Infrared small target detection method based on hierarchical feature fusion

Abstract

The invention provides an infrared small target detection method based on hierarchical feature fusion. As the features of the infrared weak target are limited, exploration and construction of additional target features are crucial, so that the weak target can effectively distinguish an accurate target area from numerous interferences. The method combines global, global-local and local features to perform fusion from different levels of target features. The method comprises the following steps: firstly, constructing improved global contrast measurement to quickly extract candidate target points; secondly, weight functions of candidate target points are designed in combination with local and global view angles; and finally, local features of the candidate target points are further mined from a local perspective. The method aims to solve the problem of infrared weak target detection of an infrared search and tracking system in a complex scene, and is suitable for the related fields of infrared security and protection and the like.

Description

Technical Field

[0001] The invention relates to the technical field of image processing, and in particular to a method for detecting small infrared targets based on hierarchical feature fusion. Background Art

[0002] Infrared search and tracking systems have attracted attention from all over the world for their advantages of all-day detection and high angular accuracy in the high-precision identification and long-distance detection of low-altitude infrared small targets. As one of the key technologies of infrared search and tracking systems, infrared small target detection has always been a hot topic of concern for many scholars. However, compared with visible light images, infrared images have a narrow grayscale dynamic range, low signal-to-noise ratio, low spatial resolution, and multiple noises introduced into the system. In addition, due to long-distance imaging, the target image size is very small and lacks shape and texture information. It is easy for infrared small targets to be submerged in clutter and difficult to detect, which poses a huge challenge to the detection performance of the detection algorithm. In summary, infrared small target detection under complex backgrounds is still a major challenge for current infrared search and tracking systems.

[0003] Methods aimed at improving the accuracy and real-time performance of infrared weak target detection continue to emerge. The filtering-based method suppresses the image background or enhances the target through the prior knowledge of the target structure, thereby achieving the purpose of target detection. Its advantages are small computational complexity and low complexity, but it can only suppress the uniform background to a certain extent, and cannot solve the problems of low detection rate and poor robustness under complex backgrounds. The method based on low-rank decomposition uses the non-local autocorrelation of the background to transform target detection into an optimization problem with a low-rank matrix and a sparse matrix. These methods converge through global optimization, which takes a long time and relies heavily on the designer to provide a large number of hyperparameters. The false alarm rate is high in complex scenes, and the false alarm rate of dark targets is also high. The HVS-based method starts from a local perspective, manually designs feature descriptors, and uses the difference between the local area of ​​the target and the background to highlight the target. Its advantages are simple principle, more sufficient expression of weak target features, and more flexibility. However, there are many interferences in complex backgrounds. Among them, target similarity interference will also be enhanced, making it difficult to suppress false alarms, and the algorithm detection performance still needs to be improved.

[0004] In summary, each category of methods has its own unique perspective and thus shows obvious advantages in different performance indicators. Summary of the invention

[0005] The purpose of the present invention is to provide a method for infrared small target detection based on hierarchical feature fusion, which is used to solve the problem of high false alarm rate when target information is limited. The method utilizes target feature fusion at three levels: global, global-local, and local, and performs infrared weak small target detection in two stages from coarse to fine. The method of the present invention is aimed at the problem of infrared weak small target detection in infrared search and tracking systems under complex scenes, and is particularly suitable for related fields such as infrared security.

[0006] The technical solution of the present invention is as follows:

[0007] A method for infrared small target detection based on hierarchical feature fusion, characterized in that it comprises the following steps:

[0008] A. The original image is median filtered and then dilated:

[0009]

[0010] Among them, f med is the image after median filtering, ΔB is the structural element, is the expansion operation; I dilation is the image after the dilation operation, and then a global contrast is performed between it and the original image to obtain the global feature map:

[0011]

[0012] IGCM is a global contrast saliency map, I raw is the original infrared image;

[0013] B. Coarsely extract candidate target points (CTP) through adaptive threshold:

[0014] Th=m IGCM +k*std IGCM . (3)

[0015] Th is the segmentation threshold, m IGCM and std IGCM They are the mean and standard deviation of IGCM, k is a hyperparameter, usually ranging from 0 to 3;

[0016] C. Perform density peak search for candidate target points to obtain global-local features and use them as weighting functions (Density Peaks Weighting, DPW):

[0017]

[0018] The nearest correlation distance is the minimum Euclidean distance between the candidate target point i and the pixel point j with higher density than it, d ij is the Euclidean distance between two pixels, is the coordinate of candidate target point i, (x j ,y j ) is the coordinate of pixel j; is the density peak value of candidate target point i; DPW is the density peak weight function;

[0019] D. Extract local features (LF) based on the local window template for the candidate target point set:

[0020]

[0021] Directional Suppression Operator (DSO), m T and are the average values ​​of the pixels in the center area of ​​the candidate target point and the pixels related to the edges in eight directions;

[0022]

[0023] LF=DSO*LC. (10)

[0024] LC is the local contrast, and LF is the local feature obtained by weighting DSO and LC;

[0025] E. Finally, integrate target information at different levels to confirm the real target:

[0026] HFF=DPW*LF. (11);

[0027] HFF is the hierarchical feature fusion result, which is obtained by weighting DPW and LF.

[0028] Compared with the prior art, the invention is beneficial in that it constructs multi-level features of the target, enriches the target information, and improves the detection performance. The two-stage operation has a lower false alarm rate in processing complex scenes. The main contributions of the invention are:

[0029] 1. Through global contrast enhancement, candidate target points can be quickly obtained, reducing a large number of invalid calculation operations.

[0030] 2. Through density peak search, construct global-local features as weight functions to effectively reduce false alarm rates

[0031] 3. Combining the multi-level attributes of the target, the detection performance of the target is enhanced. The present invention is robust to the target size and effectively improves the detection rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1It is a flowchart of the implementation process of the present invention.

[0033] Figure 2 is the original infrared image, and the box in the figure is the target to be detected.

[0034] Figure 3 It is the global contrast saliency map IGCM, and the box in the figure is the target to be detected.

[0035] Figure 4 It is a local template window, the inner area is the target area, and the outer area is divided into a cross shape. DETAILED DESCRIPTION

[0036] The present invention discloses a method for infrared small target detection based on hierarchical feature fusion. Figure 1 As shown: It includes the following steps:

[0037] A. The original image is median filtered and then dilated:

[0038]

[0039] Among them, f med is the image after median filtering, ΔB is the structural element, is the expansion operation; I dilation is the image after the dilation operation, and then a global contrast is performed between it and the original image to obtain the global feature map:

[0040]

[0041] IGCM is a global contrast saliency map, I raw is the original infrared image;

[0042] B. Coarsely extract candidate target points (CTP) through adaptive threshold:

[0043] Th=m IGCM +k*std IGCM . (3)

[0044] Th is the segmentation threshold, m IGCM and std IGCM They are the mean and standard deviation of IGCM, k is a hyperparameter, usually ranging from 0 to 3;

[0045] C. Perform density peak search for candidate target points to obtain global-local features and use them as weighting functions (Density Peaks Weighting, DPW):

[0046]

[0047] The nearest correlation distance is the minimum Euclidean distance between the candidate target point i and the pixel point j with higher density than it, d ij is the Euclidean distance between two pixels, is the coordinate of candidate target point i, (x j ,y j ) is the coordinate of pixel j; is the density peak value of candidate target point i; DPW is the density peak weight function;

[0048] D. Extract local features (LF) based on the local window template for the candidate target point set:

[0049]

[0050] Directional Suppression Operator (DSO), m T and are the average values ​​of the pixels in the center area of ​​the candidate target point and the pixels related to the edges in eight directions;

[0051]

[0052] LF=DSO*LC. (10)

[0053] LC is the local contrast, and LF is the local feature obtained by weighting DSO and LC;

[0054] E. Finally, integrate target information at different levels to confirm the real target:

[0055] HFF=DPW*LF. (11);

[0056] HFF is the hierarchical feature fusion result, which is obtained by weighting DPW and LF.

[0057] The technical scheme in the embodiment of the present invention will be described in detail below in conjunction with the accompanying drawings in the embodiment of the present invention. Several parameters are involved, and these parameters need to be adjusted according to the specific processing environment to achieve good performance. Based on the embodiment of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work belong to the scope of protection of the present invention.

[0058] The test images used in the present invention come from a dataset of small and weak aircraft target detection and tracking in infrared sequence images of the National University of Defense Technology.

[0059] Simulation environment: Matlab2024b;

[0060] Test image: medium-wave infrared image, resolution 256×256, background near the ground;

[0061] Target information: UAV.

[0062] Test steps:

[0063] (1) Figure 1 As shown, first, the original infrared image Figure 2 Input, obtained after global contrast enhancement Figure 3 .

[0064] (2) Figure 3 Adaptive threshold segmentation is used to obtain candidate target points.

[0065] (3) Taking the candidate target point as input, the density peak is globally searched to obtain the density peak construction weight function.

[0066] (4) Taking the candidate target point as input, design local feature enhancement for its local area. The local enhancement template window is as follows: Figure 4 .

[0067] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A method for infrared small target detection based on hierarchical feature fusion, characterized in that: It includes the following steps: A. The original image is median filtered and then dilated: Among them, f med is the image after median filtering, ΔB is the structural element, is the expansion operation; I dilation is the image after the dilation operation, and then a global contrast is performed between it and the original image to obtain the global feature map: IGCM is a global contrast saliency map, I raw is the original infrared image; B. Coarsely extract candidate target points (CTP) through adaptive threshold: Th=m IGCM +k*std IGCM . (3) Th is the segmentation threshold, m IGCM and std IGCM They are the mean and standard deviation of IGCM, k is a hyperparameter, usually ranging from 0 to 3; C. Perform density peak search for candidate target points to obtain global-local features and use them as weighting functions (Density Peaks Weighting, DPW): The nearest correlation distance is the minimum Euclidean distance between the candidate target point i and the pixel point j with higher density than it, d ij is the Euclidean distance between two pixels, is the coordinate of candidate target point i, (x j ,y j ) is the coordinate of pixel j; is the density peak value of candidate target point i; DPW is the density peak weight function; D. Extract local features (LF) based on the local window template for the candidate target point set: Directional Suppression Operator (DSO), m T and are the average values ​​of the pixels in the center area of ​​the candidate target point and the pixels related to the edges in eight directions; LF=DSO*LC. (10) LC is the local contrast, and LF is the local feature obtained by weighting DSO and LC; E. Finally, integrate the target information at different levels to confirm the real target: HFF=DPW*LF. (11); HFF is the hierarchical feature fusion result, which is obtained by weighting DPW and LF.

Images