CN105869156A - Infrared small target detection method based on fuzzy distance - Google Patents

Abstract

The invention aims at effectively detecting an infrared small target in a complex background, discloses an infrared small target detection method based on fuzzy distance, and relates to the technical field of digital image processing. The method firstly proposes a fuzzy distance concept for a characteristic that the appearance of a small target will cause a bigger change of local texture, and converts the change of the local texture into the measurement of the fuzzy distance. The method secondly proposes a multi-scale fuzzy distance and a multi-scale fuzzy distance map for a characteristic that the size of the small target will change along with the change of an imaging distance, thereby removing a large amount of background noise and noise interference. The method thirdly effectively inhibits the residual background and noise through iterative operation, and enhances the target. The method finally detects the target through an adaptive threshold value. The method is simple and efficient.

Description

A kind of infrared small target detection method based on fuzzy distance

Technical field

The present invention relates to digital image processing techniques field, a kind of infrared small target detection based on fuzzy distance Method.

Background technology

Infrared target detection technology is used widely at many civil areas, such as medical science infrared imaging, remote sensing and gloomy Woods detection, early warning detection etc..Detection performance quality directly determines EFFECTIVE RANGE and the complexity of equipment of infrared system Degree.Remote infrared system is remote because of image-forming range, thus causes that target size is little, intensity is weak, and be easily submerged in very noisy and In background clutter.In this case, effectively detect that the Small object of unknown position/speed/size/shape faces very disaster Degree, thus this kind of technology receives and continues and universal concern.

Existing small target deteection technology can simply be divided into detection before follow the tracks of (Track before Detect, TBD) and Detection (Detect before Track, DBT) two classes before following the tracks of.TBD method the most first searches for target all possible motion rail Mark, and it is cumulative to complete target energy, thus obtain the posterior probability of every movement locus, finally utilize the real mesh of threshold decision Mark movement locus, such as three-dimensional matched filtering, three-dimensional filtering.TBD method be prone to set up relatively more completely theoretical model and Processing method, but calculate complexity, hardware realize trouble, apply in Practical Project less (C.Q.Gao, D.Y.Meng, Y.Yang,Y.T.Wang,X.F.Zhou,A.G.Hauptmann,Infrared patch-image model for small target detection in a single image,IEEE Transactions on Image Processing,22 (12):4996-5009,2013).DBT method typically first detects candidate target according to the gamma characteristic in short-term of single-frame images, then Kinetic characteristic in short-term according to target rejects false target, thus obtains the real movement locus of target.Compared with TBD method, DBT algorithm is simple, it is simple to program modularity realizes, thus play a significant role in real-time target detection field (H.Deng, X.P.Sun,M.L.Liu,C.H.Ye,X.Zhou,Infrared small-target detection using multiscale gray difference weighted image entropy,IEEE Transactions on Aerospace and electronic Systems,52(1):60-72,2016).Be given according to SPIE Small object defines, and the size of Small object is usually no more than the 0.12% of entire image size, thus the appearance of target is to view picture figure As texture structure impact is less, but the impact of localized region texture structure is bigger.Based on this feature, some describe local grain and become Change operator to be suggested, can effectively detect infrared small target, as multiple dimensioned intensity-weighted image entropy, sparse ring represent, the main one-tenth of probability Analyze, local contrast estimates (C.L.Philip, H.Li, Y.T.Wei, T.Xia, and Y.Y.Tang, A local contrast method for small infrared target detection,IEEE Transactions on Geoscience and Remote Sensing,51(1):574-581,2014).The application patented method is under the jurisdiction of DBT side Method.Compared with conventional DBT method, the application patented method proposes a kind of fuzzy distance concept, can effectively portray target internal, Distance inside background, between target and background, thus the local grain caused because of the appearance of Small object change is converted into mould Stick with paste the tolerance of distance, it is achieved background suppression, targets improvement, be conducive to improving the detection probability of target, reduce false-alarm probability.

Although infrared small target detection field is achieved with a lot of achievements, and existing a lot of TBD and DBT algorithms are in engineering Application has obtained good realization.But for low signal-to-noise ratio infrared small target image under complex background, object detection system work Cheng Yiran faces the biggest difficulty and complexity.Therefore, how to design that result is simple, the infrared small target detection of strong robustness Algorithm is the key issue of target detection technique applied research.

Summary of the invention

The present invention be directed to the above-mentioned technical problem that existing infrared small target detection method exists, it is provided that a kind of based on mould Stick with paste the infrared small target detection method of distance.

A kind of infrared small target detection method based on fuzzy distance, comprises the following steps:

Step 1, initialization relevant parameter:

Arranging maximum iteration time L, wherein L is positive integer；Initialize iterations index k=0；Maximum local window is set Mouth size m × n, wherein m and n is the positive odd number more than 1；

Step 2, solve the fuzzy distance of each pixel of infrared image I, comprise the following steps:

Step 2.1, acquisition each pixel of Single Infrared Image Frame I (x, neighborhood space collection { Ω y)_l| l=1, 2 ..., s}, wherein s=min{0.5 (m-1), 0.5 (n-1) }, Ω_lSize be (2l+1) × (2l+1), pixel (x, neighborhood space Ω y)_lDefinition be Ω_l=(p, q) | max (| p-x |, | q-y |)≤l}, (p q) is neighborhood space Ω_l Interior pixel；

Step 2.2, calculate each pixel (x, each neighborhood space Ω y)_lThe gray average D of interior pixel_l(x, y):

$D_l(x,y)=\frac{1}{\#{\mathrm{\Ω}}_l}\underset{(\mathrm{\α},b)\∈{\mathrm{\Ω}}_l}{\Σ}I(a,b),l=1,2,...,s$

Wherein, # Ω_lRepresent neighborhood space Ω_lThe number of interior pixel, (a b) represents neighborhood space Ω to I_lInterior pixel (a, b) gray value at place.

Step 2.3, calculate each pixel (x, y) corresponding to maximum neighborhood space Ω_sEmpty with other each neighborhood Between Ω_i, i=1,2 ..., the fuzzy distance E between s-1_i:

$E_i=1-\frac{1}{e-1}(e^{1-\left|D_i+D_s\right|\·\left|D_i-D_s\right|}-1)$

Wherein e is natural constant, D_sRepresent maximum neighborhood space Ω_sThe gray average of interior pixel, D_iRepresent i-th neighborhood Space Ω_iThe gray average of interior pixel；

Step 3, solve multiple dimensioned fuzzy distance figure:

Traversal infrared image I in each pixel, obtain each pixel multiple dimensioned fuzzy distance E (x, y), so According to the multiple dimensioned fuzzy distance E of each pixel, (x y) and by method for normalizing obtains the multiple dimensioned of infrared image I afterwards Fuzzy distance figure E；

Step 4, iteration stopping criterion judge:

Iterations index k adds 1, it is judged that the relation between iterations index k and maximum iteration time L, if k < L, The multiple dimensioned fuzzy distance figure E that step 3 is obtained, as new infrared image I, returns step 2；If k >=L, stop iteration, The multiple dimensioned fuzzy distance figure E that step 3 is obtained, as final filter result, carries out step 5；

Step 5, solve adaptive threshold T:

To the final filter result obtained through step 4, the most multiple dimensioned fuzzy distance figure E, solve adaptive threshold T, And by adaptive threshold T, multiple dimensioned fuzzy distance figure E is carried out binaryzation, detect infrared small target.

(x, multiple dimensioned fuzzy distance y) is expressed as step 3 mid-infrared each pixel of image I as above

E (x, y)=max{0, E₁,E₂,…,E_s-1}。

In step 5 as above, the determination method of adaptive threshold T is

T=α E_max+β·m_t

Wherein, α and β is positive constant, m_tFor the gray average of multiple dimensioned fuzzy distance figure E, E_maxFor multiple dimensioned fuzzy away from Gray scale maximum from figure E.

The present invention compared with prior art, has the advantage that

1. the appearance for Small object can cause local grain this feature of generation large change, and the present invention proposes one Fuzzy distance concept, can effectively portray the distance inside target internal, background, between target and background, thus by local grain Change is converted into the tolerance of fuzzy distance.

2. the size for Small object can occur this feature of respective change with the change of image-forming range, and the present invention proposes A kind of multiple dimensioned fuzzy distance measure, thus it is effectively improved the detection probability of target, reduce false-alarm probability.

3. first the present invention builds the multiple dimensioned fuzzy distance figure of infrared image under complex background, can reject a large amount of background miscellaneous Ripple and noise jamming；Secondly by interative computation, the background of effectively suppression residual and noise, strengthen target；Then utilize adaptive Answer threshold value separation target, simply and efficiently detect target.

Accompanying drawing explanation

Fig. 1 is the FB(flow block) of the present invention.

Fig. 2 is the multiple dimensioned fuzzy distance figure after successive ignition, and wherein, A figure is the Small object figure under original sky background As (white rectangle frame represents target region), B figure is the multiple dimensioned fuzzy distance figure after an iteration, and C figure is to change for twice Multiple dimensioned fuzzy distance figure after Dai, D figure is the multiple dimensioned fuzzy distance figure after three iteration, and E figure is many after four iteration Yardstick fuzzy distance figure, F figure is the testing result for using adaptive threshold.

Fig. 3 is the filter result schematic diagram of the single infrared small target image using the present embodiment method to obtain.

A, B, C and D: (white rectangle frame represents target to the single infrared small target image being followed successively by under original different background Region), wherein, A figure is the Small object image under sky background, and B figure is the Small object image under clutter background, and C figure is The Small object image under water of marine background, D figure is the Small object image under surface feature background；

E, F, G and H: be corresponding in turn to the filter result using the present embodiment method to obtain in A, B, C and D.

Fig. 4 is the filter result schematic diagram of two the infrared small target images using the present embodiment method to obtain.

A, B, C and D: (white rectangle frame represents target to be followed successively by under original different background two infrared small target images Region), wherein, A figure is the Small object image under sky background, and B figure is the Small object image under sky background, and C figure is Small object image under the water surface background of ocean, D figure is the Small object image under surface feature background；

E, F, G and H: be corresponding in turn to the filter result using the present embodiment method to obtain in A, B, C and D.

Fig. 5 is the filter result schematic diagram using existing method to obtain for A, B, C and the D in Fig. 3.

A1, B1, C1 and D1: be corresponding in turn in Fig. 3 A, Fig. 3 B, Fig. 3 C and Fig. 3 D based on local contrast (Local Contrast measure, LCM) filter result of method；

A2, B2, C2 and D2: be corresponding in turn in Fig. 3 A, Fig. 3 B, Fig. 3 C and Fig. 3 D based on Largest Mean filter (Maxmean Filter, MME) filter result of method；

A3, B3, C3 and D3: be corresponding in turn in Fig. 3 A, Fig. 3 B, Fig. 3 C and Fig. 3 D based on max-medium filter The filter result of (Maxmedian filter, MED) method；

A4, B4, C4 and D4: be corresponding in turn in Fig. 3 A, Fig. 3 B, Fig. 3 C and Fig. 3 D based on top cap filter (Top-hat Filter, THT) filter result of method.

Fig. 6 is the filter result schematic diagram using existing method to obtain for A, B, C and the D in Fig. 4.

A1, B1, C1 and D1: be corresponding in turn in Fig. 4 A, Fig. 4 B, Fig. 4 C and Fig. 4 D based on local contrast (Local Contrast measure, LCM) filter result of method；

A2, B2, C2 and D2: be corresponding in turn in Fig. 4 A, Fig. 4 B, Fig. 4 C and Fig. 4 D based on Largest Mean filter (Maxmean Filter, MME) filter result of method；

A3, B3, C3 and D3: be corresponding in turn in Fig. 4 A, Fig. 4 B, Fig. 4 C and Fig. 4 D based on max-medium filter The filter result of (Maxmedian filter, MED) method；

A4, B4, C4 and D4: be corresponding in turn in Fig. 4 A, Fig. 4 B, Fig. 4 C and Fig. 4 D based on top cap filter (Top-hat Filter, THT) filter result of method.

Detailed description of the invention

Below by embodiment, and combine accompanying drawing, technical scheme is described in further detail.

Embodiment:

Fig. 1 is the structural schematic block diagram of embodiment of the present invention, mainly includes that image inputs: the input infrared little mesh of single frames Logo image；Fuzzy distance solves: solve the fuzzy distance between input picture current region and neighborhood collection；Multiple dimensioned fuzzy distance Figure solves: solve the multiple dimensioned fuzzy distance figure of input picture, portrays that may be present because of caused by the change of image-forming range Target size changes this characteristic；Iteration stopping judges: judge the relation between iterations index and maximum iteration time, weight Multiple iteration, it is thus achieved that final filter result；Threshold value solves: solve the segmentation threshold of final filter result；Binaryzation: divided by threshold value From target, it is thus achieved that target centroid position.

Particularly as follows:

Step 1, initialization relevant parameter:

Arranging maximum iteration time L, wherein L is positive integer, generally 2,3 or 4；Initialize iterations index k=0； Arranging maximum local window size m × n, wherein m and n is the positive odd number more than 1, and the value of general m and n is disposed as 7,9 or 11。

Step 2, solves the fuzzy distance of each pixel of infrared image I:

Infrared small target image under complex background is typically made up of target, complex background and noise three part.By degree Distance inside amount target internal, background, between target and background, thus by because of the local grain caused by the appearance of Small object Change is converted into the tolerance of fuzzy distance, it is achieved background suppression and targets improvement.

The solution procedure of the fuzzy distance of each pixel of infrared image I is as follows:

(1) each pixel of Single Infrared Image Frame I (x, neighborhood space collection { Ω y) are obtained_l| l=1,2 ..., s}, its Middle s=min{0.5 (m-1), 0.5 (n-1) }, Ω_lSize be (2l+1) × (2l+1), (x, neighborhood y) is empty for pixel Between Ω_lDefinition be Ω_l=(p, q) | max (| p-x |, | q-y |)≤l}, (p q) is neighborhood space Ω_lInterior pixel.

(2) each pixel (x, each neighborhood space Ω y) are calculated_lThe gray average D of interior pixel_l(x, y):

$D_l(x,y)=\frac{1}{\#{\mathrm{\&Omega;}}_l}\underset{(\mathrm{\&alpha;},b)\&Element;{\mathrm{\&Omega;}}_l}{\&Sigma;}I(a,b),l=1,2,...,s---\left(1\right)$

Wherein, # Ω_lRepresent neighborhood space Ω_lThe number of interior pixel, (a b) represents neighborhood space Ω to I_lInterior pixel (a, b) gray value at place.

(3) calculate each pixel (x, y) corresponding to maximum neighborhood space Ω_sWith other each neighborhood space Ω_i,i =1,2 ..., the fuzzy distance E between s-1_i:

$E_i=1-\frac{1}{e-1}(e^{1-\left|D_i+D_s\right|\&CenterDot;\left|D_i-D_s\right|}-1)---\left(2\right)$

Wherein e is natural constant, D_sRepresent maximum neighborhood space Ω_sThe gray average of interior pixel, D_iRepresent i-th neighborhood Space Ω_iThe gray average of interior pixel.

Step 3, solves multiple dimensioned fuzzy distance figure:

Although lack the prioris such as infrared small target size, size, but along with the change of image-forming range, the chi of target The features such as very little, size can occur to change to a certain extent.Multiple dimensioned fuzzy distance is utilized to portray that may be present because of image-forming range The target size caused by change change this character.

(x, multiple dimensioned fuzzy distance y) is expressed as each pixel of infrared image I

E (x, y)=max{0, E₁,E₂,...,E_s-1} (3)

Wherein, E_i, i=1,2 ..., s-1 represents pixel (x, a series of fuzzy distances y).

Each pixel in traversal infrared image I, obtains the multiple dimensioned fuzzy distance of each pixel, then passes through Method for normalizing obtains multiple dimensioned fuzzy distance figure E (as shown in the B of Fig. 2) of infrared image I.From the B of Fig. 2 it can be seen that Homogenizing sky background and the cloud layer interior background of infrared image I are inhibited, and target is strengthened.

Step 4, iteration stopping criterion judges:

Iterations index k adds 1, it is judged that the relation between iterations index k and maximum iteration time L, if k < L, will The multiple dimensioned fuzzy distance figure E that step 3 is obtained, as new infrared image I, returns step 2；If k >=L, stop iteration, will The multiple dimensioned fuzzy distance figure that step 3 is obtained, as final filter result, carries out step 5.

The complex background border of infrared image has the thermal imaging feature similar to target, permissible by iteration is repeated several times Eliminate residual background and effect of noise (as shown in Figure 2).The B of Fig. 2 represents the multiple dimensioned fuzzy distance figure after an iteration.From It can be seen that the homogeneous background (inside homogenizing sky and homogenizing cloud layer) in the B of Fig. 2 is well suppressed in the B of Fig. 2, but Remain more cloud layer edge.The multiple dimensioned fuzzy distance figure (as shown in the C of Fig. 2) of the B of Fig. 2 can remove overwhelming majority residual Cloud layer edge, and the multiple dimensioned fuzzy distance figure (as shown in the D of Fig. 2) of the C of Fig. 2 can remove remaining cloud layer edge so that multiple Miscellaneous cloud layer border is suppressed further, and target is further enhanced from.The multiple dimensioned fuzzy distance figure of the D of Fig. 2 is (such as figure Shown in the E of 2) little with the D difference of Fig. 2, illustrate to use the most limited iterations to be obtained with comparatively ideal filtering knot Really.

Step 5, solves adaptive threshold T:

The final filter result (the most multiple dimensioned fuzzy distance figure E) obtained through step 4 is solved adaptive threshold T, And by adaptive threshold T, multiple dimensioned fuzzy distance figure E is carried out binaryzation, (binaryzation result is such as to detect infrared small target Shown in the F of Fig. 2).The determination method of adaptive threshold T is

T=α E_max+β·m_t (4)

Wherein, α and β is positive constant, m_tFor the gray average of multiple dimensioned fuzzy distance figure E, E_maxFor multiple dimensioned fuzzy distance figure The gray scale maximum of E.

Use the filter result of different infrared small target detection method as shown in Figure 5 and Figure 6.Comparison diagram 3, Fig. 4, Fig. 5 and Fig. 6, the filtering performance that the present embodiment method obtains is best, wherein, based on local contrast (Local contrast Measure, LCM) method comes from document C.L.Philip, H.Li, Y.T.Wei, T.Xia, and Y.Y.Tang, A local contrast method for small infrared target detection,IEEE Transactions on Geoscience and Remote Sensing, 51 (1): 574-581,2014, LCM methods first pass through local contrast measure Dissimilarity between current region and neighborhood, then by threshold value separation target；(Maxmean is filtered based on Largest Mean Filter, MME) or max-medium filter (Maxmedian filter, MED) method come from document S.Deshpande, M.Er,and R.Venkateswarlu,Maxmean and Maxmedian filters for detection of Small-targets, Proceeding of SPIE, 1999,3809:74-83, MME/MED method be first pass through Largest Mean/ Median filter wiping out background noise jamming, then determines threshold value according to the statistical property of image, separates target；Filter based on top cap Ripple (Top-hat filter, THT) method comes from document X.Z.Bai and F.G.Zhou, Analysis of new top- hat transformation and the application for infrared dim small target Detection, Pattern Recognition, 2010,43 (6): 2145-2156, THT method first passes through top-cap operator suppression Background and noise, then use threshold value to separate target from filtered image.LCM, MME, MED, THT are all under the jurisdiction of DBT method.

Use Background suppression factor (Background suppression factor, BSF) objective evaluation infrared small target The filtering performance of detection method.The definition of BSF is:

BSF=σ_I/σ_O (5)

Wherein, σ_IRepresent the gray standard deviation of filtered image, σ_ORepresent the gray standard deviation of filter wavefront image.Use The BSF numerical value that LCM, MME, MED, THT and the present embodiment method are obtained is shown in Table 1.As it can be seen from table 1 the present embodiment method Obtain the highest BSF value, illustrate that the present embodiment method can suppress complex background and the noise of infrared small target image effectively, with The conclusion that Fig. 5 with Fig. 6 is obtained is consistent.

Table 1 uses the Background suppression factor (BSF) of different infrared small target detection method to compare

Image	LCM method	MME method	MED method	THT method	The present embodiment method
						Fig. 3 A	0.8134	0.5638	0.5826	1.0157	5.8339
Fig. 3 B	0.9348	0.7368	0.7396	1.8542	8.2408
						Fig. 3 C	0.5686	0.3288	0.3521	0.7993	6.0785
Fig. 3 D	0.4272	0.3800	0.3843	1.6548	6.3409
						Fig. 4 A	0.4179	0.2640	0.2476	1.6413	1.7025
Fig. 4 B	0.1698	0.1259	0.1232	0.3681	0.5156
						Fig. 4 C	0.2227	0.2171	0.2164	1.0775	4.7691
Fig. 4 D	0.5685	0.5246	0.5227	1.6320	7.5393

Specific embodiment described herein is only to present invention spirit explanation for example.Technology neck belonging to the present invention Described specific embodiment can be made various amendment or supplements or use similar mode to replace by the technical staff in territory Generation, but without departing from the spirit of the present invention or surmount scope defined in appended claims.

Claims (3)

1. an infrared small target detection method based on fuzzy distance, it is characterised in that comprise the following steps:

Step 1, initialization relevant parameter:

Arranging maximum iteration time L, wherein L is positive integer；Initialize iterations index k=0；It is big that maximum local window is set Little m × n, wherein m and n is the positive odd number more than 1；

Step 2, solve the fuzzy distance of each pixel of infrared image I, comprise the following steps:

Step 2.1, acquisition each pixel of Single Infrared Image Frame I (x, neighborhood space collection { Ω y)_l| l=1,2 ..., s}, Wherein s=min{0.5 (m-1), 0.5 (n-1) }, Ω_lSize be (2l+1) × (2l+1), pixel (x, neighborhood y) Space Ω_lDefinition be Ω_l=(p, q) | max (| p-x |, | q-y |)≤l}, (p q) is neighborhood space Ω_lInterior pixel；

Step 2.2, calculate each pixel (x, each neighborhood space Ω y)_lThe gray average D of interior pixel_l(x, y):

$D_l(x,y)=\frac{1}{\#{\mathrm{\&Omega;}}_l}\underset{(a,b)\&Element;{\mathrm{\&Omega;}}_l}{\&Sigma;}I(a,b),l=1,2,...,s$

Wherein, # Ω_lRepresent neighborhood space Ω_lThe number of interior pixel, (a b) represents neighborhood space Ω to I_lInterior pixel (a, b) The gray value at place.

Step 2.3, calculate each pixel (x, y) corresponding to maximum neighborhood space Ω_sWith other each neighborhood space Ω_i, I=1,2 ..., the fuzzy distance E between s-1_i:

$E_i=1-\frac{1}{e-1}(e^{1-\left|D_i+D_s\right|\&CenterDot;\left|D_i-D_s\right|}-1)$

Wherein e is natural constant, D_sRepresent maximum neighborhood space Ω_sThe gray average of interior pixel, D_iRepresent i-th neighborhood space Ω_iThe gray average of interior pixel；

Step 3, solve multiple dimensioned fuzzy distance figure:

Traversal infrared image I in each pixel, obtain each pixel multiple dimensioned fuzzy distance E (x, y), then root According to the multiple dimensioned fuzzy distance E of each pixel, (x y) and by method for normalizing obtains the multiple dimensioned fuzzy of infrared image I Distance map E；

Step 4, iteration stopping criterion judge:

Iterations index k adds 1, it is judged that the relation between iterations index k and maximum iteration time L, if k < L, step 3 The multiple dimensioned fuzzy distance figure E obtained, as new infrared image I, returns step 2；If k >=L, stop iteration, step 3 The multiple dimensioned fuzzy distance figure E obtained, as final filter result, carries out step 5；

Step 5, solve adaptive threshold T:

To the final filter result obtained through step 4, the most multiple dimensioned fuzzy distance figure E, solve adaptive threshold T, and lead to Cross adaptive threshold T and multiple dimensioned fuzzy distance figure E is carried out binaryzation, detect infrared small target.

A kind of infrared small target detection method based on fuzzy distance the most according to claim 1, it is characterised in that described Step 3 mid-infrared each pixel of image I (x, multiple dimensioned fuzzy distance y) is expressed as E (x, y)=max{0, E₁, E₂,...,E_s-1}。

A kind of infrared small target detection method based on fuzzy distance the most according to claim 1, it is characterised in that described Step 5 in the determination method of adaptive threshold T be

T=α E_max+β·m_t

Wherein, α and β is positive constant, m_tFor the gray average of multiple dimensioned fuzzy distance figure E, E_maxFor multiple dimensioned fuzzy distance figure E Gray scale maximum.