarity, and this method thinks that illumination is independently for object structures, and lighting change is mainly derived from the brightness and contrast; So it is separated the brightness and contrast from the structural information of image, and integrated structure information is estimated picture quality.Its algorithm implementation complexity is lower, but has also masked other physiological characteristic of HVS simultaneously, and evaluation procedure is not easy to resolve.In addition, need provide the original reference image with reference to evaluation method entirely, though and half can obtain good performance through after the training of great amount of samples with reference to algorithm, need still to extract that the part statistic is used for comparison in the reference picture.And in actual applications, remote sensing satellite all can't or difficultly obtain reference picture and compares under a lot of situations, and this just requires to use the non-reference picture quality appraisement standard.With respect to full reference with partly with reference to evaluation method, nothing still is in the starting stage with reference to the research of evaluation method, its research is to estimate at the certain distortion type mostly, and mainly concentrates on blocking effect and fuzzy the grade on the coding distortion, less than the universality to the factor that degrades.

Summary of the invention

Technology of the present invention is dealt with problems and is: overcome the deficiencies in the prior art, a kind of method for judging quality of image based on the remote sensing images phase correlation is provided, adopting does not have with reference to method of discrimination, overcome that present method for judging quality of image is subject to blur, the problem of noise and ring influence, the factor that degrades is had certain universality.

Technical solution of the present invention is: a kind of method for judging quality of image based on the remote sensing images phase correlation may further comprise the steps:

(1) the remote sensing images u cycle of carrying out of a width of cloth spatial domain is decomposed with level and smooth, resolve into the periodic component that comprises most image informations and in spatial domain, change level and smooth slowly component, then with level and smooth component filtering, periodic component behind the level and smooth component of filtering is carried out the frequency spectrum that discrete Fourier transformation obtains periodic component, and the obtain manner of periodic component frequency spectrum is:

$\∀(q,r)\∈\mathrm{\Ω}\backslash \left(\mathrm{0,0}\right)$

$\widehat{\mathrm{per}\left(u\right)}(q,r)=\hat{u}(q,r)-\frac{\hat{v}(q,r)}{4-2\cos \left(\frac{2\mathrm{\πq}}{M}\right)-2\cos \left(\frac{2\mathrm{\πr}}{N}\right)}$

$\widehat{\mathrm{peru}(q,r)}=\left|\widehat{\mathrm{peru}(q,r)}\right|e^{\mathrm{i\φ}(q,r)}$

Wherein

And v ₁, v ₂Satisfy:

$\∀(x,y)\∈\mathrm{\Ω}$

$\∀(x,y)\∈\mathrm{\Ω}$

Wherein:

Q, r are remote sensing images frequency domain coordinate

Ω (0,0) for removing the real number field of true origin;

Frequency spectrum for periodic component;

(q r) is the phase place of periodic component to φ;

Frequency spectrum for remote sensing images;

v ₁, v ₂Difference for image pixel;

M is the length of remote sensing images, and unit is a pixel;

N is the width of remote sensing images, and unit is a pixel;

X, y are remote sensing images spatial domain coordinate;

(2) in the phase place of periodic component frequency spectrum, add N random phase and form the remote sensing images that the N width of cloth adds random phase; The described concrete mode that adds N random phase in the phase place of periodic component frequency spectrum is: phase is increased a random offset ε S, obtain a new phase function ψ=φ+ε S, wherein ε is a fixed value, S is at (π, satisfy the equally distributed stochastic variable of independent same distribution condition π), wherein N is a natural number;

(3) image that the N width of cloth that generates in the step (2) is added random phase carries out inverse discrete Fourier transform, obtains N width of cloth time domain remote sensing images peru ₁, peru ₂..., peru _N

(4) calculate N width of cloth time domain remote sensing images peru ₁, peru ₂..., peru _NThe total variation value, the computing method of wherein every width of cloth time domain remote sensing images total variation value are:

$\mathrm{TV}\left(u_N\right)=\underset{x,y\∈\mathrm{\Ω}}{\mathrm{\Σ}}|\▿{\mathrm{peru}}_N(x,y)|$

(5) N width of cloth time domain remote sensing images total variation value TV (u in the calculation procedure (4) _N) average μ and variances sigma, utilize N width of cloth time domain remote sensing images total variation value TV (u _N) average μ and variances sigma calculate N width of cloth time domain remote sensing images peru ₁, peru ₂..., peru _NPhase place correlation IPC (u) promptly obtain the phase place correlation of these width of cloth spatial domain remote sensing images u;

$\mathrm{IPC}\left(u\right)=-\mathrm{lo}{g}_{10}\mathrm{\Φ}\left(\frac{\mathrm{\μ}-\mathrm{TV}\left(u_N\right)}{\mathrm{\σ}}\right)$

Wherein Φ is a normal distyribution function, promptly

(6) utilize step (1) to calculate the phase place correlation IPC (u) of the different width of cloth remote sensing images of Same Scene to the computing method of step (5), phase place correlation IPC (u) to the different width of cloth remote sensing images of Same Scene compares, utilize the size of remote sensing images phase place correlation IPC (u) value that Remote Sensing Image Quality is differentiated, method of discrimination is that the big more Remote Sensing Image Quality of remote sensing images phase place correlation IPC (u) value is good more.

The present invention's beneficial effect compared with prior art is: utilization of the present invention is based on the method computed image acutance of remote sensing images phase correlation, as a kind of new non-reference picture quality discrimination standard, without any need for reference image data, overcome in the present most judging quality of image based on entropy and gradient, blured easily, the problem of noise and ring influence, with respect to existing non-reference picture quality appraisement method, the present invention has more universality to the factor that degrades, not only can be used as an effectively fuzzy indicator and noise indicator, also can be applied to aliasing detection and ring and detect.

Description of drawings

Fig. 1 is the judging quality of image block diagram of analyzing based on error-sensitivity;

Fig. 2 is the judging quality of image block diagram based on structural similarity;

Fig. 3 is a differentiation process flow diagram of the present invention;

Fig. 4 is the remote sensing images image of two out of phase correlations of a certain scene correspondence;

Fig. 4 (a) is the remote sensing images image of phase place correlation 4360.31 correspondences;

Fig. 4 (b) is the remote sensing images image of phase place correlation 607.256 correspondences;

Fig. 5 is the remote sensing images image of two out of phase correlations of another scene correspondence;

Fig. 5 (a) is the remote sensing images image of phase place correlation 318.252 correspondences;

Fig. 5 (b) is the remote sensing images image of phase place correlation 294.126 correspondences.

Embodiment

The present invention is further illustrated below in conjunction with concrete embodiment.

As shown in Figure 3, the present invention includes following steps:

(1) at first remote sensing images is carried out pre-service

The satellite remote sensing images of one width of cloth spatial domain is carried out discretize can produce very strong artificial interference on the frequency domain after the Fourier transform, this is that the hypothesis image is continuous because of discrete Fourier transformation, but the real image border has uncontinuity usually.This gap can make the quality assessment effect of utilizing image phase deviation occur, so we need decompose with level and smooth each u cycle of carrying out, remote sensing images are decomposed into and will not influenced by boundary effect, comprise the periodic component of most image informations and in the spatial domain, change level and smooth slowly component, and the level and smooth component of filtering.

Concrete grammar is: at first the remote sensing images u cycle of carrying out of a width of cloth spatial domain is decomposed with level and smooth, resolve into the periodic component that comprises most image informations and in spatial domain, change level and smooth slowly component, then with level and smooth component filtering, periodic component behind the level and smooth component of filtering is carried out the frequency spectrum that discrete Fourier transformation obtains periodic component, and the obtain manner of periodic component frequency spectrum is:

$\∀(q,r)\∈\mathrm{\Ω}\backslash \left(\mathrm{0,0}\right)$

$\widehat{\mathrm{per}\left(u\right)}(q,r)=\hat{u}(q,r)-\frac{\hat{v}(q,r)}{4-2\cos \left(\frac{2\mathrm{\πq}}{M}\right)-2\cos \left(\frac{2\mathrm{\πr}}{N}\right)}$

Wherein

And v ₁, v ₂Satisfy:

$\∀(x,y)\∈\mathrm{\Ω}$

$\∀(x,y)\∈\mathrm{\Ω}$

Wherein:

Q, r are remote sensing images frequency domain coordinate

Ω (0,0) for removing the real number field of true origin;

Frequency spectrum for periodic component;

(q r) is the phase place of periodic component to φ;

Frequency spectrum for remote sensing images;

v ₁, v ₂Difference for image pixel;

M is the length of remote sensing images, and unit is a pixel;

N is the width of remote sensing images, and unit is a pixel;

X, y are remote sensing images spatial domain coordinate;

So just, can obtain the frequency domain distribution of piece image

$\widehat{\mathrm{peru}(q,r)}=\left|\widehat{\mathrm{peru}(q,r)}\right|e^{\mathrm{i\φ}(q,r)}$

Here we need to prove, the phase information of image all is included in the periodic component, and level and smooth component does not have phase information, so even if we have removed level and smooth component, the processing of next image phase being carried out can be do not had influence on yet, the precision of image discrete Fourier transformation can be improved on the contrary;

(2) number of times of supposing Monte-Carlo Simulation is N, then needs N such calculating of repetition, generates the image that the N width of cloth adds random phase.

In the phase place of periodic component frequency spectrum, add N random phase and form the remote sensing images that the N width of cloth adds random phase; The described concrete mode that adds N random phase in the phase place of periodic component frequency spectrum is: phase is increased a random offset ε S, obtain a new phase function ψ=φ+ε S, wherein ε is a fixed value, S is at (π, satisfy the equally distributed stochastic variable of independent same distribution condition π), wherein N is a natural number;

(3) image that the N width of cloth that generates in the step (2) is added random phase carries out inverse discrete Fourier transform, obtains N width of cloth time domain remote sensing images peru ₁, peru ₂..., peru _N

(4) calculate N width of cloth time domain remote sensing images peru ₁, peru ₂..., peru _NThe total variation value, the computing method of wherein every width of cloth time domain remote sensing images total variation value are: $\mathrm{TV}\left(u_N\right)=\underset{x,y\∈\mathrm{\Ω}}{\mathrm{\Σ}}|\▿{\mathrm{peru}}_N(x,y)|$

(5) N width of cloth time domain remote sensing images total variation value TV (u in the calculation procedure (4) _N) average μ and variances sigma, utilize N width of cloth time domain remote sensing images total variation value TV (u _N) average μ and variances sigma calculate N width of cloth time domain remote sensing images peru ₁, peru ₂..., peru _NPhase place correlation IPC (u),

$\mathrm{IPC}\left(u\right)=-\mathrm{lo}{g}_{10}\mathrm{\Φ}\left(\frac{\mathrm{\μ}-\mathrm{TV}\left(u_N\right)}{\mathrm{\σ}}\right)$

Wherein Φ is a normal distyribution function, promptly

The formula applicable elements is: N random phase S ₁..., S _NIndependent same distribution, then TV (u ₁) ..., TV (u _N) independent same distribution, TV (u) meets Gaussian distribution and has limited mathematical expectation and variance.

(6) utilize step (1) to calculate the phase place correlation IPC (u) of several remote sensing images of Same Scene to the computing method of step (5), phase place correlation IPC (u) to several remote sensing images of Same Scene compares, and wherein the big more picture quality of remote sensing images phase place correlation IPC (u) value is good more.

It should be noted that, the IPC value does not have absolute sense, the single IPC value of utilizing that can not be simple is differentiated the quality of a width of cloth Remote Sensing Image Quality, has only when obtaining several groups of remote sensing images IPC values of Same Scene different conditions, just can carry out picture quality according to the size of these values and pass judgment on.

Handle according to above six steps in sequence, can finish quality discrimination several remote sensing images.Shown in Fig. 4,5, the remote sensing images of two out of phase correlations of each scene of the two groups of scenes correspondence that obtains according to step of the present invention (1)-(5), Fig. 4 (a) is the remote sensing images image of a certain scene phase place correlation 4360.31 correspondences; Fig. 4 (b) is the remote sensing images image of a certain scene phase place correlation 607.256 correspondences; Fig. 5 (a) is the remote sensing images image of another scene phase place correlation 318.252 correspondences; Fig. 5 (b) is the remote sensing images image of another scene phase place correlation 294.126 correspondences; Compare as seen by above-mentioned two group of four width of cloth image, the remote sensing images good imaging quality that the phase place correlation is big, the remote sensing images image quality that the phase place correlation is little is poor.

The present invention not detailed description is a technology as well known to those skilled in the art.

Claims (1)

1. method for judging quality of image based on the remote sensing images phase correlation is characterized in that may further comprise the steps:

(1) the remote sensing images u cycle of carrying out of a width of cloth spatial domain is decomposed with level and smooth, resolve into the periodic component that comprises most image informations and in spatial domain, change level and smooth slowly component, then with level and smooth component filtering, periodic component behind the level and smooth component of filtering is carried out the frequency spectrum that discrete Fourier transformation obtains periodic component, and the obtain manner of periodic component frequency spectrum is:

$\∀(q,r)\∈\mathrm{\Ω}\backslash \left(\mathrm{0,0}\right)$

$\widehat{\mathrm{per}\left(u\right)}(q,r)=\hat{u}(q,r)-\frac{\hat{v}(q,r)}{4-2\cos \left(\frac{2\mathrm{\πq}}{M}\right)-2\cos \left(\frac{2\mathrm{\πr}}{N}\right)}$

$\widehat{\mathrm{peru}(q,r)}=\left|\widehat{\mathrm{peru}(q,r)}\right|e^{\mathrm{i\φ}(q,r)}$

Wherein

And v ₁, v ₂Satisfy:

$\∀(x,y)\∈\mathrm{\Ω}$

$\∀(x,y)\∈\mathrm{\Ω}$

Wherein:

Q, r are remote sensing images frequency domain coordinate;

Ω (0,0) for removing the real number field of true origin;

Frequency spectrum for periodic component;

(q r) is the phase place of periodic component to φ;

Frequency spectrum for remote sensing images;

v ₁, v ₂Difference for image pixel;

M is the length of remote sensing images, and unit is a pixel;

N is the width of remote sensing images, and unit is a pixel;

X, y are remote sensing images spatial domain coordinate;

(2) in the phase place of periodic component frequency spectrum, add N random phase and form the remote sensing images that the N width of cloth adds random phase; The described concrete mode that adds N random phase in the phase place of periodic component frequency spectrum is: phase is increased a random offset ε S, obtain a new phase function ψ=φ+ε S, wherein ε is a fixed value, S is at (π, satisfy the equally distributed stochastic variable of independent same distribution condition π), wherein N is a natural number;

(3) image that the N width of cloth that generates in the step (2) is added random phase carries out inverse discrete Fourier transform, obtains N width of cloth time domain remote sensing images peru ₁, peru ₂..., peru _N

(4) calculate N width of cloth time domain remote sensing images peru ₁, peru ₂..., peru _NThe total variation value, the computing method of wherein every width of cloth time domain remote sensing images total variation value are:

(5) N width of cloth time domain remote sensing images total variation value TV (u in the calculation procedure (4) _N) average μ and variances sigma, utilize N width of cloth time domain remote sensing images total variation value TV (u _N) average μ and variances sigma calculate N width of cloth time domain remote sensing images peru ₁, peru ₂..., peru _NPhase place correlation IPC (u) promptly obtain the phase place correlation of these width of cloth spatial domain remote sensing images u; $\mathrm{IPC}\left(u\right)=-{\log }_{10}\mathrm{\Φ}\left(\frac{\mathrm{\μ}-\mathrm{TV}\left(u_N\right)}{\mathrm{\σ}}\right)$

Wherein Φ is a normal distyribution function, promptly

(6) utilize step (1) to calculate the phase place correlation IPC (u) of the different width of cloth remote sensing images of Same Scene to the computing method of step (5), phase place correlation IPC (u) to the different width of cloth remote sensing images of Same Scene compares, utilize the size of remote sensing images phase place correlation IPC (u) value that Remote Sensing Image Quality is differentiated, method of discrimination is that the big more Remote Sensing Image Quality of remote sensing images phase place correlation IPC (u) value is good more.

Images