An Airborne Infrared Small Target Detection Method Based on Target Polarization Feature Reconstruction

Abstract

The present invention discloses an airborne infrared small target detection method based on target polarization feature reconstruction, belonging to the technical field of infrared target detection, and comprising the following steps: S1, acquiring polarization images of three channels of 0°, 60°, and 120°; S2, calculating the corresponding Stokes vectors according to the polarization images of the three channels; S3, calculating gradient features for suppressing edges according to the Stokes vectors, and performing saliency feature extraction; S4, acquiring the final gradient features for suppressing background edges; S5, acquiring the reconstructed polarization features; S6, calculating the local variance stability level; S7, acquiring a weight function to obtain the final detection result. By adopting the above-mentioned airborne infrared small target detection method based on target polarization feature reconstruction, the present invention can accurately detect small targets, improve the suppression efficiency of background clutter, and has good detection performance under both simple sky backgrounds and complex sky backgrounds.

Description

Technical Field

[0001] The present invention relates to the technical field of infrared target detection, and in particular to an airborne infrared small target detection method based on target polarization feature reconstruction. Background Art

[0002] At present, the detection methods for small and weak targets under clouds can be mainly divided into two categories: single-frame image-based detection methods and frame-sequence-based detection methods. The frame-sequence-based methods utilize the state and motion trajectory of the target to achieve target detection. Currently, relatively advanced multi-frame image detection methods include Markov random field, spatio-temporal filtering, and spatio-temporal contrast filtering, etc. These methods perform well in removing complex background interference, but usually rely on prior information of the target or background, and at the same time require a certain correlation between frame sequences.

[0003] The target detection methods based on infrared polarization imaging mainly distinguish the differences between the target and the background through the degree of linear polarization and polarization angle, highlighting the characteristics of the target to be detected. For example, Forssell, Y. Gronau, etc. respectively established models using the degree of linear polarization to achieve the detection of buried landmines and cars in the field; Zhao Yongqiang et al. realized the detection of vehicle targets under low signal-to-noise ratio and complex backgrounds by designing a method of polarization-weighted local contrast; Ratlif et al. verified that infrared polarization imaging has stronger detection ability for targets under complex sky backgrounds; Hao et al. combined polarization with Top-Hat transform to achieve the detection of weak targets in the air; Song Minmin et al. carried out a large number of experiments and verified that infrared polarization imaging has a significant effect on suppressing sea clutter.

[0004] Although small target detection under the sky background can be well achieved through infrared imaging, when facing a scene with a low contrast in the target neighborhood, there are still many difficulties and challenges. Polarization information can describe the physical and chemical properties of the target and the background itself, and can still reflect the polarization information difference between the target and the background even in a low-contrast scene, improving the contrast between the target and the background, and providing a new idea for small target detection under complex sky backgrounds. At the same time, although the infrared polarization imaging technology improves the local contrast in the target neighborhood, it also brings an increase in the contrast of the background highlight edges, bringing new difficulties and challenges to the target detection work. Summary of the Invention

[0005] The object of the present invention is to provide an airborne infrared small target detection method based on target polarization feature reconstruction, which can accurately achieve the detection of small targets, improve the suppression efficiency of background clutter, and has good detection performance both in simple sky backgrounds and complex sky backgrounds.

[0006] To achieve the above object, the present invention provides an airborne infrared small target detection method based on target polarization feature reconstruction, including the following steps:

[0007] Step S1: Obtain the polarization images of three channels at 0°, 60°, and 120°;

[0008] Step S2: Calculate the corresponding Stokes vectors based on the polarization images of the three channels;

[0009] Step S3: Calculate the gradient features for suppressing edges based on the Stokes vectors, and perform significant feature extraction;

[0010] Step S4: Obtain the final gradient features for suppressing background edges according to Step S3;

[0011] Step S5: Obtain the reconstructed polarization features;

[0012] Step S6: Calculate the local variance stability level;

[0013] Step S7: Obtain the weight function to get the final detection result.

[0014] Preferably, the formula for calculating the gradient features for suppressing edges based on the Stokes vectors in Step S3 is as follows:

[0015]

[0016] where Gm(I)(x,y) represents the gradient feature for suppressing edges calculated from the Stokes vector, and I represents Stokes vector S1 or Stokes vector S2; respectively represent the directional derivatives of (x,y) along α i , β i directions; where α i , β i satisfy β i =α i +π / 2.

[0017] Preferably, the formula for significant feature extraction in Step S3 is as follows:

[0018]

[0019] where SP(I) represents the significant feature; ||·||2 represents the 2-norm; mean ω (I) represents the mean value of the local region ω of the image; represents the image after smoothing filtering with a Gaussian filter of size ω; the size of ω is 5×5, and the standard deviation σ of the Gaussian filtering is 1 - 3.

[0020] Preferably, the significant features obtained through step S3 are used to calculate the gradient weighting of Stokes vectors S1 and S2 using formula (1) respectively, and the final gradient feature image GT for suppressing background edges is obtained:

[0021]

[0022] Preferably, the suppression PF of the background polarization feature is achieved through the gradient feature image GT for suppressing background edges, and its definition is:

[0023] PF = GT·S1 + GT·S2 (4);

[0024] Then the calculation formula for the reconstructed polarization feature PF is as follows:

[0025]

[0026] Among them, RDoLP represents the reconstructed polarization feature; S0 represents the first component in the Stokes vector, representing the intensity image of the scene.

[0027] Preferably, the local variance stability level is calculated through the local stability VSL of variance, and its definition is:

[0028]

[0029] Among them, represents the local variance stability level; i represents the pixel number in the nested window, i = 0 to 8, when i = 0 it represents the inner window of the nested structure, and i = 1 - 8 represents the outer window of the nested structure; represents the set of all pixels in the pixel numbered i in the nested window at the jth scale; represents the set of the first N largest pixels in the pixel.

[0030] Preferably, for any window, the weight W j (x, y) of the central image is expressed as:

[0031]

[0032] Among them, represents the variance stability level of the inner window; represents the standard deviation of the VSL between the pixels in the outer window; ζ is a positive number, and its value is 0.5; TED j represents the minimum value of the difference in variance stability level between the central pixel and the surrounding pixels, and its definition is:

[0033]

[0034] Then the final target saliency image SupD in the final detection result is expressed as:

[0035] SupD = max{W j ·RDoLP}(9).

[0036] Therefore, the present invention adopts the above-mentioned method for detecting small infrared targets in the air based on target polarization feature reconstruction. By analyzing a large amount of infrared polarization data of the sky background, the differences between the polarization features of the target and the background are analyzed. Specifically, the polarization feature images S1 and S2 of the target area are both isotropic, while the gradient features of the background edge area are all anisotropic. Therefore, by using this characteristic to constrain the polarization Stokes parameters, the polarization features of the scene are further reconstructed, effectively suppressing the background information and highlighting the target.

[0037] At the same time, this method realizes the design of the weight function by defining the variance stability level, and realizes the detection of small targets under complex sky backgrounds. Compared with other methods, the signal-to-clutter ratio gain (SCRG) of the detection result is increased by about 40%; in different types of scenes (large continuous clouds, complex clouds), the experimental results show that this method has better robustness.

[0038] Next, through the attached drawings and embodiments, the technical solution of the present invention will be further described in detail. Brief Description of the Drawings

[0039] Figure 1 is a flowchart of an embodiment of the method for detecting small infrared targets in the air based on target polarization feature reconstruction of the present invention;

[0040] Figure 2 is the original image (left) and its three-dimensional intensity distribution diagram (right) of Scene 1 in the embodiment of the present invention;

[0041] Figure 3 is the detection result (left) and its three-dimensional intensity distribution diagram (right) of Scene 1 in the embodiment of the present invention;

[0042] Figure 4 is the original image (left) and its three-dimensional intensity distribution diagram (right) of Scene 2 in the embodiment of the present invention;

[0043] Figure 5 is the detection result (left) and its three-dimensional intensity distribution diagram (right) of Scene 2 in the embodiment of the present invention.

[0044] Figure 6 is a schematic diagram of the nested structure in steps S6 and S7 of the embodiment of the present invention, Figure 6 where (a) represents the reconstructed polarization feature; Figure 6 where (b) represents the residual edge window; Figure 6 where (c) represents the target area window. Detailed Embodiment

[0045] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0046] Unless otherwise defined, the technical terms or scientific terms used in the present invention shall have the ordinary meanings understood by those of ordinary skill in the field to which the present invention pertains.

[0047] Embodiment 1

[0048] As Figure 1 shown, the present invention provides an airborne infrared small target detection method based on target polarization feature reconstruction, including the following steps:

[0049] Step S1: Obtain polarization images of three channels of 0°, 60°, and 120°;

[0050] Step S2: Calculate the corresponding Stokes vector based on the polarization images of the three channels;

[0051] Step S3: Calculate the gradient feature for suppressing edges based on the Stokes vector, and perform significant feature extraction;

[0052] Step S4: Obtain the final gradient feature for suppressing background edges according to Step S3;

[0053] Step S5: Obtain the reconstructed polarization feature;

[0054] Step S6: Calculate the local variance stability level;

[0055] Step S7: Obtain the weight function to obtain the final detection result.

[0056] Preferably, the formula for calculating the gradient feature for suppressing edges based on the Stokes vector in Step S3 is as follows:

[0057]

[0058] where Gm(I)(x,y) represents the gradient feature for suppressing edges calculated from the Stokes vector, and I represents Stokes vector S1 or Stokes vector S2; respectively represent the directional derivatives of (x,y) along α i and β i directions; where α i and β i satisfy β i = α i + π / 2.

[0059] The directional derivative can characterize the rate of change of pixels in an image along a certain direction. Therefore, in a large-area continuous background region, the modulus of the directional derivative in each direction is a very small positive value; at the edge of the cloud background, the modulus of the directional derivative is only large in the gradient change direction and the opposite direction; while in the target region, the modulus of the directional derivative is greater than 0 in each direction and is evenly distributed. Therefore, the gradient obtained by formula (1) shows a high gradient in the target region; while at the background edge, since there must be a direction with a very small directional derivative in each set of orthogonal directions, it shows a low gradient.

[0060] The formula for extracting the significant features in step S3 is as follows:

[0061]

[0062] where SP(I) represents the significant feature; ||·||2 represents the 2-norm; mean ω (I) represents the mean value of the local region ω of the image; represents the image after smoothing filtering with a Gaussian filter of size ω; the size of ω is 5×5, and the standard deviation σ of the Gaussian filtering is 1 to 3.

[0063] In step S4, the significant features obtained in step S3 are used to weight the gradients obtained by calculating Stokes vectors S1 and S2 using formula (1) respectively, to obtain the final gradient feature image GT for suppressing the background edge:

[0064]

[0065] In step S5, the background polarization feature suppression PF is achieved through the gradient feature image GT for suppressing the background edge, and its definition is:

[0066] PF = GT·S1 + GT·S2 (4);

[0067] Then the calculation formula for the reconstructed polarization feature PF is as follows:

[0068]

[0069] where RDoLP represents the reconstructed polarization feature; S0 represents the first component in the Stokes vector, representing the intensity image of the scene.

[0070] Compared with the original linear polarization degree image, the reconstructed polarization feature suppresses a large number of non-target features, and only retains the polarization features of the target and a small amount of residual background edge features.

[0071] In step S6, although most of the background edges are suppressed by the reconstructed polarization feature RDolp, there are still some weak edges and sparse noises that may affect the accuracy of target detection. Therefore, a two-layer nested window optimization weighting function is introduced to further suppress the weak edge background while retaining the real target. To accurately estimate the weight, the local variance stability level VSL is calculated through the local stability of variance, which is defined as:

[0072]

[0073] where, represents the local variance stability level; i represents the pixel number in the nested window, i = 0~8, when i = 0, it represents the inner window of the nested structure, and i = 1-8 represents the outer window of the nested structure; represents the set of all pixels in the pixel numbered i in the nested window at the jth scale; represents the pixel (i.e., Figure 6 a total of nine pixels numbered 0-8 in) the set of the first N largest pixels.

[0074] Considering that the size of small targets cannot be predicted in advance in the actual scenario, five different scales of pixels are set in the algorithm to cover the vast majority of small target sizes, namely 5×5, 7×7, 9×9, 11×11, and 13×13. The two-layer nested sliding window consists of 3×3 pixels, which are divided into a central pixel (numbered 0) and surrounding pixels (numbered 1-8), as shown in Figure 6 shown, where each pixel contains N×N pixels, and N should be slightly larger than the size of the target.

[0075] In step S7, for any window, the weight W of the central image j (x,y) is expressed as:

[0076]

[0077] where, represents the variance stability level of the inner window; represents the standard deviation of the VSL between the pixels of the outer window; to prevent the denominator from being zero, ζ is set to a very small positive number, and its value is 0.5; TED j represents the minimum value of the difference in variance stability level between the central pixel and the surrounding pixels, which is defined as:

[0078]

[0079] The weight obtained through formula (8) has the following characteristics: First, when the center of the inner window is a real target, is larger, the pixels of the outer window are the background area, and the pixel distribution is relatively stable, satisfying Smaller, TED j Larger, so the weight of the central pixel at this time is very large; when the center of the inner window is located at the continuous background, is approximately equal to 0, while TED j must be a finite positive value. Therefore, regardless of whether all the pixels in the outer window are at the continuous background, the weight at the continuous background is close to 0; finally, when the center of the inner window is located at the background edge, due to the continuity of the edge, among the surrounding pixels, there may be some pixels whose VSL satisfies TED j is a positive value close to 0. Similarly, is also a finite value, that is, the background edge weight value is extremely small. In summary, the weight function defined based on the polarization local variance characteristic can effectively suppress the residual background of the image and highlight the real target.

[0080] In the final detection result, the final target saliency image SupD is expressed as:

[0081] SupD = max{W j ·RDoLP}(9).

[0082] Figure 2 , Figure 3 Scene 1 in Figure 4 , Figure 5 shows a single simple cloud background, Figure 2 , Figure 3 Scene 2 in Figure 4 , Figure 5 shows the target detection effect under a complex cloud background. By comparing the three-dimensional intensity distribution maps in Figure 3 , Figure 5 and the three-dimensional intensity distribution maps in

[0083] It can be seen that the target information in

[0084] This invention can not only effectively suppress noise and extract target information in a simple background, but also effectively highlight the target in a complex background and low-contrast scene, greatly reducing the background residue.

[0085] It should be noted that the content not elaborated in detail in this invention is all prior art and is well-known to those skilled in the art.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions of the present invention or make equivalent replacements, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. An airborne infrared small target detection method based on target polarization feature reconstruction, characterized in that Including the following steps: Step S1: Obtain the polarization images of three channels at 0°, 60°, and 120°; Step S2: Calculate the corresponding Stokes vectors based on the polarization images of the three channels; Step S3: Calculate the gradient features for suppressing edges based on the Stokes vectors and perform significant feature extraction. The formula for calculating the gradient features for suppressing edges based on the Stokes vectors is as follows: （1）； Among them, represents the gradient feature of the Stokes vector calculation for suppressing the edge I represents the Stokes vector S1 or the Stokes vector S2; , respectively represent along , the directional derivatives in the directions; where , satisfy ; The formula for significant feature extraction is as follows: （2）； Among them, represents the significant feature; represents the 2-norm; represents the local region of the image mean value; represents the region is the image after smooth filtering by a Gaussian filter with a size of; The size of is 5 5, and the standard deviation of the Gaussian filter is 1 to 3; By using the obtained significant features, the gradient weighting obtained by calculating the Stokes vectors S1 and S2 respectively using formula (1) is used to obtain the final gradient feature image for suppressing background edges. GT : （3）； Step S4: Obtain the final gradient features for suppressing background edges according to Step S3; Step S5: Obtain the reconstructed polarization features; Step S6: Calculate the local variance stability level; Step S7: Obtain the weight function to get the final detection result.

2. The method for detecting small infrared targets in the air based on target polarization feature reconstruction according to claim 1, wherein: By suppressing the gradient feature image of the background edge GT to achieve the suppression of the background polarization feature PF which is defined as: （4）； The reconstructed polarization characteristics PF The calculation formula is as follows: （5）； Among them, represents the reconstructed polarization feature; represents the first component in the Stokes vector and represents the intensity image of the scene.

3. The air infrared small target detection method based on target polarization feature reconstruction according to claim 2, characterized in that: Calculate the local variance stability level through the local stability of variance VSL, which is defined as: （6）； Among them, represents the local variance stability level; i represents the pixel number in the nested window, , i= when it is 0, it represents the inner window of the nested structure ，i= 1 - 8 represents the outer window of the nested structure; represents the th scale, and is the set of all pixels in the pixel numbered in the nested window; represents the set of the first largest pixels in the pixel.

4. The air infrared small target detection method based on target polarization feature reconstruction according to claim 3, wherein: For any window, the weight of the central image is expressed as: （7）； Among them, represents the variance stability level of the inner window; represents the standard deviation of the VSL between the pixels of the outer window; is a positive number, and its value is 0.5; represents the minimum value of the difference in variance stability level between the central pixel and the surrounding pixels, and its definition is: （8）； The final target saliency image in the final detection result is expressed as: （9）。

Images