CN106960436A - A kind of remote sensing image process performance appraisal procedure - Google Patents

Abstract

A kind of remote sensing image process performance appraisal procedure, belongs to remote sensing image processing and assessment technique field.Methods described step is as follows：The image objective evaluation index system of gradation of image, texture and marginal information is set up, Processing Algorithm is assessed and process performance enhanced with definition is kept to details；With reference to the modulation transfer function method of estimation based on edge of a knife picture and natural image, overall improving performance of the Processing Algorithm to each frequency band MTF is assessed using the integral area of MTF (modulation transfer function) curve；Signal noise ratio (snr) of image SNR evaluation index and method is proposed, the integrated treatment performance that Processing Algorithm keeps enhancing and noise suppressed to details is assessed；The evaluation method of ring and aliasing effect is proposed, the process performance that Processing Algorithm suppresses to pseudomorphism is assessed.The method have the advantages that the performance for being applicable to comprehensive objective appraisal Image Restoration Algorithm is good and bad, Processing Algorithm parameter optimization and model refinement are instructed.

Description

A kind of remote sensing image process performance appraisal procedure

Technical field

The invention belongs to remote sensing image processing and assessment technique field, and in particular to a kind of remote sensing image processing Performance estimating method.

Background technology

Link, due to being influenceed by factors such as atmospheric environment, camera, Platform Vibrations, remote sensing images are imaged for optical remote sensing Degenerated in the presence of fuzzy, noise iseikonia matter, in-orbit image quality is generally lifted using image restoration treatment technology, while the system of reduction Development difficulty.Research at present on remote sensing image Processing Algorithm is a lot, and these algorithms can be obviously improved image tune Modulation trnasfer function (MTF) and definition, but the problem of still suffer from certain, while being mainly reflected in MTF liftings, in fact it could happen that make an uproar There is the problems such as grain details holding capacity is not strong while pseudomorphism, noise suppressed in sound amplification, adjacent edges, therefore for above-mentioned The image synthesis processing lifting of recovering quality problem has turned into one of important development direction that image restoration is handled.And obvious image The comprehensive assessment of processing improving performance is premise and the guarantee of the research of image synthesis processing boosting algorithm, checking and optimization, therefore Be badly in need of setting up more perfect image procossing Performance Evaluation index system and method, this method should be able to realize image MTF liftings, The image synthesises such as noise suppressed, grain details are kept, pseudomorphism suppression handle the assessment of improving performance, so as to assess more fully hereinafter The process performance of restoration algorithm, while instructing the optimization design of algorithm, also may be directly applied to image quality evaluation, instruct simultaneously The optimization design of optical remote sensing imaging system.

The content of the invention

The purpose of the present invention is to assess and instruct algorithm design for remote sensing image facture process performance with changing Enter the demand of optimization, propose a kind of remote sensing image process performance appraisal procedure.

The present invention considers phenogram as the objective parameter of every picture quality of quality good or not sets up evaluation index system With method, in index system in addition to comprising image modulation transmission function (MTF) and signal noise ratio (snr) of image (SNR), also comprising figure As details is kept and definition enhancing, noise suppressed, the objective evaluation index of pseudomorphism rejection ability, while in view of practical application Problem proposes the appraisal procedure of indices, compared with the objective evaluation index of traditional image and its recovering quality, Neng Gougeng Plus comprehensively, objectively reflect integrated treatment improving performance of the Image Restoration Algorithm to picture quality, while more efficiently instructing The optimization design of Processing Algorithm and optical remote sensing imaging system.

To achieve the above object, the technical scheme that the present invention takes is as follows：

A kind of remote sensing image process performance appraisal procedure, methods described step is as follows：

Step one：The image objective evaluation index system of gradation of image, texture and marginal information is set up, Processing Algorithm is assessed Process performance enhanced with definition is kept to details；

Step 2：With reference to the modulation transfer function method of estimation based on edge of a knife picture and natural image, MTF curve is utilized Integral area assess Processing Algorithm to each frequency band MTF overall improving performance；

Step 3：Propose signal noise ratio (snr) of image SNR evaluation index and method, assess Processing Algorithm details is kept enhancing and The integrated treatment performance of noise suppressed；

Step 4：The evaluation method of ring and aliasing effect is proposed, the process performance that Processing Algorithm suppresses to pseudomorphism is assessed.

The present invention is relative to the beneficial effect of prior art：

(1) compared with traditional method, this method can realize image MTF liftings, noise suppressed, grain details keep, The image synthesises such as pseudomorphism suppression handle the comprehensive assessment of improving performance, so that reflection Processing Algorithm is excellent more completely, exactly It is bad, the optimization design of algorithm is and guided, image quality evaluation is also may be directly applied to, while guiding opticses remote-sensing imaging system Optimization design；

(2) consider not including in the in-orbit image of Optical remote satellite the edge of a knife as characteristic feature scenery such as, bridges The actual application problem that MTF is assessed, the present invention is using the MTF methods of estimation combined based on edge of a knife picture and based on any natural image MTF evaluations are carried out, while lifting energy of the Processing Algorithm to each frequency band of scenery can be reflected more fully hereinafter in view of MTF areas Power, using the area of MTF curved surfaces institute envelope as evaluation index, when in image comprising the edge of a knife as when be based on using internationally recognized The MTF extracting methods of edge of a knife picture, otherwise using the MTF methods of estimation based on natural image, by comparing image MTF after before processing Area, assesses improving performance of the Processing Algorithm to different frequency bands MTF, so as to ensure that integrality and engineering that image MTF is assessed Feasibility；

(3) consider to suppress noise, while it is to investigate the good and bad important finger of Processing Algorithm to keep image texture details ability Mark, and be only capable of reflecting rejection of the Processing Algorithm to noise using traditional signal noise ratio (snr) of image evaluation index, the present invention is proposed The method of joint image flat site and non-planar regions (edge, texture relatively enrich region) comprehensive assessment signal noise ratio (snr) of image, puts down The signal noise ratio (snr) of image that smooth region is defined can reflect process performance of the Processing Algorithm to noise suppressed, the figure that non-planar regions are defined The ability of grain details is kept while can reflecting Processing Algorithm to noise suppressed as signal to noise ratio；

(4) image artifacts that Processing Algorithm is caused include ringing effect and frequency alias, influence the subjective interpretation matter of image Amount, therefore lifting image MTF, while the ability for suppressing pseudomorphism is to investigate the good and bad important indicator of Processing Algorithm, the present invention is directed to The adjacent edges of image are likely to occur the vibration texture problem of class Gibbs Distribution after processing, based on edge response function, according to It assesses the ring rejection of Processing Algorithm through the overshoot height at edge, while high frequency blocks caused frequency when being directed to imaging Rate is aliasing in processing procedure the problem of may being exaggerated, mixed than inhibition of the assessment algorithm to aliasing effect based on image letter Can, so as to effectively assess the pseudomorphism rejections of Processing Algorithm lifting image MTF simultaneously.

To sum up, the invention provides a kind of remote sensing image process performance appraisal procedure.Consider restoration algorithm pair Factor in terms of image detail holding and definition enhancing performance, pseudomorphism rejection ability and MTF and SNR winding levels, is applicable It is good and bad in the performance of comprehensive objective appraisal Image Restoration Algorithm, instruct Processing Algorithm parameter optimization and model refinement.

Brief description of the drawings

Fig. 1 is a kind of remote sensing image process performance evaluation system figure.

Embodiment

Technical scheme is further described with reference to embodiment and accompanying drawing, but is not limited thereto, It is every technical solution of the present invention to be modified or equivalent, without departing from the spirit and scope of technical solution of the present invention, It all should cover in protection scope of the present invention.

Embodiment one：As shown in figure 1, present embodiment record is that a kind of remote sensing image process performance is commented Estimate method, methods described step is as follows：

Step one：The image objective evaluation index system of gradation of image, texture and marginal information is set up, Processing Algorithm is assessed Process performance enhanced with definition is kept to details；

Step 2：With reference to the modulation transfer function method of estimation based on edge of a knife picture and natural image, MTF (modulation is utilized Transmission function) integral area of curve assesses Processing Algorithm to each frequency band MTF overall improving performance；

Step 3：Propose signal noise ratio (snr) of image SNR evaluation index and method, assess Processing Algorithm details is kept enhancing and The integrated treatment performance of noise suppressed；

Step 4：The evaluation method of ring and aliasing effect is proposed, the process performance that Processing Algorithm suppresses to pseudomorphism is assessed.

Embodiment two：A kind of remote sensing image process performance appraisal procedure described in embodiment one, The step one is comprised the following steps that：

The details retentivity of remote sensing image Processing Algorithm is assessed according to gradation of image level, texture and marginal information Performance can be strengthened with definition, mainly evaluated using the characterization parameter of gradation of image, texture and marginal information, according to image The objective parameter representative technique study of image of restoration algorithm, the image components of texture and edge details can be characterized by drawing, described Image components mainly include：Grey-scale contrast, definition, spatial frequency, texture entropy；

(1) grey-scale contrast

Grey-scale contrast refers to most bright part and most the ratio between dark-part density or logarithmic difference in image, is the different atural objects of reflection Between contrast and image clearly degree important indicator, the rill of texture is deeper in the picture, and crestal line is more prominent, and contrast is got over Greatly, image visual effect is better；Its expression formula is：

Wherein, f₁Represent grey-scale contrast；L represents the gray level of image；I, j represent the gray value of two pixels respectively；P (i, j) represents gray level co-occurrence matrixes；It is P (i, j) normalized value；

It can be seen that f from formula (1)₁Significantly amplification gray scale larger gray scale pair can be differed, so that texture is thin Clearly picture contrast is larger for section；When Same Scene image comparison, the picture contrast that Texture Boundaries tend to be fuzzy is smaller, The more visible picture contrast of Texture Boundaries is larger；It can be seen that, contrast can reflect detailed information before and after image enhancement processing Situation of change；

(2) definition

Definition refers to the obvious degree between the adjacent image that human eye can be felt；Clearly image is included than blurred picture More details and information content, this is the foundation that image quality evaluation is carried out using definition；It is evaluation image scenery thin portion The index of ability to express；Definition is high, and the effect of image is all right, and gray scale is sharper with the change of position, and image detail change is fast, The distinguishable degree at the edge of image is high；It can be seen that, definition can directly reflect that image enhancement processing algorithm details keeps and strengthened Performance；The numerical value of definition is in itself a kind of index for being compared to each other, computational methods are as follows without absolute sense：By Individual every in image takes 8 neighborhood points to subtract each other therewith, first seeks the weighted sum of 8 differences, and the size of power depends on distance, then will There is an income value to be added divided by pixel total number；Its expression formula is：

In formula (2), f₂Represent definition；w_αFor weight factor；(x, y) is pixel position coordinates in the picture；Dp tables Show that two pixel gray levels are poor；Dl represents pixel spacing；| | it is absolute value sign；M, n represent picturedeep, columns respectively；α is Eight neighborhood points numbering of pixel；

(3) spatial frequency

In original image, adjacent pixel gray value is obtained into difference image g as difference along row, column direction respectively₁(x,y)、g₂ (x, y), then the expression formula of spatial frequency be：

In formula (3), f₃Representation space frequency；

(4) texture entropy

Texture entropy is the measurement for the information content that image has, and can preferably reflect the number of information expressed by image, It determines the uncertainty degree of image, is mainly used to weigh the randomness of image；When the texture of image is extremely inconsistent, gray scale is common The value of each element will be less than normal in raw matrix GLCM, and this will mean that the texture has the complexity of larger entropy, i.e. texture higher Mean that amount of image information is bigger, its entropy is also bigger；It can be seen that, texture entropy can reflect image information before and after image enhancement processing Situation of change；Its expression formula is：

Wherein, f₄Represent texture entropy, when all P (i, j) are equal, f₄Take maximum；Ln () refers to natural logrithm computing； Because P (i, j) is the statistics based on GLCM points pair, the more traditional gray level entropy formula of its value can more reflect image texture details and The situation of edge mutation；

Finally, These parameters are added by certain weight proportion：

f₀=α₁f₁+α₂f₂+α₃f₃+α₄f₄ (5)

Wherein, f₀For details retention property and definition enhancing Evaluation results, α₁,α₂,α₃,α₄For weight factor.

Embodiment three：A kind of remote sensing image process performance appraisal procedure described in embodiment one, The step 2 is comprised the following steps that：

According to the scene features included in image, using the MTF methods of estimation based on edge of a knife picture or based on natural scene MTF methods of estimation extract the MTF of image, and then assess the MTF improving performances of remote sensing image Processing Algorithm；Work as original image Scenery in when including edge of a knife picture (sword side) feature, select MTF methods of estimation based on edge of a knife picture to extract the area of MTF curve； Otherwise, the in-orbit system MTF algorithm for estimating for being taken based on natural scene is extracted；

(1) the MTF methods of estimation based on edge of a knife picture

First look for the knife-edge point in given image, be fitted knife-edge curve, by pixel and knife-edge Distance carries out resampling, obtains edge-spread function (ESF) data distribution with sub-pixed mapping interval, afterwards the side to collecting Edge spread function data make smoothing processing and fit corresponding curve, carry out differential to the curve that fitting is obtained, just can obtain To the line spread function of system, carry out be exactly after Fourier transformation optical system one-dimensional MTF；

(2) the MTF methods of estimation based on natural scene

This method is intended to be directed to the problem of in-orbit MTF has coupling with object scene in observed image, by target Scenery and in-orbit MTF carry out parametrization sign respectively, and the remotely sensed image statistics for setting up association scenery parameter and MTF parameters characterizes mould Type, and the estimation for realizing MTF is solved to model parameter progress；

The transmission capacity of each frequency bandwidth characteristics can be characterized in view of the area of MTF curve institute envelope, the area of envelope is got over Greatly, treatment effect is better, than being evaluated more fully only with the mtf value at nyquist frequency, therefore uses MTF areas Carry out hoisting power of the evaluation image processing method to image MTF.

Embodiment four：A kind of remote sensing image process performance appraisal procedure described in embodiment one, The step 3 is comprised the following steps that：

The enhancing of suppression and image non-planar regions (edge, texture-rich) echo signal from image flat site noise Two aspects of degree assess the SNR improving performances of remote sensing image Processing Algorithm；Image SNR is carried out using two kinds of evaluation indexes The assessment of improving performance；

(1) the SNR improving performances towards image flat site noise suppressed are assessed

The SNR evaluation indexes and conventional method are consistent；Appraisal procedure is：In image to be assessed select one piece it is flat In region, the region should comprising few edge as far as possible and textural characteristics (in such region, the change master of gradation of image distribution If as caused by noise), noise level can be assessed by extracting the gradation of image variance in the region；Image in SNR indexs Signal section is then represented using the average of the area grayscale；In the assessment of noise characteristic and noise level, selected areas exists Accounting in entire image should try one's best greatly, truly to reflect the noise situations of entire image；

After image quality inverting enhancing before processing, the gray average of image flat site should be held essentially constant, because If this index is same approximate constant, show that image quality inverting enhancing algorithm has stronger rejection ability to noise；

(2) the SNR improving performances towards image non-planar regions signal enhancing are assessed

Included in remote sensing image, in image roof, road surface, waters, desert flat site and vehicle, house, Military target edge non-planar regions (region of edge and texture-rich), non-flat forms information for identification target provide it is important according to According to；Therefore, it is proposed to which a kind of SNR evaluation indexes specifically designed for image non-planar regions signal enhancing, are defined as：

Wherein,Represent the local variance of image, i.e., the gray variance calculated in pixel (i, j) neighborhood；I Represent the set of entire image coordinate points；Max () is maximum symbol；Min () is minimum value symbol；

Circular is：Piecemeal (usual window is taken as 3 × 3~7 × 7) is carried out to original image, statistics calculates figure The local variance in each region as in, finds the corresponding image-region of local variance maximum and minimum value；Obviously, local variance is maximum The variance in region characterizes the edge and texture-rich degree of image object signal, and the variance of local variance Minimum Area is illustrated Noise level；In fact, minimum variance is the noise variance measured in flat site, this with traditional noise evaluation method still It is so consistent, signal noise ratio (snr) of image can be accessed by the ratio of the two；

Because minimum variance is the noise variance that measures in flat site, maximum variance characterizes picture signal, therefore The index is that can effectively assess image quality inverting to strengthen algorithm to the enhancing degree of image object signal, again can be preferable Assess inhibition level of the enhancing algorithm to noise in ground.

Embodiment five：A kind of remote sensing image process performance appraisal procedure described in embodiment one, The step 4 is comprised the following steps that：

Although most image processing algorithm can numerically preferably approximate original image, due to using model Inaccuracy and influenceed by noise that inappropriate property is selected with Processing Algorithm parameter, all can be attached at more obvious edge Closely produce in original image and non-existent deceptive information, we are referred to as image artifacts.Image artifacts rejections is commented Estimating mainly includes：Ring suppression and the evaluation of aliasing inhibition level after remote sensing image before processing.

(1) the ring rejection of remote sensing image Processing Algorithm is assessed based on normalization edge response function

It is that the high frequency of image in a frequency domain is blocked that ring, which should be approximately considered, shows as the neighbour in gradation of image acute variation There is the vibration of class Gibbs Distribution, i.e. restored image edge ringing effect and is usually expressed as in domain：It is particularly terraced in image border The larger edge of angle value, can occur repeating vibration texture with artwork edge shape identical after recovery in the adjacent edges of image； Image ringing effect can be reflected by edge response function (ER)：

Wherein：S_net(f_x,f_y) represent whole system transmission function；u_xRepresent that the normalization of optical system on x directions is cut Only frequency；u_yRepresent the normalization cut-off frequency of optical system on y directions；D represents one horizontal pixel center response of distance Position；f_xAnd f_yX and y directions frequency is represented respectively；π is pi；ER_xAnd ER (d)_y(d) edge for referring to x and y directions respectively rings Answer function；

(2) based on the mixed aliasing rejection than assessing remote sensing image Processing Algorithm of letter

Image aliasing is due to that signal has been carried out into periodic continuation in detector sampling process, during continuation, The information of some low frequencies may replicate into radio-frequency component, cause frequency aliasing, and these false compositions will be in image procossing During be exaggerated；Image aliasing effect can be reflected by letter is mixed than (SAR)；

First, find with the strong sub-block of Energy-orientation；Content is less complicated in image but includes edge again simultaneously Sub-block is object to be found；Define local contrast C_y(m,n)：

Wherein, μ_y(m,n)、σ_y(m, n) be respectively centered on (m, n), window size for 7 × 7 sub-images average and Standard deviation, and calculate the variance of correspondence sub-block；The big explanation region of variance is complex, the figure included in the big explanation region of contrast Picture marginal information is more, and variance is small and there is the sub-block region of larger local contrast to be likely to contain substantial amounts of edge；Choosing Take and so meet after the sub-block of condition, the sub-block pixel value of the condition that is unsatisfactory for all is zeroed, then find out big big of contrast It is small be 64 × 64 sub-block；

Then the direction of each fritter is found out；The strong direction of energy comparison is there may be aliasing, and not by low frequency energy With the influence of the strong aliased portion of any medium-high frequency；Accurate direction estimation can accurately find aliased portion in frequency domain；To every One fritter is estimated with Blackman-Harris windows travel direction；Calculate the discrete Fourier transform width for 64 × 64 sub-blocks selected Value square is designated as Z^p(i, j) (wherein-p/2≤i, j ＜ p/2, p=64), and low frequency is moved into center；M_s(i, j) is used as frequency domain In from (- s ,-p/2+1) arrive (s, p/2-1) direction template, wherein variable s values be-p/2 ＜ s ＜ p/2；Ms (i, j) is wide The two-value template of ω=5；Similar establishment aliasing template A_s(i, j), A_s(i, j) is the continuation of direction template；Direction finding algorithm is such as Under：

Define primary iteration control variable k=1 and initial Q=(i, j) | 3≤| i |, | j |≤α (p/2) }；Initial E= ∑_(i,j)∈QZ^p(i,j)；Its correspondence direction energy d is calculated each s_s, as shown in formula：

If maxd_s＜ β E, make k_max=k-1, then terminates, otherwise, will make d_sMaximized s saves as s (k)；Then By M_s(i, j) and A_sThe frequency of (i, j) removes Q；K=k+1 is finally set, top-operation is repeated, through Setup Experiments α=0.4, β= 0.05；

The signal without aliasing is carried out to each sub-block to estimate；In this link, the most strong side first found out based on previous step The signal energy containing aliasing is estimated to s (k), it is M respectively to set two set^PAnd A^P, it is all k≤k respectively_maxM_s(i, J), A_sThe set of (i, j)；M^PFrequency domain include the very strong signal energy for covering aliased energy；Therefore, aliased energy is measured Frequency domain selection is included in A^PIn without in M^PIn；

The energy of isotropic is calculated, from Z^pRemoved in (i, j) and be included in M^PStrong oriented energy corresponding to middle frequency is obtainedThis constitutes estimation in A^PIn without in M^PIn frequency signal energy Low threshold, high threshold is Z^p(i,j)； Estimating signal energy It is Z^p(i's, j) is each The average value in individual direction；Finally, the ratio between estimating signal and aliasing are：

Wherein, P is the set of all sub-blocks；The size of the aliasing effect of piece image can be evaluated by SAR；But need Be pointed out that this method be according at image strong edge, what the aliasing of power spectrum performance was solved, and in the absence of The flat site of strong edge is not applied to simultaneously, and in the cities and towns scene comprising strong edge, whether complicated or simple scene, all The mixed ratio of letter can be extracted with the method.

Claims (5)

1. a kind of remote sensing image process performance appraisal procedure, it is characterised in that：Methods described step is as follows：

Step one：The image objective evaluation index system of gradation of image, texture and marginal information is set up, Processing Algorithm is assessed to thin Section keeps process performance enhanced with definition；

Step 2：With reference to the modulation transfer function method of estimation based on edge of a knife picture and natural image, the product of MTF curve is utilized Facet product assesses overall improving performance of the Processing Algorithm to each frequency band MTF；

Step 3：Signal noise ratio (snr) of image SNR evaluation index and method is proposed, Processing Algorithm is assessed and enhancing and noise is kept to details The integrated treatment performance of suppression；

Step 4：The evaluation method of ring and aliasing effect is proposed, the process performance that Processing Algorithm suppresses to pseudomorphism is assessed.

2. a kind of remote sensing image process performance appraisal procedure according to claim 1, it is characterised in that：The step One comprises the following steps that：

Assessed according to gradation of image level, texture and marginal information remote sensing image Processing Algorithm details retention property and Definition strengthens performance, is mainly evaluated using the characterization parameter of gradation of image, texture and marginal information, according to image quality The objective parameter representative technique study of the enhanced image of inverting, the image components of texture and edge details can be characterized by drawing, described Image components mainly include：Grey-scale contrast, definition, spatial frequency, texture entropy；

(1) grey-scale contrast

Grey-scale contrast refers to most bright part and most the ratio between dark-part density or logarithmic difference in image, is reflected between different atural objects Contrast and image clearly degree important indicator, the rill of texture is deeper in the picture, and crestal line is more prominent, and contrast is bigger, Image visual effect is better；Its expression formula is：

$f_1=\underset{i=0}{\overset{L-1}{\Σ}}\underset{j=0}{\overset{L-1}{\Σ}}{(i-j)}^2\hat{p}(i,j)---\left(1\right)$

Wherein, f₁Represent grey-scale contrast；L represents the gray level of image；I, j represent the gray value of two pixels respectively；P(i,j) Represent gray level co-occurrence matrixes；It is P (i, j) normalized value；

It can be seen that f from formula (1)₁Significantly amplification gray scale larger gray scale pair can be differed, so that grain details are clear Picture contrast it is larger；When Same Scene image comparison, the picture contrast that Texture Boundaries tend to be fuzzy is smaller, texture side The more visible picture contrast in boundary is larger；It can be seen that, contrast can reflect the change feelings of detailed information before and after image enhancement processing Condition；

(2) definition

Definition refers to the obvious degree between the adjacent image that human eye can be felt；Clearly image includes more than blurred picture Details and information content, this be using definition carry out image quality evaluation foundation；It is the expression of evaluation image scenery thin portion The index of ability；Definition is high, and the effect of image is all right, and gray scale is sharper with the change of position, and image detail change is fast, image Edge distinguishable degree it is high；It can be seen that, definition can directly reflect that image enhancement processing algorithm details is kept and enhanced property Energy；The numerical value of definition is in itself a kind of index for being compared to each other, computational methods are as follows without absolute sense：It is right one by one Every in image takes 8 neighborhood points to subtract each other therewith, first seeks the weighted sum of 8 differences, and the size of power depends on distance, then will be all Point income value is added divided by pixel total number；Its expression formula is：

$f_2=\frac{\underset{x=1}{\overset{m}{\Σ}}\underset{y=1}{\overset{n}{\Σ}}\underset{\mathrm{\α}=1}{\overset{8}{\Σ}}{w}_{\mathrm{\α}}|dp/dl|}{m\×n}---\left(2\right)$

In formula (2), f₂Represent definition；w_αFor weight factor；(x, y) is pixel position coordinates in the picture；M × n is represented Image size；Dp represents that two pixel gray levels are poor；Dl represents pixel spacing；| | it is absolute value sign；M, n represent image respectively Line number, columns；α numbers for eight neighborhood points of pixel；

(3) spatial frequency

In original image, adjacent pixel gray value is obtained into difference image g as difference along row, column direction respectively₁(x,y)、g₂(x, y), Then the expression formula of spatial frequency is：

$f_3=\sqrt{\frac{\underset{x=1}{\overset{m}{\Σ}}\underset{y=1}{\overset{n}{\Σ}}\[g_1^2(x,y)+g_2^2(x,y)\]}{m\×n}}---\left(3\right)$

In formula (3), f₃Representation space frequency；

(4) texture entropy

Texture entropy is the measurement for the information content that image has, and the number of information expressed by image can be preferably reflected, it determines Determine the uncertainty degree of image, be mainly used to weigh the randomness of image；When the texture of image is extremely inconsistent, gray scale symbiosis square The value of each element will be less than normal in battle array GLCM, and this will mean that the texture has the higher just meaning of the complexity of larger entropy, i.e. texture Taste that amount of image information is bigger, and its entropy is also bigger；It can be seen that, texture entropy can reflect the change of image information before and after image enhancement processing Change situation；Its expression formula is：

$f_4=-\underset{i=0}{\overset{L-1}{\Σ}}\underset{j=0}{\overset{L-1}{\Σ}}\hat{P}(i,j)ln\hat{P}(i,j)---\left(4\right)$

Wherein, f₄Represent texture entropy, when all P (i, j) are equal, f₄Take maximum；Ln () refers to natural logrithm computing；Due to P (i, j) is the statistics based on GLCM points pair, and the more traditional gray level entropy formula of its value can more reflect image texture details and edge The situation of mutation；

Finally, These parameters are added by certain weight proportion：

f₀=α₁f₁+α₂f₂+α₃f₃+α₄f₄ (5)

Wherein, f₀For details retention property and definition enhancing Evaluation results, α₁,α₂,α₃,α₄For weight factor.

3. a kind of remote sensing image process performance appraisal procedure according to claim 1, it is characterised in that：The step Two comprise the following steps that：

According to the scene features included in image, estimated using the MTF methods of estimation based on edge of a knife picture or the MTF based on natural scene Meter method extracts the MTF of image, and then assesses the MTF improving performances of remote sensing image Processing Algorithm；When the scape of original image When including the edge of a knife as feature in thing, the MTF methods of estimation based on edge of a knife picture are selected to extract the area of MTF curve；Otherwise, base is taken Extracted in the in-orbit system MTF algorithm for estimating of natural scene；

(1) the MTF methods of estimation based on edge of a knife picture

The knife-edge point in given image is first looked for, knife-edge curve is fitted, by pixel and the distance of knife-edge Resampling is carried out, the edge-spread function data distribution with sub-pixed mapping interval is obtained, afterwards the edge-diffusion letter to collecting Number data make smoothing processing and fit corresponding curve, carry out differential to the curve that fitting is obtained, just can obtain system It is exactly the one-dimensional MTF of optical system after line spread function, progress Fourier transformation；

(2) the MTF methods of estimation based on natural scene

This method is intended to be directed to the problem of in-orbit MTF has coupling with object scene in observed image, by object scene Parametrization sign is carried out respectively with in-orbit MTF, sets up the remotely sensed image statistics characterization model of association scenery parameter and MTF parameters, And model parameter is carried out to solve the estimation for realizing MTF；

The transmission capacity of each frequency bandwidth characteristics can be characterized in view of the area of MTF curve institute envelope, the area of envelope is bigger, located Manage effect better, commented than being evaluated more fully only with the mtf value at nyquist frequency, therefore using MTF areas Hoisting power of the valency image processing method to image MTF.

4. a kind of remote sensing image process performance appraisal procedure according to claim 1, it is characterised in that：The step Three comprise the following steps that：

Light is assessed in terms of the suppression of image flat site noise and the enhancing degree two of image non-planar regions echo signal Learn the SNR improving performances of remote sensing image processing algorithm；The assessment of image SNR improving performances is carried out using two kinds of evaluation indexes；

(1) the SNR improving performances towards image flat site noise suppressed are assessed

The SNR evaluation indexes and conventional method are consistent；Appraisal procedure is：One piece of flat region is selected in image to be assessed It should be extracted the gradation of image variance in the region comprising few edge and textural characteristics as far as possible and can assess and make an uproar in domain, the region Sound level；Image signal part in SNR indexs is then represented using the average of the area grayscale；In noise characteristic and noise water In flat assessment, accounting of the selected areas in entire image should try one's best greatly, truly to reflect the noise situations of entire image；

After image quality inverting enhancing before processing, the gray average of image flat site should be held essentially constant, if therefore The index is same approximate constant, then shows that image quality inverting enhancing algorithm has stronger rejection ability to noise；

(2) the SNR improving performances towards image non-planar regions signal enhancing are assessed

Roof, road surface, waters, desert flat site and vehicle, house, military affairs are included in remote sensing image, in image Object edge non-planar regions, non-flat forms information provides important evidence for identification target；Therefore, it is proposed to which one kind is specifically designed for figure As the SNR evaluation indexes of non-planar regions signal enhancing, it is defined as：

$SNR=\frac{\underset{(i,j)\∈I}{max}{\mathrm{\σ}}_{local}^2(i,j)}{\underset{(i,j)\∈I}{\min }{\mathrm{\σ}}_{local}^2(i,j)}---\left(6\right)$

Wherein,Represent the local variance of image, i.e., the gray variance calculated in pixel (i, j) neighborhood；I is represented The set of entire image coordinate points；Max () is maximum symbol；Min () is minimum value symbol；

Circular is：Piecemeal is carried out to original image, statistics calculates the local variance in each region in image, find local The corresponding image-region of variance maximum and minimum value；Obviously, the variance of local variance maximum region characterizes image object signal Edge and texture-rich degree, the variance of local variance Minimum Area illustrates noise level；In fact, minimum variance is The noise variance measured in flat site, this is still consistent with traditional noise evaluation method, is by the ratio of the two Signal noise ratio (snr) of image can be obtained.

5. a kind of remote sensing image process performance appraisal procedure according to claim 1, it is characterised in that：The step Four comprise the following steps that：

(1) the ring rejection of remote sensing image Processing Algorithm is assessed based on normalization edge response function

It is that the high frequency of image in a frequency domain is blocked that ring, which should be approximately considered, shows as going out in the neighborhood of gradation of image acute variation The vibration of existing class Gibbs Distribution, i.e. restored image edge ringing effect is usually expressed as：In image border, particularly Grad Larger edge, can occur repeating vibration texture with artwork edge shape identical after recovery in the adjacent edges of image；Image Ringing effect can be reflected by edge response function：

${\mathrm{ER}}_x\left(d\right)=0.5+\frac{1}{\mathrm{\π}}{\∫}_0^{\left(u_x\right)}\[\frac{S_{net}(f_x,0)}{f_x}\×sin\left(2{\mathrm{\πf}}_{x}d\right)\]{\mathrm{df}}_x---\left(7\right)$

${\mathrm{ER}}_y\left(d\right)=0.5+\frac{1}{\mathrm{\π}}{\∫}_0^{\left(u_y\right)}\[\frac{S_{net}(0,f_y)}{f_y}\×sin\left(2{\mathrm{\πf}}_{y}d\right)\]{\mathrm{df}}_y---\left(8\right)$

Wherein：S_net(f_x,f_y) represent whole system transmission function；u_xRepresent the normalization cutoff frequency of optical system on x directions Rate；u_yRepresent the normalization cut-off frequency of optical system on y directions；D represents the position of one horizontal pixel center response of distance； f_xAnd f_yX and y directions frequency is represented respectively；π is pi；ER_xAnd ER (d)_y(d) the skirt response letter in x and y directions is referred to respectively Number；

(2) based on the mixed aliasing rejection than assessing remote sensing image Processing Algorithm of letter

Image aliasing is due to that signal has been carried out into periodic continuation in detector sampling process, during continuation, some The information of low frequency may replicate into radio-frequency component, cause frequency aliasing, and these false compositions will be in image processing process In be exaggerated；Image aliasing effect can be reflected by the mixed ratio of letter；

First, find with the strong sub-block of Energy-orientation；Content is less complicated in image but includes the sub-block at edge again simultaneously It is object to be found；Define local contrast C_y(m,n)：

$C_{y(m,n)}=\frac{{\mathrm{\σ}}_y(m,n)}{{\mathrm{\μ}}_y(m,n)}---\left(9\right)$

Wherein, μ_y(m,n)、σ_y(m, n) is that centered on (m, n), window size is the average and standard of 7 × 7 sub-images respectively Difference, and calculate the variance of correspondence sub-block；The big explanation region of variance is complex, the image side included in the big explanation region of contrast Edge information is more, and variance is small and there is the sub-block region of larger local contrast to be likely to contain substantial amounts of edge；Choose this Sample is met after the sub-block of condition, the sub-block pixel value for the condition that is unsatisfactory for all is zeroed, then finding out the big size of contrast is 64 × 64 sub-block；

Then the direction of each fritter is found out；By low frequency energy and appoint there may be aliasing, and not in the strong direction of energy comparison The influence of what strong aliased portion of medium-high frequency；Accurate direction estimation can accurately find aliased portion in frequency domain；To each small Block is estimated with Blackman-Harris windows travel direction；Calculate the discrete Fourier transform amplitude for 64 × 64 sub-blocks selected Square it is designated as Z^p(i, j) (wherein-p/2≤i, j ＜ p/2, p=64), and low frequency is moved into center；M_s(i, j) is as in frequency domain (s, p/2-1) direction template is arrived from (- s ,-p/2+1), wherein variable s values are-p/2 ＜ s ＜ p/2；M_s(i, j) is wide ω=5 Two-value template；Similar establishment aliasing template A_s(i, j), A_s(i, j) is the continuation of direction template；Direction finding algorithm is as follows：

Define primary iteration control variable k=1 and initial Q=(i, j) | 3≤| i |, | j |≤α (p/2) }；Initial E= ∑_(i,j)∈QZ^p(i,j)；Its correspondence direction energy d is calculated each s_s, as shown in formula：

$d_s=\underset{(i,j)\∈Q}{\Σ}{M}_s(i,j)Z^p(i,j)---\left(10\right)$

If max d_s＜ β E, make k_max=k-1, then terminates, otherwise, will make d_sMaximized s saves as s (k)；Then will M_s(i, j) and A_sThe frequency of (i, j) removes Q；K=k+1 is finally set, top-operation is repeated, through Setup Experiments α=0.4, β= 0.05；

The signal without aliasing is carried out to each sub-block to estimate；In this link, the most strong direction s first found out based on previous step (k) signal energy containing aliasing is estimated, it is M respectively to set two set^PAnd A^P, it is all k≤k respectively_maxM_s(i, j), A_sThe set of (i, j)；M^PFrequency domain include the very strong signal energy for covering aliased energy；Therefore, the frequency domain of aliased energy measurement Selection is included in A^PIn without in M^PIn；

The energy of isotropic is calculated, from Z^pRemoved in (i, j) and be included in M^PStrong oriented energy corresponding to middle frequency is obtainedThis constitutes estimation in A^PIn without in M^PIn frequency signal energy Low threshold, high threshold is Z^p(i,j)； Estimating signal energy It is Z^p(i's, j) is each The average value in individual direction；Finally, the ratio between estimating signal and aliasing are：

$SAR=10{\log }_{10}\left(\frac{{\mathrm{\Σ}}_{\{p\∈P\}}{\mathrm{\Σ}}_{i,j}{Z}_{sig}^p(i,j)}{{\mathrm{\Σ}}_{\{p\∈P\}}{\mathrm{\Σ}}_{i,j}\left(Z^p\right(i,j)-Z_{sig}^p(i,j\left)\right)}\right)---\left(11\right)$

Wherein, P is the set of all sub-blocks；The size of the aliasing effect of piece image can be evaluated by SAR.