CN1484193A - Multi-frame fuzzy detection method for point object in low s/n ratio infrared image sequence - Google Patents

Abstract

The invention discloses a kind of multi-frames fuzzy examination method for point target in low signal-noise ratio infrared picture sequence, it gains picture background through mathematics morphology filter algorithm, subtracts the background from the source picture, obtains de-average picture including target and noise, estimates the noise average value and the variance of local part through several frames. Carries on normalized processing to noise picture through the noise average value and the variance, uses noise probability distribution function as single frame fuzzy examination subordination function, retains as far as possible goals amplitude information, uses the velocity filter group method to accumulate the target energy, enhances the signal to noise ratio, gets the distributed characteristic curve with or without target, according to the examination performance which is assigned, uses the permanent false alarm rate criterion, affirms the target threshold value, thus makes the decision whether the target exists or not, finally simplifies the computation of examination threshold with central limit theorem of probability theory.

Description

The fuzzy detection method of the multiframe of point target in the low signal-to-noise ratio infrared image sequence

Technical field:

The present invention relates to the fuzzy detection method of multiframe of point target in a kind of low signal-to-noise ratio infrared image sequence, be a core technology of Infra-Red Search ﹠ Track System, precise guidance system, infraed early warning system, big visual field targeted surveillance system, satellite remote sensing system, safety check system etc., in all kinds of military, civilian systems, all can be widely used.

Background technology:

Precision guided weapon is to adopt detected with high accuracy, control and guidance technology, detection from complex background, tracking and identification effectively, select target and high precision hit critical position, the weaponry of final prang.Imaging detection can intuitively obtain target informations such as abundant target shape or basic structure, suppresses background interference, the critical position of recognition objective and target, thereby become the important development direction of precision guided weapon.

The modernization advanced technology weapon requires the remote target of finding in order to increase operational distance.Have only and find target, tracking target in time, catch and lock onto target in time, could realize effectively attacking.Infrared imaging sensor is because its good concealment, imaging resolution height, identification camouflage ability is strong and the ability that can work under rugged surroundings such as night, makes it become photoelectric sensor indispensable in the modern war.

Research for the detection problem of the little target of low signal-to-noise ratio in the infrared image originates from Reconnaissance, Long Range and supervision.When target range was far away, target showed as point target in infrared image, can't utilize features such as shape, size, texture to detect target.In this stage, signal noise ratio (snr) of image is lower, promptly relative background of the amplitude of echo signal and noise a little less than, flooded by noise basically.Single frames detects may produce a lot of false targets, can't obtain desired detection probability and false-alarm probability.Study the detection algorithm of the medium and small target of low signal-to-noise ratio infrared image, can realize expanding the purpose of their operating distance, for increasing operational distance and increasing the reaction time, the survival probability that improves one's own side has great importance.

Though the detection of Weak target has had the history of nearly half a century with the research of following the tracks of in radar and sonar signal processing, the detection of Weak target then is the extensive attention that just obtained the researcher before and after 1980 with following the tracks of in the image.For over ten years, the detection problem of the little target of low signal-to-noise ratio is the research focus in optics and infrared image field always in the image.

Can't guarantee to detect performance because single-frame images is handled, and target travel has regularity and continuity, therefore handle, utilize temporal information further to confirm real goal, become effective solution route by the sequence image multiframe.

Multiframe target detection problem can be expressed as: whether the existence of detection target in given 3-D view space, and the track of estimating target.The multiframe algorithm of target detection can be divided into two classes basically: a class is for detecting the back track algorithm earlier, it will detect with tracking and be divided into two independently problems, at first every two field picture is all made the judgement whether target exists, in track algorithm, observed reading and flight path are carried out relevant, realize initial, the affirmation of flight path of targetpath and the termination of flight path.The algorithm of this detection earlier back tracking is simple, but requires the signal to noise ratio (S/N ratio) of target higher.Another kind of is to follow the tracks of the back detection algorithm earlier, in 3-D view, a large amount of possible tracks is followed the tracks of, but originally whether the authentic representative target is not made judgement to these tracks, but the track of every tracking is calculated its posteriority probability function, if the posterior probability function of certain bar track surpasses a certain thresholding, just think that this track represents a target.

The detection of Weak target and tracking mainly contain three-dimensional matched filtering algorithm, multistage test of hypothesis algorithm, dynamic programming method etc. in the low SNR images sequence, they are suitable for the environment of low signal-to-noise ratio, but all hypothetical target and noise are obeyed independent identically distributed Gaussian distribution respectively, and the detection algorithm of having derived on this basis, analyzed the performance of algorithm.Yet the unevenness of the temperature of air, humidity, pressure and density can cause atmospheric turbulence, and when infrared radiation is propagated by turbulent atmosphere, beam intensity, frequency, phase place, degree of polarization and the angle of arrival will be subjected to the influence of turbulent flow, produces jitter phenomenon.In addition, a kind of random motion of atmospheric molecule group in three dimensions also can cause scintillation, and these all can make the picture element variation of actual infrared image.Therefore the not necessarily always independent identically distributed Gaussian distribution of actual infrared image noise, so the resulting result of above-mentioned algorithm can not guarantee optimum in actual applications.

The research of small point target detection problem is that a difficulty is bigger in the low signal-to-noise ratio infrared image, and the problem of significant application value is arranged.Because single-frame images is handled the reliable detection that can not realize point target; must adopt the sequence image disposal route of based target motion feature, frontier inspection is surveyed the limit and is followed the tracks of, and manages along targetpath accumulation flight path energy; improve signal to noise ratio (S/N ratio), to reach the purpose that suppresses noise, track.Present research is primarily aimed at the method that how to improve the performance of detection and track algorithm under the low signal-to-noise ratio complex background condition.

Summary of the invention:

The objective of the invention is at the deficiencies in the prior art, a kind of new method that can detect small point target in actual low signal-to-noise ratio infrared image sequence is proposed, solve the uncertain problem of noisiness in the actual infrared image sequence, avoid the slow-action track rejection, improve signal to noise ratio (S/N ratio), improve processing speed.

For realizing such purpose, the present invention will blur to detect and combine with velocity filtter, propose the fuzzy detection method of multiframe of point target in a kind of low signal-to-noise ratio infrared image sequence.At first obtain image background by the mathematical morphology filter algorithm, source images is deducted the image background of obtaining, what obtain containing target and noise removes the average image, by noise average and the variance that the front and back multiple image comes each regional area of estimated image, utilize noise average and the variance estimated that noise image is carried out normalized then.Fuzzy theory is applied in the low SNR images target detection,, makes single frames detect and be output as one in the equally distributed stochastic variable of [0,1] interval obedience with the fuzzy membership function that detects of noise probability distribution function conduct.Adopt the velocity filtter group of methods that the fuzzy testing result of single frames may be added up by trajectory direction at each then, analytic derivation the probability density function of the statistic that adds up, and utilize the constant false alarm rate criterion to detect target.Utilize the central limit theorem in the theory of probability to simplify the calculating that multiframe is blured detection threshold in the detection design process at last.

Method of the present invention comprises following concrete steps:

1. infrared image goes average to handle:

Adopt the mathematical morphology filter algorithm to obtain the background of image, its principle is with complete zero flat-top shape structural element infrared image to be opened and then closed procedure, and all kinds of bright, dark noises and little target obtain image background in the removal image.Remove the average image with what source images and the background subtracting that obtains obtained comprising Weak target and noise.

2. picture noise normalization:

To go the average image division to be local wicket one by one, multiple image comes the noisiness in each local window of estimated image before and after utilizing, and the noise average and the standard deviation of going the average image to estimate according to each wicket are carried out the noise normalized.

3. single frames is fuzzy detects:

Utilize fuzzy theory to be applicable to the characteristics that accurately to know the mathematical model of controlled plant occasion, solve the uncertain problem of noisiness in the actual infrared image sequence, keep target amplitude information as much as possible by the detection of fuzzy membership function single frames, postponed hard judgement time.The present invention adopts the fuzzy membership function that detects of noise probability distribution function conduct, has kept target amplitude information as much as possible, and makes the single frames detection be output as one in the equally distributed stochastic variable of [0,1] interval obedience.

4. the velocity filtter group is handled:

Movement velocity size and the direction possible according to target, off-line is constructed one group of velocity filtter, corresponding to all possible target trajectory, and then the energy that on all possible target trajectorys, adds up.

5. constant false alarm rate target detection:

Through the energy accumulation of velocity filtter group, can obtain, the distribution character curve of statistic during driftlessness, according to given detection performance, utilize the constant false alarm rate criterion, determine the threshold value of target detection.If the statistic of certain point judges then that greater than detection threshold there is target in this point; Otherwise then driftlessness exists.

6. design process is simplified:

Above-mentioned five steps have been finished the fuzzy testing process of multiframe, and the detection threshold that this step utilizes the central limit theorem in the theory of probability to simplify step 5 in the fuzzy detection method of multiframe calculates.When number of image frames was big, the fuzzy detection algorithm of multiframe was utilizing the constant false alarm rate criterion to carry out threshold value when determining, can directly utilize normal distyribution function definite threshold size of tabling look-up.

The present invention adopts mathematical morphology filter to obtain image background, and its algorithm can Parallel Implementation, has improved processing speed greatly.Source images and background image subtracted each other obtain noise image, the shortcoming of the slow-action track rejection that may cause when having avoided the image difference computing to obtain noise image.Utilize fuzzy theory to be applicable to that the characteristics that can't accurately know the mathematical model of controlled plant occasion solve the uncertain problem of noisiness in the actual infrared image sequence.Keep target amplitude information as much as possible by the detection of fuzzy membership function single frames, postponed hard judgement time.Adopt the velocity filtter group of methods to accumulate target energy, improve signal to noise ratio (S/N ratio), thereby improve the performance of algorithm of target detection.Utilize the central limit theorem in the theory of probability to simplify the calculating that multiframe is blured the detection threshold of detection algorithm.The present invention can detect the point target in the low SNR images sequence, can be widely used in all kinds of military, civilian systems, has vast market prospect and using value.

Description of drawings:

Fig. 1 is a disposal route The general frame of the present invention.

As shown in Figure 1, infrared image sequence at first passes through the image pre-service, comprises image and goes average to handle and picture noise normalization, and result is the normalization noise image that comprises Weak target.Adopt the noise probability distribution function as the fuzzy membership function that detects of single frames, make the single frames testing result be one and obey equally distributed stochastic variable in that [0,1] is interval.Adopt the velocity filtter group of methods that the fuzzy testing result of single frames may be added up by trajectory direction at each then, analytic derivation the probability density function of the statistic that adds up, and utilize the constant false alarm rate criterion to detect target.

Fig. 2 is the pretreated detailed step of infrared image sequence.

As shown in Figure 2, the pre-service of infrared image sequence comprises that infrared image goes average to handle and picture noise normalization.Adopt the mathematical morphology filter algorithm to obtain the background of image, remove the average image with what source images and the background subtracting that obtains obtained comprising Weak target and noise, to going the average image to carry out noisiness analysis and normalized, obtain the approximate picture noise that satisfies standardized normal distribution again.

Fig. 3 is for to become each local wicket with the infrared average image division of going.

Fig. 4 is the fuzzy membership function that detects of typical single frames.

Fig. 5 was respectively 2,3,4,5 o'clock for the image sequence frame number, and the fuzzy testing result of single frames is along the probability density function curve of possibility target trajectory energy accumulation statistic.

Fig. 6 is that the multiframe of utilizing the present invention to propose is blured detection method is carried out Point Target Detection to the actual infrared sequence image of 12 frames result.

Wherein, Fig. 6 (a) is a frame of infrared image sequence, detected point target in Fig. 6 (b) 12 frame sequence images.

Fig. 7 is under different images sequence frame said conditions, resolve the uniformly distributed random variable try to achieve and probability density function and corresponding Density Function of Normal Distribution curve ratio.

Wherein, Fig. 7 (a) for the image sequence frame number be 2,3,4,5 o'clock both curve ratio, Fig. 7 (b) for the image sequence frame number be 10,20 o'clock both curve ratio.

Embodiment:

In order to understand technical scheme of the present invention better, embodiments of the present invention are further described below in conjunction with accompanying drawing.

Fig. 1 is the fuzzy detection method The general frame of multiframe of point target in a kind of low signal-to-noise ratio infrared image sequence of the present invention's proposition.The concrete implementation detail of each several part is as follows:

1. infrared image goes average to handle:

Infrared image goes average to handle shown in Fig. 2 first half.

Adopt mathematical morphology filter to obtain image background.Morphology is opened computing and is used to eliminate and compares the less bright noise of size with structural element:

f _l=fob=(f b) b (1) morphology closure operation is used to eliminate compares the less dark noise of size with structural element:

f _B=f _lB=(f _l b) b (2) and keep the integral image gray-scale value and big bright, dark areas is unaffected substantially. is respectively expansion of morphology gray scale and erosion operation in the formula, and b is a structural element, and f is the source infrared image, f ₁For opening operation result, f _BBe the image background that obtains.

Because distant object shows as little bright spot or speck in infrared image, therefore, infrared image is opened and then closed procedure with 5 * 5 complete zero flat-top shape structural elements, can remove all kinds of bright, dark noises and little target, the acquisition image background.

Source infrared image and the image background of obtaining are subtracted each other the noise image that just obtains comprising little target:

f _n=f-f _B(3) in the formula, f _nBe noise image.

2. picture noise normalization:

Picture noise normalization is shown in Fig. 2 latter half.Noise image is divided into wicket (for example 8 * 8 windows as shown in Figure 3) one by one, and multiple image (for example 10 frames) is estimated the standard deviation sigma of noise in each window before and after utilizing _Ij(k):

${\mathrm{\σ}}_{\mathrm{ij}}\left(k\right)=\sqrt{\frac{1}{N}\underset{k=1}{\overset{N}{\mathrm{\Σ}}}{\left(f_n\left(k\right)\right)}^2}---\left(4\right)$

In the formula, N is the number of pixels of the same wicket of front and back multiframe, and k is a pixel label.The standard deviation that noise image is estimated according to each wicket is carried out normalization:

$S(x_i,y_j,k)=\frac{f_n(x_i,y_j,k)}{{\mathrm{\σ}}_{\mathrm{ij}}\left(k\right)}---\left(5\right)$

3. single frames is fuzzy detects

In order to reduce the loss of hard judgement to amplitude information, obfuscation is carried out in whole judgement territory divided, judge having or not of target by the tendentiousness information of extracting observation signal, this tendentiousness information is a kind of fuzzy message, it can be characterized by membership function.

Choosing the noise probability distribution function and have a critical nature as detecting membership function, is exactly that testing result goes up evenly distribution of obedience in [0,1].The advantage of this character is exactly that make to detect performance irrelevant with concrete noise profile, and no matter noise is what to be obeyed distribute, as long as the probability distribution function of choosing this noise is as detecting membership function, and just can be so that testing result is obeyed even distribution.In practice,, can draw the empirical distribution function of noise, try to achieve the output valve of each point then by interpolation algorithm by the method for statistics if distribution character the unknown of noise can't obtain the analytical expression of its probability distribution function.Typical single frames is fuzzy to detect membership function as shown in Figure 4.

Probability distribution function to each local wicket statistical noise of normalization noise image carries out obfuscation to all the pixel amplitudes in this window then, obtains the fuzzy testing result of single frames:

T (x _i, y _j, k)=μ (S (x _i, y _j, k)) and in (6) formula, μ () is the probability distribution function of noise.

4. the velocity filtter group is handled

Movement velocity size and the direction possible according to target, off-line is constructed one group of velocity filtter, corresponding to all possible target trajectory.

In this example, choosing the movement velocity scope 0-1.5 pixel/frame of target, is a possible target speed value every 0.25 pixel/frame, and per 15 degree that are chosen as of velocity reversal are a possible target velocity direction.In application, also can choose different spacing values according to actual conditions.

With the institute of the fuzzy testing result of single frames in the velocity filtter group energy that adds up on might target trajectory:

$u=\underset{k=1}{\overset{n}{\mathrm{\Σ}}}t(x_i,y_j,k)(x_i,y_j,k)\∈S---\left(7\right)$

S is the set of all possible target trajectory in the formula, i.e. velocity filtter group, and n is the frame number of image sequence.

When target did not exist, the probability density function that statistic u obeys was:

$h_n\left(u\right)=\frac{1}{(n-1)!}[u^{n-1}-C_n^1{(u-1)}^{n-1}+C_n^2{(u-2)}^{n-1}+\·\·\·$

$+(-1{)}^{\mathrm{floor}\left(u\right)}{C}_n^{\mathrm{floor}\left(u\right)}{(u-\mathrm{floor}\left(u\right))}^{n-1}]---\left(8\right)$

In the formula, floor () function representation rounds to negative infinity.Fig. 5 was respectively the curve of statistic u probability density function at 2,3,4,5 o'clock for the image sequence frame number.

Corresponding probability distribution function is:

$H_n\left(u\right)=\underset{i=0}{\overset{\mathrm{floor}\left(u\right)}{\mathrm{\Σ}}}{(-1)}^i{C}_n^i\frac{{(u-i)}^n}{n!}---\left(9\right)$

5. constant false alarm rate target detection

Given false-alarm probability P _F, detection threshold can be obtained by through type (10):

P _F＝1-H _n(T) (10)

For example, if given detection false-alarm probability is P _F=10 ^-5, the image sequence frame number is 12 o'clock, the detection threshold that calculating can get this moment is about T=10.26.If the value along certain statistic u that may target trajectory calculates in the step 3 surpasses detection threshold T, judge that then it is a target.

Figure 6 shows that the fuzzy multiframe detection algorithm that utilizes the present invention to propose carries out the result of Point Target Detection to the actual infrared sequence image of 12 frames.The image size of actual infrared image sequence is 80 * 80, and the image pixel gray-scale value is 8, and the signal to noise ratio (S/N ratio) of image is about 2.8, and target is made approximate linear uniform motion, and speed is less than 1 pixel/frame, shown in Fig. 6 (a).Institute along step 3 medium velocity bank of filters might accumulate target energy by track, then is judged as target when adding up and surpassing detection threshold.Real image Sequence Detection result makes 0.5 pixel/frame linear uniform motion for target, shown in Fig. 6 (b).

6. design process is simplified

Above each step finished the testing process of Weak target in the infrared image sequence, next utilize central limit theorem in the theory of probability to simplify the design process of the fuzzy detection algorithm of multiframe.

For the even distribution on [0,1] interval, its average ${\mathrm{M\ξ}}_i=\frac{1}{2},$ Variance ${\mathrm{D\ξ}}_i=\frac{1}{12}.$ According to central limit theorem, when n increases, the uniformly distributed random variable on n obedience [0, the 1] interval and trend towards normal distribution gradually $N(\frac{n}{2},\sqrt{\frac{n}{12}}).$

Fig. 7 (a) be depicted as image sequence frame number n get respectively 2,3,4,5 o'clock through type (8) uniformly distributed random variables of trying to achieve and probability density function and corresponding normal distyribution function curve, Fig. 7 (b) be depicted as image sequence frame number n get respectively 10,20 o'clock through type (8) uniformly distributed random variables of trying to achieve and probability density function and corresponding normal distyribution function curve.As can be seen along with the increase of n, the uniformly distributed random variable that through type (8) is tried to achieve and probability density function more and more littler with corresponding normal distyribution function curve error.

When number of image frames was big, the fuzzy detection algorithm of multiframe was utilizing the constant false alarm rate criterion to carry out threshold value when determining, can directly utilize normal distyribution function definite threshold size of tabling look-up, and calculate without through type (10).

The present invention utilizes fuzzy theory to solve the uncertain problem of noisiness in the actual infrared image sequence.Kept target amplitude information as much as possible by the detection of fuzzy membership function single frames.Adopt the velocity filtter group of methods to accumulate target energy, improve signal to noise ratio (S/N ratio), thereby improve the performance of algorithm of target detection.Utilize the central limit theorem in the theory of probability to simplify the calculating that multiframe is blured the detection threshold of detection algorithm.Experiment to true infrared image sequence shows that the fuzzy detection method of multiframe of the present invention can detect the point target in the low SNR images sequence.

Claims (1)

1, the fuzzy detection method of the multiframe of point target in a kind of low signal-to-noise ratio infrared image sequence is characterized in that comprising following concrete steps:

1) infrared image goes average to handle: adopt the mathematical morphology filter algorithm to obtain the background of image, with complete zero flat-top shape structural element infrared image is opened and then closed procedure, remove all kinds of bright, dark noises and little target in the image, obtain image background, remove the average image what source images and the background subtracting that obtains obtained comprising Weak target and noise;

2) picture noise normalization: will go the average image division to be local wicket one by one, multiple image is estimated the noisiness in each window before and after utilizing, and the noise criteria difference of going the average image to estimate according to each wicket is carried out normalized;

3) single frames is fuzzy detects: adopt the noise probability distribution function as the fuzzy membership function that detects, detect reservation target amplitude information as much as possible by the fuzzy membership function single frames, and make the single frames detection be output as one in the equally distributed stochastic variable of [0,1] interval obedience;

4) movement velocity size and the direction possible according to target, off-line is constructed one group of velocity filtter, corresponding to all possible target trajectory, and then the energy that adds up on all possible target trajectorys, obtains to have, the distribution character curve of statistic during driftlessness;

5) constant false alarm rate target detection: the given detection performance of distribution character curve of statistic during according to having of obtaining of velocity filtter group energy accumulation, driftlessness, utilize the constant false alarm rate criterion, determine the threshold value of target detection, make the judgement whether target exists, if the statistic of certain point judges then that greater than detection threshold there is target in this point; Otherwise then driftlessness exists;

6) simplification of design process: utilize the threshold calculations in the central limit theorem simplification constant false alarm rate target detection in the theory of probability, when number of image frames is big, utilize the constant false alarm rate criterion to carry out threshold value when determining, directly utilizing normal distyribution function definite threshold size of tabling look-up.

Images