An infrared small target detection method based on density peak search and local features

Abstract

The application discloses an infrared small target detection method based on density peak value search and local features, which comprises the following steps: S1, preprocessing an image through Gaussian difference filtering and suppressing high-brightness clutter; S2, determining the position of a candidate target in the preprocessed image through a density peak value global search method; S3, performing local contrast enhancement gradient characteristic on the candidate target and suppressing background clutter; S4, calculating the multi-directional gradient characteristics of each candidate target through a Facet model; S5, capturing the gradient characteristics in eight directions of different size targets with the candidate target point as the center; S6, suppressing the gradient characteristics of the background clutter by making difference to the regions in the symmetric direction of the candidate target point and using the standard deviation weighted gradient response value of the symmetric region difference; and S7, obtaining the infrared weak small target through an adaptive threshold segmentation method on the gradient characteristic response value of the candidate target. According to the application, the detection rate is high in various scenes, and the operation process is relatively simpler.

Description

Technical Field

[0001] This invention relates to the technical field of infrared small target detection, and in particular to an infrared small target detection method based on density peak search and local features. Background Technology

[0002] Currently, infrared imaging systems are widely used in fault diagnosis, target detection, and video surveillance. As a key technology in infrared imaging systems, infrared small target detection remains a research challenge in this field. Due to the distance of the target from the detection system and severe atmospheric attenuation, true infrared small targets occupy only a few pixels in the image field, resulting in a lack of local structural information. Furthermore, they are often submerged in complex background clutter and heavy noise, leading to a low signal-to-clutter ratio, making it difficult to distinguish small targets from background clutter and noise. Therefore, separating small targets from complex backgrounds in infrared images is a challenging problem. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the present invention aims to provide an infrared small target detection method based on density peak search and local features, which exhibits a high detection rate in various scenarios and has a relatively simpler computational process. To achieve the above-mentioned objectives and other advantages of the present invention, an infrared small target detection method based on density peak search and local features is provided, comprising:

[0004] S1. Preprocess the image using Gaussian difference filtering to suppress high-brightness noise;

[0005] S2. The location of candidate targets is determined in the preprocessed image using the density peak global search method;

[0006] S3. Perform local contrast analysis on candidate targets to enhance gradient characteristics and suppress background clutter;

[0007] S4. Calculate the multi-directional gradient characteristics of each candidate target using the Facet model;

[0008] S5. Using the candidate target point as the center, set different thresholds for the size range of different regions, optimize the division scheme of the surrounding regions, and capture the gradient characteristics of targets of different sizes in eight directions.

[0009] S6. The gradient characteristics of background clutter are suppressed by subtracting regions along the symmetrical direction of the candidate target point and the gradient response value is weighted by the standard deviation of the symmetrical region difference.

[0010] S7. The infrared weak targets are obtained by using an adaptive threshold segmentation method on the gradient characteristic response value of the candidate targets.

[0011] Preferably, in step S1, a low-pass filtered image is obtained by convolving the image with a Gaussian function. By optimizing and selecting different parameters, the two Gaussian filtered image results under different parameters are subtracted to obtain a Gaussian difference filtered image, which suppresses high-brightness clutter near the target and obtains a Gaussian difference filtered image.

[0012] Preferably, in step S2, by defining the density and distance of each pixel in the infrared image, the density peak of each pixel in the image is obtained, the density peak of each pixel is arranged from largest to smallest, and the pixel points with the largest density peak after optimization are selected as candidate target points by iteratively optimizing the threshold comparison.

[0013] Preferably, in step S3, a 3×3 image patch slider is used to slide a local image patch centered on the candidate target point. By comparing the grayscale difference between the central region and the surrounding region of the slider, the gradient characteristics of the selected real target point are enhanced and optimized, and background clutter is suppressed.

[0014] Preferably, in step S4, the multi-directional gradient of each point is calculated by using the Facet model. The Facet model considers that any small neighborhood of a pixel in the image can be fitted by a polynomial function. By optimizing the polynomial function, the key parameters in the polynomial can be directly obtained by performing convolution operations using a fixed filter. Based on the neighborhood polynomial, the gradient characteristics of each pixel in different directions in the neighborhood can be quickly obtained.

[0015] Preferably, the adaptive threshold in step S7 can be calculated based on the mean and standard deviation of the fused gradient response values. By processing the gradient response values ​​of candidate target points according to a certain threshold and cyclic segmentation calculation, non-real target points under a certain threshold are filtered out, and the precise location of the infrared weak target is obtained.

[0016] Compared with existing technologies, the advantages of this invention are as follows: It improves the accuracy of density peak search by adding Gaussian difference filtering, enhances gradient characteristics using local contrast, employs the Facet model to efficiently and accurately calculate the multi-directional gradient features of each point, and proposes an effective gradient feature extraction framework. This framework effectively extracts the gradient response values ​​of candidate target points of different sizes while suppressing background clutter. Furthermore, it uses the standard deviation of symmetrical region differences to weight the gradient characteristics of candidate targets, further enhancing the gradient response values ​​of the targets. The gradient response values ​​are then subjected to adaptive threshold segmentation to obtain the detection results, thereby achieving infrared weak target detection based on frequency domain residual analysis. By utilizing the local and global feature information of infrared small targets, it exhibits good robustness, background suppression effect, and detection performance in complex scenes across various scenarios. This overcomes the problems of existing detection methods, such as ineffective detection, high false alarm rate, and difficulty in suppressing background clutter, when detecting small, dark targets in complex scenes. Moreover, the computational process of this invention is simpler. Attached Figure Description

[0017] Figure 1 The flowchart shows the infrared small target detection method based on density peak search and local features according to the present invention.

[0018] Figure 2 Images obtained from each step of the infrared small target detection method based on density peak search and local features according to the present invention;

[0019] Figure 3 This is a diagram illustrating the surrounding region scheme centered on the candidate target point in the infrared small target detection method based on density peak search and local features according to the present invention.

[0020] Figure 4 The results of detecting multiple input infrared images are presented in the infrared small target detection method based on density peak search and local features according to the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Reference Figure 1-4 An infrared small target detection method based on density peak search and local features includes: such as Figure 1 As shown, the infrared small target detection method based on density peak search and local features of the present invention mainly includes the following steps: inputting an infrared image, such as... Figure 2 a → Gaussian difference filtering, such as Figure 2 Middle b → density peak search, such as Figure 2 In the middle, c and d → local contrast is performed on candidate target points to enhance gradient characteristics and suppress background clutter, such as Figure 2 The three-dimensional plot of the gradient response characteristics of the candidate target in eight directions is shown in Figure e→. Figure 2 f → Adaptive thresholding of candidate target gradient response values ​​→ Detection result is an infrared image of weak targets, such as Figure 2 The process described above is explained in detail below.

[0023] The infrared small target detection method based on density peak search and local features of the present invention is implemented by the following steps:

[0024] Step S1: Use Gaussian difference filtering to preprocess the image and suppress high-brightness clutter;

[0025] A low-pass filtered image is obtained by convolving the image with a Gaussian function. Then, by optimizing and selecting different parameters, the two Gaussian filtered image results under different parameters are subtracted to obtain a Gaussian difference filtered image. This process suppresses high-brightness clutter near the target, resulting in the Gaussian difference filtered image. Figure 2 b.

[0026] Step S2: Use the density peak global search method to determine the location of candidate targets in the preprocessed image;

[0027] By defining the density and distance of each pixel in the infrared image, the density peak of each pixel in the image is obtained, such as... Figure 2 In step c, the density peaks of each pixel are arranged from largest to smallest. Through iterative optimization and threshold comparison, the top n pixels with the largest optimized density peaks are selected as candidate target points. Figure 2 The candidate target points are those within the circle in the middle d.

[0028] Step S3: Perform local contrast enhancement on candidate target points to improve gradient characteristics and suppress background clutter;

[0029] Using a 3×3 image patch slider, a local image patch centered on the candidate target point is created. By comparing the grayscale difference between the central and surrounding areas of the slider, the gradient characteristics of the real target point are enhanced and optimized while suppressing background clutter. The results of the enhancement are as follows: Figure 2 e.

[0030] Step S4: Calculate the multi-directional gradient features of each point using the Facet model;

[0031] By using the Facet model to calculate the multi-directional gradient at each point, the facet model considers that any small neighborhood of a pixel in the image can be fitted by a polynomial function. Compared with direct calculation methods, it can fully consider the gray-level influence of other points and more fully fit the gray-level changes of the image. By optimizing the polynomial function, key parameters in the polynomial can be directly obtained by performing convolution operations with a fixed filter. Based on the neighborhood polynomial, the gradient characteristics of each pixel in different directions in the neighborhood can be quickly calculated. For the gradient characteristics in eight directions of the image, such as... Figure 1 Three-dimensional plot of gradient characteristics in eight directions;

[0032] Step S5: Construct a surrounding region partitioning scheme to capture gradient characteristics in eight directions for targets of different sizes;

[0033] This includes setting different region size thresholds centered on candidate target points, optimizing the selection of surrounding region partitioning schemes, establishing two different region size thresholds, and three effective region extraction and partitioning schemes to adapt to the gradient characteristics of targets of different sizes, such as... Figure 3Region d in the figure shows a good gradient response when the target pixel size is smaller than 3×3, while region D in the figure shows a good gradient response when the target pixel size is larger than 3×3. Different extraction and partitioning schemes for different directions are shown below. Figure 3 When α = 0°, 45°, and 90°, the three extraction planning schemes effectively extract the gradient characteristic values ​​of the target in eight directions;

[0034] Step S6: Use a symmetric region contrast mechanism to suppress background clutter gradient characteristics and use the standard deviation of the symmetric region difference to weight the candidate target gradient characteristics;

[0035] The gradient characteristics of suppressing background clutter are achieved by subtracting regions along the symmetrical direction of candidate target points. Figure 3 As shown, for example, when α = 0°, for small targets of different sizes, the difference between D1-D2 or d1-d2 is taken as the gradient response value in that direction, and the gradient response value is weighted by the standard deviation of the symmetric region difference, as shown in the figure. Figure 3 The standard deviations of D1-D2 and d1-d2 are calculated to weight the gradient response value, which greatly distinguishes the gradient response values ​​of the target and non-target.

[0036] Step S7: Adaptive threshold segmentation is performed on the gradient characteristic values ​​of candidate targets to obtain infrared weak targets;

[0037] S7 employs an adaptive threshold segmentation method to process the gradient response values ​​of candidate target points according to a certain threshold and cyclic segmentation calculation. This filters out non-realistic target points below a certain threshold, obtaining the precise location of infrared weak targets and achieving infrared weak target detection. The resulting infrared weak target image is shown below. Figure 2 As shown in g. The adaptive threshold can be calculated based on the mean and standard deviation of the fused gradient response values.

[0038] Example 1

[0039] The infrared small target detection method based on density peak search and local features of this invention was used to detect four input infrared images, and the detection results are as follows: Figure 4 As shown. This infrared image is obtained by analyzing a background with sharp, complex cloud edge structures, as illustrated below. Figure 4 As shown in A1, its three-dimensional diagram is as follows: Figure 4 As shown in A2, a density peak search is performed to obtain candidate target points, as follows. Figure 4 In section A3, the gradient characteristics of candidate target points are enhanced through local contrast, and an effective region partitioning scheme is used to extract the gradient characteristic response values ​​of the target points in eight directions. A symmetrical region contrast mechanism is used to suppress background clutter gradient characteristics, and the standard deviation of the symmetrical region difference is used to weight the gradient characteristic response values ​​of the candidate targets. The final gradient response values ​​are obtained, and their 3D plot is shown below. Figure 4As shown in A4, adaptive threshold segmentation is performed on the gradient characteristic response value of the target point to obtain infrared weak targets, such as... Figure 4 As shown in A5.

[0040] Example 2

[0041] Similarly, by analyzing infrared images with complex ground building backgrounds, such infrared images... Figure 4 As shown in B1, its three-dimensional plot is as follows: Figure 4 As shown in B2, a density peak search is performed to obtain candidate target points, as follows. Figure 4 In section B3, the gradient characteristics of candidate target points are enhanced through local contrast, and an effective region partitioning scheme is used to extract the gradient characteristic response values ​​of the target points in eight directions. A symmetrical region contrast mechanism is used to suppress background clutter gradient characteristics, and the standard deviation of the symmetrical region difference is used to weight the gradient characteristic response values ​​of the candidate targets. The final gradient response values ​​are obtained, and their 3D plot is shown below. Figure 4 As shown in B4, adaptive thresholding is performed on the gradient characteristic response value of the target point to obtain infrared weak targets, such as... Figure 4 As shown in B5.

[0042] Example 3

[0043] Similarly, by examining an infrared image with a large, bright, edged sea surface background, this infrared image, as shown... Figure 4 As shown in C1, its three-dimensional plot is as follows: Figure 4 As shown in C2, a density peak search is performed to obtain candidate target points, as follows. Figure 4 In C3, the gradient characteristics of candidate target points are enhanced through local contrast, and an effective region division scheme is used to extract the gradient characteristic response values ​​of the target points in eight directions. A symmetrical region contrast mechanism is used to suppress background clutter gradient characteristics, and the standard deviation of the symmetrical region difference is used to weight the gradient characteristic response values ​​of the candidate targets. Finally, the gradient response values ​​are obtained, and their 3D plot is shown below. Figure 4 As shown in C4, adaptive thresholding is performed on the gradient characteristic response value of the target point to obtain infrared weak targets, such as... Figure 4 As shown in C5.

[0044] Example 4

[0045] Similarly, by analyzing infrared images with strong ground interference and complex cloud structures as background, this infrared image, like... Figure 4 As shown in D1, its three-dimensional plot is as follows: Figure 4 As shown in D2, a density peak search is performed to obtain candidate target points, as follows. Figure 4In D3, local contrast enhancement of candidate target points is performed to enhance gradient characteristics, and an effective region partitioning scheme is used to extract gradient response values ​​in eight directions. Background clutter gradient characteristics are suppressed using a symmetrical region contrast mechanism, and the standard deviation of the symmetrical region difference is used to weight the candidate target gradient response values. The final gradient response values ​​are obtained, and their 3D plot is shown below. Figure 4 As shown in D4, adaptive thresholding is performed on the gradient characteristic response value of the target point to obtain infrared weak targets, such as... Figure 4 As shown in D5.

[0046] The number of devices and processing scale described herein are for simplification of the invention. Applications, modifications, and variations of this invention will be readily apparent to those skilled in the art. Although embodiments of the invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. It can be applied to various fields suitable for this invention, and further modifications can be readily implemented by those skilled in the art. Therefore, without departing from the general concept defined by the claims and their equivalents, this invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A method for detecting small infrared targets based on density peak search and local features, characterized in that, Includes the following steps: S1. Preprocess the image using Gaussian difference filtering to suppress high-brightness noise; S2. The location of candidate targets is determined in the preprocessed image using the density peak global search method; S3. Perform local contrast analysis on candidate targets to enhance gradient characteristics and suppress background clutter; S4. Calculate the multi-directional gradient characteristics of each candidate target using the Facet model. In step S4, the multi-directional gradient of each point is calculated using the Facet model. The facet model considers that any small neighborhood of a pixel in the image can be fitted by a polynomial function. By optimizing the polynomial function, the key parameters in the polynomial can be directly obtained by performing convolution operations using a fixed filter. Based on the neighborhood polynomial, the gradient characteristics of each pixel in different directions in the neighborhood can be quickly obtained. S5. Using the candidate target point as the center, set different thresholds for the size range of different regions, optimize the division scheme of the surrounding regions, and capture the gradient characteristics of targets of different sizes in eight directions. S6. The gradient characteristics of background clutter are suppressed by subtracting regions along the symmetrical direction of the candidate target point and the gradient response value is weighted by the standard deviation of the symmetrical region difference. S7. Adaptive threshold segmentation method is used to obtain infrared weak targets by analyzing the gradient characteristic response values ​​of candidate targets. In step S7, the adaptive threshold can be calculated based on the mean and standard deviation of the fused gradient response values. By processing the gradient response values ​​of candidate target points according to a certain threshold and cyclic segmentation calculation, non-real target points under a certain threshold are filtered out to obtain the precise location of infrared weak targets.

2. The infrared small target detection method based on density peak search and local features as described in claim 1, characterized in that, In step S1, a low-pass filtered image is obtained by convolving the image with a Gaussian function. By optimizing and selecting different parameters, the two Gaussian filtered image results under different parameters are subtracted to obtain the Gaussian difference filtered result, which suppresses high-brightness clutter near the target and obtains the Gaussian difference filtered image.

3. The infrared small target detection method based on density peak search and local features as described in claim 2, characterized in that, In step S2, by defining the density and distance of each pixel in the infrared image, the density peak of each pixel in the image is obtained. The density peaks of each pixel are arranged from largest to smallest. Through iterative optimization of threshold comparison, the pixels with the largest optimized density peaks are selected as candidate target points.

4. The infrared small target detection method based on density peak search and local features as described in claim 3, characterized in that, In step S3, a 3×3 image patch slider is used to slide a local image patch centered on the candidate target point. By comparing the gray-level difference between the central region and the surrounding region of the slider, the gradient characteristics of the selected real target point are enhanced and optimized, and background clutter is suppressed.

Images