An anti-interference infrared small target detection method

Abstract

The invention relates to the field of image processing, and discloses an anti-interference infrared small target detection method. Traditional algorithms are stable in simple environments according to target prior modeling, but the presence of similar interference sources in complex environments will lead to a large number of false alarms. The main scheme includes first establishing a new infrared image representation method suitable for complex backgrounds, treating the infrared image as a linear superposition of a foreground image and a background image with a target, and controlled by a control matrix; secondly, constructing a three-color constraint graph and solving the control matrix; the difference between the real target and the interference source is that the control values ​​of the pixels around them are inconsistent, the surrounding pixels of the real target are more represented by the background, and the control value will be low, while the surrounding pixels of the interference source are more represented by the interference source itself, and its control value will be high, so finally the sparse interference source is suppressed according to the control matrix, and the real target is enhanced. The model has a very strong anti-interference ability for complex environments.

Description

Technical Field

[0001] The present invention relates to the field of image processing and machine vision applications, in particular to the problem of small target detection under infrared imaging, and is of great significance to the military defense system. Background Art

[0002] Long-range infrared small target detection is the core of infrared search and tracking system. The algorithm based on low-rank sparse decomposition has become one of the current mainstream algorithms due to its scene adaptability and high target detection rate. It regards the target as a sparse component and the background as a low-rank component, and realizes small target detection through the idea of ​​low-rank sparse decomposition. However, there may be interference sources with sparse characteristics in complex backgrounds, such as cirrus cloud edges, high-brightness parts of complex ground, etc. Therefore, how to distinguish sparse interference sources from real sparse targets is the key. For complex environments, we first establish a new infrared image representation method suitable for complex backgrounds, which regards the infrared image as a linear superposition of the foreground image containing the target and the background image containing the background and interference sources. This linear superposition is controlled by the control matrix α. This representation method is more robust than the traditional method of directly superimposing the target image and the background image. We regard the sparse areas of the image as the absolute foreground, which contains the real target and sparse interference sources. The essential difference between them is that the real target often exists independently of the background, and its components have no relationship with the surrounding background, while the sparse components of the interference source are local areas of the interference source, so these sparse pixels have stronger similarity with the surrounding pixels. We define an affinity relationship to describe the relationship between pixels. This affinity relationship is also closely related to the control matrix in the infrared image representation method proposed in the present invention. We regard the surrounding pixel area of ​​the sparse area as a mixed area. For the real target in the sparse component, the affinity between the surrounding pixels and the target is low, so its mixed area is more represented by the background, and the α value is close to 0. For the mixed area of ​​the interference source, they may all be part of the interference source, so the α value will be high. Therefore, the control matrix α can be used to design an algorithm to measure the affinity between each sparse candidate area and the surrounding background to distinguish between real targets and sparse interference sources. Summary of the invention

[0003] The present invention provides an anti-interference infrared small target detection method. The technical problem to be solved is how to distinguish real targets and sparse interference sources in sparse areas, thereby ensuring target detection while reducing the false alarm rate, and realizing a set of anti-interference infrared small target detection algorithms suitable for complex environments.

[0004] In order to solve the above technical problems and achieve the above objectives, the technical solutions adopted by the present invention are as follows.

[0005] An infrared small target detection method suitable for complex backgrounds includes the following steps:

[0006] Step 1: Express the infrared small target image in a complex environment and proceed to step 2;

[0007] Step 2: Input the original infrared image to extract the sparse component image, and construct a three-color constraint graph as the constraint for model solution, and then proceed to step 3;

[0008] Step 3: Solve the model and obtain the control matrix α with affinity relationship, and then go to step 4;

[0009] Step 4: Use the control matrix α obtained in step 3 to construct a sparse interference source suppression method, and suppress and enhance the interference sources and targets in the sparse component image of step 2 respectively to obtain the final target image.

[0010] In the above technical solution, step 1 comprises the following steps:

[0011] Step 1.1: In the traditional algorithm, the infrared small target image f O As sparse component f S and the low-rank component f D The background is regarded as a low-rank component and the target is regarded as a sparse component. However, this expression is not suitable for complex environments because there are many interference sources in complex environments, and some of their local structures are still regarded as sparse components in the global environment. The present invention first proposes a new infrared small target image expression mode. This method regards the infrared image as a linear superposition of a foreground image with a target and a background image containing both interference sources and background. The weights of the foreground and background in the linear superposition in the global image are controlled by the control matrix α. Modeling in this way can better distinguish the target from other interference sources, as shown in formula (1). Formula (1) is as follows:

[0012] f O =αf F +(1-α)f B , α(C)=k#(1)

[0013] Among them, f O 、f F 、f B Respectively represent the input infrared image, the unknown foreground image with the target, and the unknown background image. α is the unknown control matrix of the same size as the original image, and the value range is [0, 1]. Since each pixel contains 3 unknown solutions and the problem has only one constraint, the solution process lacks constraints. In order to solve this problem, we add the constraint α(C) = k, where C is the known pixel and k represents the corresponding constraint value. By setting the constraint α(C) = k, the solution of other unknown solutions can be completed.

[0014] In the above technical solution, step 2 comprises the following steps:

[0015] Step 2.1: In order to provide accurate constraints for step 1.1, the low-rank sparse decomposition method can provide accurate prior information. We no longer set the target f T Equivalent to the sparse component f in the infrared image S , but consider the sparse components as the target f T With the interference source f which also has sparse property G We use the low-rank sparse decomposition method to obtain the sparse component f S , which will be used to provide constraints later.

[0016] Step 2.2: The sparse component image f obtained in step 2.1 S Perform global threshold segmentation to obtain the sparse template M S , where the value of the pixel in the sparse area is 1, and the other values ​​are 0;

[0017] Step 2.3: For the sparse template M obtained in step 2.2 S , construct the three-color constraint map Tmap through formula (2). The three-color constraint map is used to provide constraint values ​​for formula (1) in step 1.1. The values ​​of the sparse area and the background area are known areas, and their values ​​are 1 and 0, respectively, representing the foreground and background. The surrounding area of ​​the target is regarded as an unknown area, represented by 0.5, which means that the area is a mixed superposition area of ​​the foreground and background. The α value of the unknown area can be solved through the known area. Formula (2) is as follows:

[0018] Tmap=0.5expan(M S )+0.5M S #(2)

[0019] Among them, expan(*) represents morphological expansion, Tmap is a three-color constraint map. Since it contains three values ​​0, 0.5, and 1, which correspond to the three colors black, gray, and white in the image, it is called a three-color constraint map, which respectively represent the background area, unknown area, and foreground area.

[0020] In the above technical solution, step 3 comprises the following steps:

[0021] Step 3.1: The essential difference between the real target and the sparse interference source is that the real target often exists independently of the background, and its components have no relationship with the surrounding background, while the sparse components of the interference source are the local area of ​​the interference source, so these sparse pixels have stronger similarity with the surrounding pixels. We first define the affinity relationship to describe the relationship between pixels, as shown in formula (3), which is as follows:

[0022]

[0023] Among them, a(i, j) represents the affinity between pixel i and pixel j, I(i) and P(i) represent the grayscale value and position of pixel i respectively, and S(i) is the local structure of pixel i. This item is mainly set for cluster structures such as cirrus clouds. C control parameter is used to maintain a(i, j)∈[0, 1], λ 0 is the weight coefficient.

[0024] Step 3.2: The affinity relationship between multiple pixels can be expressed by an affinity relationship matrix as shown in formula (4), which is as follows:

[0025]

[0026] Where N represents the number of pixels.

[0027] Step 3.3: The infrared image is relatively smooth as a whole, so it can be assumed that each pixel can be represented by its neighboring pixels, as shown in formula (5). Formula (5) is as follows:

[0028]

[0029] Step 3.4: Combine the control matrix α of equation (1) in step 1.1 with equation (5) in step 3.3 to obtain equation (6), which is as follows:

[0030]

[0031] Where D i To guarantee

[0032] Step 3.5: Matrix the equation (6) in step 3.4 to obtain equation (7), which is as follows:

[0033]

[0034] Among them, α table is the vector composed of all α.

[0035] Step 3.6: Use equation (7) in step 3.5 to construct the objective function with constraints, as shown in equation (8). Then, the vector α can be obtained by solving the objective function and reconstructing the complete control matrix α. Equation (8) is as follows:

[0036]

[0037] Among them, α(C) = k is a constraint, where C is a known pixel, which can be provided by the three-color constraint map Tmap in 2.3. The black area and white area of ​​Tmap are known pixels, representing the background and foreground respectively, and their corresponding k are 0 and 1 respectively. The model has sufficient constraints to solve, so the α value of the unknown pixel (gray area) can be solved to obtain the completed vector α, and the complete control matrix α can be reconstructed.

[0038] In the above technical solution, step 4 comprises the following steps:

[0039] Step 4.1: The essential difference between the real target and the sparse interference source is that they have different affinity with the surroundings. This also shows that the α value of the mixed area around the sparse area is different. For the real target in the sparse component, the affinity between the surrounding pixels and the target is low, so its mixed area is more characterized by the background, and the α value is close to 0. For the mixed area of ​​the interference source, they may all be part of the interference source, so the α value will be high. Therefore, in this step, first use the three-color constraint map Tmap obtained in step 2.3 to calculate the unknown area template M through formula (9): un The formula (9) is as follows:

[0040]

[0041] Step 4.2: Use the control matrix α obtained in step 3.6 and the neighbor template M obtained in step 4.1 un , calculate the sparse area Sparse by formula (10) k Affinity attribute Af with its neighbors k , formula (10) is as follows:

[0042]

[0043] Step 4.3: Add the affinity attribute Af obtained in step 4.2 k , the inhibition factor ξ is calculated by formula (11) k , formula (11) is as follows:

[0044] ξ k =tanh(Af k -K)#(11)

[0045] Where K represents the affinity attribute threshold, tanh(*) is the activation function, and tanh(x) = (e x -e -x ) / (e x +e -x ).

[0046] Step 4.4: Use the inhibition factor ξ obtained in step 4.3 k, get the sparse component image f from step 2.1 S The interference source is suppressed and the real target is enhanced by formula (12), and the final target detection result f is obtained. D . Formula (12) is as follows:

[0047]

[0048] Among them, C is used to control the interference source suppression degree or the target enhancement degree, b represents the sparse component image f S The value of the background area.

[0049] Because the present invention adopts the above technical solution, it has the following beneficial effects:

[0050] 1. The infrared small target detection method proposed in the present invention has strong adaptability to various complex environments. It can ensure a high target detection rate while greatly reducing the false alarm rate.

[0051] 2. The infrared image representation mode proposed in step 1 of the present invention solves the problem that the traditional mode is insufficient in representing complex environments. The method proposed in the present invention is more suitable for the detection and modeling of weak and small targets in complex environments;

[0052] 3. The three-color constraint graph proposed in step 2 of the present invention provides accurate constraint terms for solving the objective function, which is helpful for distinguishing the subsequent sparse interference sources and real targets;

[0053] 4. Step 3 of the present invention essentially describes the difference between the sparse interference source and the real target, and proposes a solution method for the control matrix in step 1, which is helpful for the design of the sparse interference source suppression algorithm;

[0054] 5. The interference source suppression and target enhancement technology based on affinity attributes proposed in step 4 of the present invention can be widely embedded in various infrared small target detection algorithms, and can be used as post-processing to further reduce the false alarm rate in the results. At the same time, since most of the entire calculation process is based on matrix operations, the speed is relatively fast. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 Design a process for the entire algorithm;

[0056] Figure 2 Indicates the affinity relationship between a sparse pixel and its neighboring pixels. (a) and (c) are the original infrared images, and the real target is marked by a red box. (b) and (d) are the sparse components in (a) and (c), respectively. The upper part is the original image block, and the lower part is the affinity relationship. The sparse pixels are marked by red dots.

[0057] Figure 3The first row is the sparse component image block, the second row is the 3D display of the first row image block, and the third row is the result after suppression. The last column is the real target;

[0058] Figure 4 These are the related result images of small target detection using this method. (a) and (d) are the original images.

[0059] (b) and (e) are sparse component images, which can be found to contain a large number of false alarms, and (c) and (f) are target images. DETAILED DESCRIPTION

[0060] The present invention is further described in detail below in conjunction with test examples and specific implementation methods. However, this should not be understood as the scope of the above subject matter of the present invention being limited to the following embodiments, and all technologies realized based on the content of the present invention belong to the scope of the present invention.

[0061] The present invention provides an anti-interference infrared small target detection method. The technical problem to be solved is how to distinguish the real target and the sparse interference source in the sparse area, thereby ensuring the target detection while reducing the false alarm rate, and realizing a set of anti-interference infrared small target detection algorithms suitable for complex environments. The whole algorithm design structure framework is as follows Figure 1 As shown, the steps include:

[0062] Step 1.1: In the traditional algorithm, the infrared small target image f O As sparse component f S and the low-rank component f D The background is regarded as a low-rank component and the target is regarded as a sparse component. However, this expression is not suitable for complex environments because there are many interference sources in complex environments, and some of their local structures are still regarded as sparse components in the global environment. The present invention first proposes a new infrared small target image expression mode. This method regards the infrared image as a linear superposition of a foreground image with a target and a background image containing both interference sources and background. The weights of the foreground and background in the linear superposition in the global image are controlled by the control matrix α. Modeling in this way can better distinguish the target from other interference sources, as shown in formula (1). Formula (1) is as follows:

[0063] f O =αf F +(1-α)f B , α(C)=k#(1)

[0064] Among them, f O 、f F 、f BRespectively represent the input infrared image, the unknown foreground image with the target, and the unknown background image. α is the unknown control matrix of the same size as the original image, and the value range is [0, 1]. Since each pixel contains 3 unknown solutions and the problem has only one constraint, the solution process lacks constraints. In order to solve this problem, we add the constraint α(C) = k, where C is the known pixel and k represents the corresponding constraint value. By setting the constraint α(C) = k, the solution of other unknown solutions can be completed.

[0065] Step 2.1: In order to provide accurate constraints for step 1.1, the low-rank sparse decomposition method can provide accurate prior information. We no longer set the target f T Equivalent to the sparse component f in the infrared image S , but consider the sparse components as the target f T With the interference source f which also has sparse property G We use the low-rank sparse decomposition method to obtain the sparse component f. , which will be used to provide constraints later. In this example, the image size is 250×250, and the most basic tensor decomposition method RIPT model is used to obtain the sparse component image f S , where the target is detected but also contains a large number of false alarms.

[0066] Step 2.2: The sparse component image f obtained in step 2.1 S Perform global threshold segmentation to obtain the sparse template M S , where the value of the pixel in the sparse area is 1, and the other values ​​are 0;

[0067] Step 2.3: For the sparse template M obtained in step 2.2 S , construct the three-color constraint map Tmap through formula (2). The three-color constraint map is used to provide constraint values ​​for formula (1) in step 1.1. The values ​​of the sparse area and the background area are known areas, and their values ​​are 1 and 0, respectively, representing the foreground and background. The surrounding area of ​​the target is regarded as an unknown area, represented by 0.5, which means that the area is a mixed superposition area of ​​the foreground and background. The α value of the unknown area can be solved through the known area. Formula (2) is as follows:

[0068] Tmap=0.5expan(M S )+0.5M S #(2)

[0069] Among them, expan(*) represents morphological expansion, Tmap is a three-color constraint map. Since it contains three values ​​0, 0.5, and 1, which correspond to the three colors black, gray, and white in the image, it is called a three-color constraint map, which respectively represent the background area, unknown area, and foreground area.

[0070] Step 3.1: The essential difference between the real target and the sparse interference source is that the real target often exists independently of the background, and its components have no relationship with the surrounding background, while the sparse components of the interference source are the local area of ​​the interference source, so these sparse pixels have stronger similarity with the surrounding pixels. We first define the affinity relationship to describe the relationship between pixels, as shown in formula (3), which is as follows:

[0071]

[0072] Among them, a(i, j) represents the affinity between pixel i and pixel j, I(i) and P(i) represent the grayscale value and position of pixel i respectively, and S(i) is the local structure of pixel i. This item is mainly set for cluster structures such as cirrus clouds. C control parameter is used to maintain a(i, j)∈[0, 1], λ 0 is the weight coefficient. In this example, λ 0 =0.6.

[0073] Step 3.2: The affinity relationship between multiple pixels can be expressed by an affinity relationship matrix as shown in formula (4), which is as follows:

[0074]

[0075] Where N represents the number of pixels.

[0076] Step 3.3: The infrared image is relatively smooth as a whole, so it can be assumed that each pixel can be represented by its neighboring pixels, as shown in formula (5). Formula (5) is as follows:

[0077]

[0078] Where D i To guarantee

[0079] Step 3.4: Combine the control matrix α of equation (1) in step 1.1 with equation (5) in step 3.3 to obtain equation (6), which is as follows:

[0080]

[0081] Step 3.5: Matrix the equation (6) in step 3.4 to obtain equation (7), which is as follows:

[0082]

[0083] Among them, α table is the vector composed of all α.

[0084] Step 3.6: Use equation (7) in step 3.5 to construct the objective function with constraints, as shown in equation (8). Then, the vector α can be obtained by solving the objective function and reconstructing the complete control matrix α. Equation (8) is as follows:

[0085]

[0086] Among them, α(C) = k is a constraint, where C is a known pixel, which can be provided by the three-color constraint map Tmap in 2.3. The black area and white area of ​​Tmap are known pixels, representing the background and foreground respectively, and their corresponding k are 0 and 1 respectively. The model has sufficient constraints to solve, so the α value of the unknown pixel (gray area) can be solved to obtain the completed vector α, and the complete control matrix α can be reconstructed.

[0087] Step 4.1: The essential difference between the real target and the sparse interference source is that they have different affinity with the surroundings. This also shows that the α value of the mixed area around the sparse area is different. For the real target in the sparse component, the affinity between the surrounding pixels and the target is low, so its mixed area is more characterized by the background, and the α value is close to 0. For the mixed area of ​​the interference source, they may all be part of the interference source, so the α value will be high. Therefore, in this step, first use the three-color constraint map Tmap obtained in step 2.3 to calculate the unknown area template M through formula (9): un The formula (9) is as follows:

[0088]

[0089] Step 4.2: Use the control matrix α obtained in step 3.6 and the neighbor template M obtained in step 4.1 un , calculate the sparse area Sparse by formula (10) k Affinity attribute Af with its neighbors k , formula (10) is as follows:

[0090]

[0091] Step 4.3: Add the affinity attribute Af obtained in step 4.2 k , the inhibition factor ξ is calculated by formula (11) k , formula (11) is as follows:

[0092] ξ k =tanh(Af k -K)#(11)

[0093] Where K represents the affinity attribute threshold, tanh(*) is the activation function, and tanh(x) = (e x -e -x) / (e x +e -x ). In this example, K=15.

[0094] Step 4.4: Use the inhibition factor ξ obtained in step 4.3 k , get the sparse component image f from step 2.1 S The interference source is suppressed and the real target is enhanced by formula (12), and the final target detection result f is obtained. D . Formula (12) is as follows:

[0095]

[0096] Among them, C is used to control the interference source suppression degree or the target enhancement degree, b represents the sparse component image f S In this example, C=1.2 and b=100.

Claims

1. An anti-interference infrared small target detection method, characterized in that: The steps include: Step 1: Express the infrared small target image in a complex environment and proceed to step 2; Step 1.1: The infrared image is considered as a linear superposition of a foreground image with a target and a background image containing both interference sources and background. The weights of the foreground image and the background image in the linear superposition in the global image are determined by the control matrix Control, as shown in formula (1): ； in, 、 、 Respectively represent the input infrared image, the unknown foreground image with the target, and the unknown background image. is an unknown control matrix of the same size as the original image, and the range of values ​​inside is ,constraint ,in is a known pixel, Indicates the corresponding constraint value, by setting the constraint , then other unknown solutions can be solved; Step 2: Input the original infrared image to extract the sparse component image, and construct a three-color constraint graph as the constraint for model solution, and then proceed to step 3; Step 3: Solve the model and obtain the control matrix with affinity relationship , go to step 4; The step 3 comprises the following steps: Step 3.1: Define affinity relationship to describe the relationship between pixels, as shown in formula (3): ； in, Represents pixel Follow the pixels affinity relationship, , Represents pixel points The gray value and position of , Represents pixel points The gray value and position of Pixel The local structure of Pixel The local structure of Control parameters to maintain , is the weight coefficient; Step 3.2: The affinity relationship between multiple pixels is represented by an affinity relationship matrix as shown in formula (4), which is as follows: in, Indicates the number of pixels; Step 3.3: The infrared image is relatively smooth as a whole, so it can be assumed that each pixel can be represented by its neighboring pixels, as shown in formula (5): ； Step 3.4: Add the Chinese formula from step 1.1 The control matrix The formula in step 3.3 Combining formula (6): ； Step 3.5: Repeat the steps in step 3.

4. Matrixization of , we get formula (7): ； in, Table for all The vector formed; Step 3.6: Use equation (7) in step 3.5 to construct the objective function with constraints, as shown in equation (8), and then solve the objective function to obtain the vector , and reconstruct the complete control matrix , formula (8) is as follows: ； in, is a constraint, where For known pixels, we can use the three-color constraint map in 2.3 supply, The black and white areas of are known pixels, representing the background and foreground, respectively. are 0 and 1 respectively, the model solves the constraints sufficiently, so the unknown pixels can be Solve the value to get the completed vector , and can be reconstructed to obtain a complete control matrix ; Step 4: Use the control matrix obtained in step 3 A sparse interference source suppression method is constructed to suppress and enhance the interference sources and targets in the sparse component image of step 2, respectively, to obtain the final target image.

2. The anti-interference infrared small target detection method according to claim 1 is characterized in that: The step 2 comprises the following steps: Step 2.1: Consider the sparse components as targets With the interference source of the same sparse nature The superposition of sparse components is obtained by low-rank sparse decomposition method. , which will be used to provide constraints later; Step 2.2: The sparse component image obtained in step 2.1 Perform global threshold segmentation to obtain a sparse template , where the value of the pixel in the sparse area is 1, and the other values ​​are 0; Step 2.3: For the sparse template obtained in step 2.2 , construct the three-color constraint graph through formula (2) The three-color constraint graph is used to represent the equation in step 1.

1. Provide constraint values. The values ​​of sparse areas and background areas are known areas, with values ​​of 1 and 0, representing foreground and background respectively. The target area is regarded as an unknown area, represented by 0.5, which means that the area is a mixed superposition area of ​​foreground and background. The unknown area can be identified through the known area. The value is solved, formula (2) is as follows: ； in, represents morphological expansion, It is a three-color constraint map. Since it contains three values ​​0, 0.5, and 1, which correspond to the three colors black, gray, and white in the image, it is called a three-color constraint map, which respectively represent the background area, unknown area, and foreground area.

3. The anti-interference infrared small target detection method according to claim 2 is characterized in that: The step 4 comprises the following steps: Step 4.1: Use the three-color constraint graph obtained in step 2.3 , calculate the unknown area template through formula (9) , formula (9) is as follows: ； Step 4.2: Use the control matrix obtained in step 3.6 and the unknown region template obtained in step 4.1 , the sparse area is calculated by formula (10) Affinity properties with its neighbors , formula (10) is as follows: ； Step 4.3: Add the affinity attribute obtained in step 4.2 , the inhibition factor is calculated by formula (11) , formula (11) is as follows: ； in, represents the affinity attribute threshold, is the activation function, ; Step 4.4: Use the inhibitory factor obtained in step 4.3 , get the sparse component image in step 2.1 The interference source is suppressed and the real target is enhanced by equation (12), and the final target detection result is obtained. ; Formula (12) is as follows: ； in, It is used to control the interference source suppression degree or target enhancement degree. Represents sparse component images The value of the background area.

Images