CN105869156B - A kind of infrared small target detection method based on fuzzy distance - Google Patents

Abstract

In order to effectively detect small infrared targets under complex backgrounds, the invention discloses a small infrared target detection method based on fuzzy distance, and relates to the technical field of digital image processing. The present invention first aims at the feature that the appearance of a small target will cause a large change in the local texture, and proposes a concept of fuzzy distance, thereby converting the change of the local texture into a measure of the fuzzy distance; secondly, the size of the small target will change with the imaging distance According to the characteristics of corresponding changes due to the change of the target, a multi-scale fuzzy distance and a multi-scale fuzzy distance map are proposed, which can eliminate a large number of background clutter and noise interference; then through iterative operations, the residual background and noise can be effectively suppressed and the target can be enhanced; finally Using adaptive threshold to detect objects, the detection method is simple and effective.

Description

A Method of Infrared Small Target Detection Based on Fuzzy Distance

technical field

The invention relates to the technical field of digital image processing, in particular to an infrared small target detection method based on fuzzy distance.

Background technique

Infrared target detection technology has been widely used in many civilian fields, such as medical infrared imaging, remote sensing and forest fire detection, early warning detection and so on. The detection performance directly determines the effective range of the infrared system and the complexity of the equipment. Due to the long imaging distance of the long-range infrared system, the target size is small, the intensity is weak, and it is easy to be submerged in strong noise and background clutter. In this case, it is very difficult to effectively detect small targets with unknown position/speed/size/shape, so this type of technology has received continuous and widespread attention.

Existing small target detection techniques can be simply divided into two categories: Track before Detect (TBD) and Detect before Track (DBT). The TBD method generally first searches all possible motion trajectories of the target, and completes the accumulation of target energy to obtain the posterior probability of each motion trajectory, and finally uses the threshold to judge the real target motion trajectory, such as three-dimensional matched filtering and three-dimensional direction filtering. The TBD method is easy to establish a relatively complete theoretical model and processing method, but the calculation is complicated, the hardware implementation is troublesome, and it is rarely used in actual engineering (C.Q.Gao, D.Y.Meng, Y.Yang, Y.T.Wang, X.F.Zhou, A.G.Hauptmann, Infrared patch-image model for smalltarget detection in a single image, IEEE Transactions on Image Processing, 22(12):4996-5009, 2013). The DBT method generally first detects candidate targets according to the short-term grayscale characteristics of a single frame image, and then eliminates false targets according to the short-term motion characteristics of the target, so as to obtain the real trajectory of the target. Compared with the TBD method, the DBT algorithm is simple and convenient for program modularization, so it plays an important role in the field of real-time target detection (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). According to the definition of small objects given by the International Society for Optical Engineering, the size of small objects generally does not exceed 0.12% of the size of the entire image, so the appearance of objects has little impact on the texture structure of the entire image, but has a greater impact on the texture structure of local areas . Based on this feature, some operators describing local texture changes have been proposed, which can effectively detect small infrared targets, such as multi-scale gray-scale weighted image entropy, sparse ring representation, probabilistic principal component analysis, and local contrast measurement (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 patented method of this application belongs to the DBT method. Compared with the conventional DBT method, the patented method of this application proposes a concept of fuzzy distance, which can effectively describe the distance between the inside of the target, the inside of the background, and the target and the background, thereby converting the local texture changes caused by the appearance of small targets into The measurement of fuzzy distance can realize background suppression and target enhancement, which is conducive to improving the detection probability of targets and reducing the probability of false alarms.

Although many achievements have been made in the field of infrared small target detection, and many TBD and DBT algorithms have been well realized in engineering applications. However, for infrared small target images with low signal-to-noise ratio in complex backgrounds, the target detection system engineering still faces great difficulty and complexity. Therefore, how to design an infrared small target detection algorithm with simple results and strong robustness is a key issue in the application research of target detection technology.

Contents of the invention

The invention aims at the above-mentioned technical problems existing in the existing infrared small target detection method, and provides an infrared small target detection method based on fuzzy distance.

An infrared small target detection method based on fuzzy distance, comprising the following steps:

Step 1. Initialize related parameters:

Set the maximum number of iterations L, where L is a positive integer; initialize the number of iterations index k=0; set the maximum local window size m×n, where m and n are both positive odd numbers greater than 1;

Step 2, solving the fuzzy distance of each pixel of the infrared image I includes the following steps:

Step 2.1, obtain the neighborhood space set {Ω _l |l=1,2,...,s} of each pixel point (x, y) of the single-frame infrared image I, where s=min{0.5 (m- 1), 0.5·(n-1)}, the size of Ω _l is (2l+1)×(2l+1), and the neighborhood space Ω _l of a pixel (x, y) is defined as Ω _l ={( p,q)|max(|px|,|qy|)≤l}, (p,q) is a pixel in the neighborhood space Ω _l ;

Step 2.2. Calculating the average gray value D _l (x, y) of pixels in each neighborhood space Ω _l of each pixel point (x, y):

Among them, #Ω _l represents the number of pixels in the neighborhood space Ω _l , and I(a, b) represents the gray value at the pixel point (a, b) in the neighborhood space Ω _l .

Step 2.3. Calculate the fuzzy distance E _i between the largest neighborhood space Ω _s corresponding to each pixel point (x, y) and other neighborhood spaces Ω _i , i=1,2,...,s-1:

Where e is a natural constant, D _s represents the average gray value of pixels in the largest neighborhood space Ω _s , and D _i represents the average gray value of pixels in the i-th neighborhood space Ω _i ;

Step 3. Solve the multi-scale fuzzy distance map:

Traverse each pixel in the infrared image I to obtain the multi-scale blur distance E(x, y) of each pixel, and then use the normalization method according to the multi-scale blur distance E(x, y) of each pixel Obtain the multi-scale fuzzy distance map E of the infrared image I;

Step 4, iteration stop criterion judgment:

Add 1 to the iteration number index k, and judge the relationship between the iteration number index k and the maximum iteration number L, if k<L, use the multi-scale fuzzy distance map E obtained in step 3 as the new infrared image I, and return to step 2; If k≥L, stop the iteration, use the multi-scale fuzzy distance map E obtained in step 3 as the final filtering result, and proceed to step 5;

Step 5. Solve the adaptive threshold T:

For the final filtering result obtained through step 4, that is, the multi-scale fuzzy distance map E, the adaptive threshold T is solved, and the multi-scale fuzzy distance map E is binarized by the adaptive threshold T to detect small infrared targets.

The multi-scale fuzzy distance of each pixel point (x, y) of the infrared image I in step 3 above is expressed as

E(x,y)=max{0,E ₁ ,E ₂ ,...,E _s-1 }.

The method for determining the adaptive threshold T in step 5 as described above is

T＝α·E _max +β·m _t

Among them, α and β are positive constants, m _t is the average gray value of the multi-scale fuzzy distance map E, and E _max is the maximum gray value of the multi-scale fuzzy distance map E.

Compared with the prior art, the present invention has the following advantages:

1. Aiming at the feature that the appearance of a small target will cause a large change in the local texture, the present invention proposes a concept of fuzzy distance, which can effectively describe the distance between the inside of the target, the inside of the background, and the distance between the target and the background, so that the local texture The change translates into a measure of blur distance.

2. In view of the characteristic that the size of small targets will change with the change of imaging distance, the present invention proposes a multi-scale fuzzy distance measurement method, thereby effectively improving the detection probability of targets and reducing the probability of false alarms.

3. The present invention first constructs a multi-scale fuzzy distance map of an infrared image under a complex background, which can eliminate a large number of background clutter and noise interference; secondly, through iterative operations, effectively suppress the residual background and noise, and enhance the target; then use adaptive threshold to separate target, detect targets easily and efficiently.

Description of drawings

Fig. 1 is a flowchart of the present invention.

Figure 2 is the multi-scale fuzzy distance map after multiple iterations, in which, picture A is the small target image under the original sky background (the white rectangle box indicates the area where the target is located), and picture B is the multi-scale fuzzy distance map after one iteration, Picture C is the multi-scale fuzzy distance map after two iterations, picture D is the multi-scale fuzzy distance map after three iterations, picture E is the multi-scale fuzzy distance map after four iterations, and picture F is the one using adaptive threshold Test results.

Fig. 3 is a schematic diagram of a filtering result of a single infrared small target image obtained by using the method of this embodiment.

A, B, C, and D: the original single infrared small target images in different backgrounds (the white rectangle box indicates the area where the target is located), among which, A is the small target image under the sky background, and B is the clutter background The image of the small target, C is the image of the underwater small target in the ocean background, and D is the image of the small target in the background of the ground object;

E, F, G, and H: corresponding to A, B, C, and D in turn, the filtering results obtained by using the method of this embodiment.

FIG. 4 is a schematic diagram of filtering results of two infrared small target images obtained by using the method of this embodiment.

A, B, C, and D: two infrared small target images under different original backgrounds (the white rectangle box indicates the area where the target is located), in which, A is the small target image under the sky background, and B is the sky background The small target image of , C is the small target image under the background of the ocean water surface, and D is the small target image under the ground object background;

E, F, G, and H: corresponding to A, B, C, and D in turn, the filtering results obtained by using the method of this embodiment.

FIG. 5 is a schematic diagram of filtering results obtained by using an existing method for A, B, C and D in FIG. 3 .

A1, B1, C1 and D1: the filtering results based on the local contrast measure (LCM) method corresponding to Fig. 3A, Fig. 3B, Fig. 3C and Fig. 3D in turn;

A2, B2, C2 and D2: the filtering results based on the Maxmean filter (MME) method corresponding to Fig. 3A, Fig. 3B, Fig. 3C and Fig. 3D in turn;

A3, B3, C3 and D3: corresponding to the filtering results of the Maxmedian filter (Maxmedian filter, MED) method in sequence corresponding to Fig. 3A, Fig. 3B, Fig. 3C and Fig. 3D;

A4, B4, C4 and D4: the filtering results based on the Top-hat filter (Top-hat filter, THT) method corresponding to FIG. 3A , FIG. 3B , FIG. 3C and FIG. 3D in sequence.

FIG. 6 is a schematic diagram of filtering results obtained by using existing methods for A, B, C and D in FIG. 4 .

A1, B1, C1 and D1: the filtering results based on the local contrast measure (LCM) method corresponding to Fig. 4A, Fig. 4B, Fig. 4C and Fig. 4D in turn;

A2, B2, C2 and D2: the filtering results based on the Maxmean filter (Maxmeanfilter, MME) method corresponding to Fig. 4A, Fig. 4B, Fig. 4C and Fig. 4D in turn;

A3, B3, C3 and D3: the filtering results based on the Maxmedian filter (MED) method corresponding to Fig. 4A, Fig. 4B, Fig. 4C and Fig. 4D in turn;

A4, B4, C4 and D4: the filtering results based on the Top-hat filter (Top-hat filter, THT) method corresponding to FIG. 4A , FIG. 4B , FIG. 4C and FIG. 4D in sequence.

Detailed ways

The technical solutions of the present invention will be further specifically described below through the embodiments and in conjunction with the accompanying drawings.

Example:

Fig. 1 is a schematic block diagram of the structure of the embodiment of the present invention, mainly including - image input: input a single frame infrared small target image; fuzzy distance solution: solve the fuzzy distance between the current region of the input image and the neighborhood set; multi-scale fuzzy distance map Solution: solve the multi-scale fuzzy distance map of the input image, and describe the possible characteristics of the target size change caused by the change of the imaging distance; iteration stop judgment: judge the relationship between the iteration number index and the maximum number of iterations, and repeat the iteration , to obtain the final filtering result; threshold solution: solve the segmentation threshold of the final filtering result; binarization: separate the target through the threshold, and obtain the target centroid position.

Specifically:

Step 1, initialize relevant parameters:

Set the maximum number of iterations L, where L is a positive integer, usually 2, 3 or 4; initialize the number of iterations index k=0; set the maximum local window size m×n, where m and n are positive odd numbers greater than 1, generally The values of m and n are both set to 7, 9 or 11.

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

Infrared small target images under complex background generally consist of three parts: target, complex background and noise. By measuring the distance between the inside of the target, the inside of the background, and between the target and the background, the local texture changes caused by the appearance of small targets are converted into the measurement of blur distance, and background suppression and target enhancement are realized.

The process of solving the fuzzy distance of each pixel in the infrared image I is as follows:

(1) Obtain the neighborhood space set {Ω _l |l=1,2,...,s} of each pixel point (x, y) of the single-frame infrared image I, where s=min{0.5·(m-1) ,0.5·(n-1)}, the size of Ω _l is (2l+1)×(2l+1), and the neighborhood space Ω _l of a pixel point (x, y) is defined as Ω _l ={(p, q)|max(|px|,|qy|)≤l}, (p,q) is a pixel in the neighborhood space Ω _l .

(2) Calculate the average gray value D _l (x, y) of pixels in each neighborhood space Ω _l of each pixel (x, y):

Among them, #Ω _l represents the number of pixels in the neighborhood space Ω _l , and I(a, b) represents the gray value at the pixel point (a, b) in the neighborhood space Ω _l .

(3) Calculate the fuzzy distance E between the largest neighborhood space Ω _s corresponding to each pixel (x, y) and other neighborhood spaces Ω _i , i=1,2,...,s-1 _i :

Where e is a natural constant, D _s represents the average gray value of pixels in the largest neighborhood space Ω _s , and D _i represents the average gray value of pixels in the i-th neighborhood space Ω _i .

Step 3, solve the multi-scale fuzzy distance map:

Although there is a lack of prior knowledge about the size and size of infrared small targets, as the imaging distance changes, the size and size of the target will change to a certain extent. Using the multi-scale fuzzy distance to characterize the property of possible object size change caused by the change of imaging distance.

The multi-scale fuzzy distance of each pixel point (x, y) in the infrared image I is expressed as

E(x,y)=max{0,E ₁ ,E ₂ ,...,E _s-1 } (3)

Wherein, E _i , i=1, 2, . . . , s-1 represent a series of blur distances of pixel points (x, y).

Traverse each pixel in the infrared image I to obtain the multi-scale fuzzy distance of each pixel, and then obtain the multi-scale fuzzy distance map E of the infrared image I through the normalization method (as shown in Figure 2B). From Figure 2B, it can be seen that the homogeneous sky background and cloud interior background of the infrared image I are suppressed, and the target is enhanced.

Step 4, iteration stop criterion judgment:

Add 1 to the iteration number index k, judge the relationship between the iteration number index k and the maximum iteration number L, if k<L, use the multi-scale fuzzy distance map E obtained in step 3 as the new infrared image I, and return to step 2; If k≥L, stop the iteration, take the multi-scale fuzzy distance map obtained in step 3 as the final filtering result, and proceed to step 5.

The complex background boundary of the infrared image has thermal imaging characteristics similar to the target, and the influence of the residual background and noise can be eliminated through repeated iterations (as shown in Figure 2). B of Fig. 2 represents the multi-scale fuzzy distance map after one iteration. It can be seen from B of FIG. 2 that the homogeneous background (homogeneous sky and homogeneous cloud interior) in FIG. 2 B is well suppressed, but more cloud edges remain. The multi-scale fuzzy distance map of B in Figure 2 (as shown in Figure 2 C) can remove most of the remaining cloud edges, while the multi-scale fuzzy distance map of Figure 2 C (shown in Figure 2 D) can Removing the remaining cloud edges enables further suppression of complex cloud boundaries and further enhancement of targets. The multi-scale fuzzy distance map of D in Fig. 2 (as shown in E in Fig. 2) is not much different from D in Fig. 2, indicating that a more ideal filtering result can be obtained by using a suitable limited number of iterations.

Step 5, solve the adaptive threshold T:

The adaptive threshold T is solved for the final filtering result obtained in step 4 (ie, the multi-scale fuzzy distance map E), and the multi-scale fuzzy distance map E is binarized by the adaptive threshold T to detect small infrared targets (two The valued results are shown in F of Figure 2). The method of determining the adaptive threshold T is

T＝α·E _max +β·m _t (4)

Among them, α and β are positive constants, m _t is the average gray value of the multi-scale fuzzy distance map E, and E _max is the maximum gray value of the multi-scale fuzzy distance map E.

The filtering results of different infrared small target detection methods are shown in Figure 5 and Figure 6. Comparing Fig. 3, Fig. 4, Fig. 5 and Fig. 6, the filtering performance obtained by the method of this embodiment is the best, wherein, based on local contrast measure (Local contrastmeasure, LCM) method comes from literature C.L.Philip, H.Li, Y.T.Wei, T .Xia, and Y.Y.Tang, A localcontrast method for small infrared target detection, IEEE Transactions on Geoscience and Remote Sensing, 51(1):574-581, 2014, the LCM method first measures the difference between the current region and the neighborhood through the local contrast Similarity, and then separate the target by threshold; based on the maximum mean filter (Maxmean filter, MME) or maximum median filter (Maxmedian filter, MED) method from the literature S.Deshpande, M.Er, and R.Venkateswarlu, Maxmean and Maxmedian filters For detection of small-targets, Proceeding of SPIE, 1999, 3809: 74-83, the MME/MED method is to first filter out the background clutter interference through the maximum mean/median filter, and then determine the threshold according to the statistical characteristics of the image to separate the target ; Based on the top-hat filter (Top-hat filter, THT) method from the literature 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, the THT method first suppresses the background and noise through the top-hat operator, and then uses the threshold to separate the target from the filtered image. LCM, MME, MED, and THT all belong to the DBT method.

The background suppression factor (Background suppression factor, BSF) is used to objectively evaluate the filtering performance of the infrared small target detection method. BSF is defined as:

BSF＝ _σI / _σO (5)

Among them, σ _I represents the gray standard deviation of the image after filtering, and σ _O represents the gray standard deviation of the image before filtering. The BSF values obtained by using LCM, MME, MED, THT and the method of this embodiment are shown in Table 1. It can be seen from Table 1 that the method of this embodiment obtains the highest BSF value, indicating that the method of this embodiment can effectively suppress the complex background and noise of the infrared small target image, which is consistent with the conclusions obtained in Figures 5 and 6.

Table 1 Comparison of background suppression factor (BSF) using different infrared small target detection methods

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

The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which the present invention belongs can make various modifications or supplements to the described specific embodiments or adopt similar methods to replace them, but they will not deviate from the spirit of the present invention or go beyond the definition of the appended claims range.

Claims (2)

1. a kind of infrared small target detection method based on fuzzy distance, is characterized in that, comprises the following steps: Step 1. Initialize related parameters: Set the maximum number of iterations L, where L is a positive integer; initialize the number of iterations index k=0; set the maximum local window size m×n, where m and n are both positive odd numbers greater than 1; Step 2, solving the fuzzy distance of each pixel of the infrared image I includes the following steps: Step 2.1, obtain the neighborhood space set {Ω _l |l=1,2,...,s} of each pixel point (x, y) of the single-frame infrared image I, where s=min{0.5 (m- 1), 0.5·(n-1)}, the size of Ω _l is (2l+1)×(2l+1), and the neighborhood space Ω _l of a pixel (x, y) is defined as Ω _l ={( p,q)|max(|px|,|qy|)≤l}, (p,q) is a pixel in the neighborhood space Ω _l ; Step 2.2. Calculating the average gray value D _l (x, y) of pixels in each neighborhood space Ω _l of each pixel point (x, y): Among them, #Ω _l represents the number of pixels in the neighborhood space Ω _l , I(a, b) represents the gray value at the pixel point (a, b) in the neighborhood space Ω _l , Step 2.3. Calculate the fuzzy distance E between the largest neighborhood space Ω _s corresponding to each pixel (x, y) and other neighborhood spaces Ω _i , i=1,2,...,s-1 _i : Where e is a natural constant, D _s represents the average gray value of pixels in the largest neighborhood space Ω _s , and D _i represents the average gray value of pixels in the i-th neighborhood space Ω _i ; Step 3. Solve the multi-scale fuzzy distance map: Traverse each pixel in the infrared image I to obtain the multi-scale blur distance E(x, y) of each pixel, and then use the normalization method according to the multi-scale blur distance E(x, y) of each pixel Obtain the multi-scale fuzzy distance map E of the infrared image I; Step 4, iteration stop criterion judgment: Add 1 to the iteration number index k, and judge the relationship between the iteration number index k and the maximum iteration number L, if k<L, use the multi-scale fuzzy distance map E obtained in step 3 as the new infrared image I, and return to step 2; If k≥L, stop the iteration, take the multi-scale fuzzy distance map E obtained in step 3 as the final filtering result, and proceed to step 5; Step 5. Solve the adaptive threshold T: For the final filtering result obtained through step 4, that is, the multi-scale fuzzy distance map E, solve the adaptive threshold T, and use the adaptive threshold T to binarize the multi-scale fuzzy distance map E to detect small infrared targets. In step 3, the multi-scale fuzzy distance of each pixel (x, y) of the infrared image I is expressed as E(x,y)=max{0,E ₁ ,E ₂ ,...,E _s-1 }. 2. a kind of infrared small target detection method based on fuzzy distance according to claim 1, is characterized in that, the determination method of self-adaptive threshold T in the described step 5 is T＝α·E _max +β·m _t Among them, α and β are positive constants, m _t is the average gray value of the multi-scale fuzzy distance map E, and E _max is the maximum gray value of the multi-scale fuzzy distance map E.