A weak and small target detection method in a complex scene

Abstract

The application discloses a weak and small target detection method in a complex scene and belongs to the technical field of signal processing. Firstly, noise filtering is performed on an image to reduce the influence of imaging system noise on target detection. Then, the distribution of image gradient vector domains is calculated, the image is lifted from a two-dimensional plane domain to a high-dimensional vector domain, the problem of weak and small target image feature deficiency is solved, signal structure expression capability is enhanced, and target and clutter feature difference is improved. Secondly, a multi-scale background suppression mapping diagram is calculated to realize background clutter suppression and highlight target signals. Meanwhile, multi-scale features can adapt to the change of target scale size. Thirdly, pre-detection is performed on a target region to extract each candidate target region in the image. Finally, a real target is extracted through a time domain filtering method. The application improves signal structure representation capability, enhances target and background clutter feature difference, and realizes high-reliability target detection.

Description

Technical Field

[0001] This invention belongs to the field of signal processing technology, and relates to weak target detection technology, specifically a method for weak target detection in complex scenarios. Background Technology

[0002] Due to the influence of target imaging distance, observation equipment, and operating environment, weak targets often exhibit characteristics such as low contrast, low signal-to-clutter ratio, and lack of feature information. These sub-ideal conditions easily lead to missed detections and false alarms for weak targets. The difficulties in weak target detection mainly lie in the following aspects:

[0003] (1) Because the imaging system is far from the target, the target occupies only a few or a dozen pixels on the camera target surface. The target texture and structural information are severely lost, which poses a great challenge to target detection and identification.

[0004] (2) Sky clouds and sea clutter have time-varying and non-stationary characteristics, which makes it difficult for weak target detection algorithms based on background modeling. At the same time, the intensity of some background clutter even exceeds the intensity of the target, and it is difficult to distinguish the target from clutter simply based on the signal strength.

[0005] (3) The target imaging intensity is weak, the texture and structural features are not obvious, there is little information available, the target's representation ability is weak, and it is difficult to reliably distinguish the target from background clutter in the two-dimensional image plane.

[0006] The paper “Infrared weak target detection by fusing visual saliency and local entropy, Chinese Optics, 2022, 15(2), 267-275” discloses a weak target detection method. First, the weak target is coarsely located using local entropy. Then, the saliency map of the coarsely located region is calculated by the visual saliency method. Finally, the weak target is detected by threshold segmentation. This method can detect weak targets under low signal-to-noise ratio. However, this method still has the problem of insufficient target representation effectiveness under strong fluctuation clutter interference and is prone to detection failure under complex background conditions.

[0007] The paper "Research on the Joint Infrared Weak Target Detection Method of Spatiotemporal Spectral Multi-Feature", Doctoral Dissertation, 2020, University of Electronic Science and Technology of China, discloses a method for constructing features in the spatial domain, temporal domain and transform domain, and achieving high-reliability detection of weak targets by fusing features from multiple domains. This method effectively distinguishes between the target and the background in the feature space. However, this method does not consider the change in target scale. When the distance to the detected target changes, the adaptability of the algorithm will decrease. Summary of the Invention

[0008] Traditional weak target detection methods suffer from reduced detection rate and increased false alarm rate in scenarios with strong clutter interference and varying target scales. Furthermore, in scenarios with strong clutter interference, the lack of feature differences between the target and clutter in the image, and the target being submerged in background clutter, are also problems encountered. This invention proposes a weak target detection method for complex scenarios. This method addresses the lack of target features in weak target images, enhances the information difference between the target and background clutter, and achieves robust detection of weak targets under extremely low signal-to-clutter ratios. This invention improves the structural representation capability of the signal, enhances the feature difference between the target and background clutter, and achieves highly reliable target detection.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] A method for detecting weak targets in complex scenarios includes the following steps:

[0011] Step 1) Filter the image using spatial filtering methods to reduce noise interference from the imaging system;

[0012] Step 2) Calculate the gradient vector domain distribution of the image, thus elevating the image from a two-dimensional planar domain to a high-dimensional vector domain;

[0013] Step 3) Calculate the multi-scale background suppression map, which is realized by convolution of the image gradient vector domain and the multi-scale closed contour template. The multi-scale closed contour template is composed of a horizontal template and a vertical template.

[0014] Step 4) Pre-detect the target region using a threshold segmentation method to extract each candidate target region in the image;

[0015] Step 5) Extract candidate targets using the target temporal filtering method. Construct the target motion trajectory and feature fusion pipeline filter based on the position of the candidate targets in the image in the previous few frames. Output the candidate targets that successfully pass through the feature fusion pipeline filter as the real targets.

[0016] The advantages of this invention compared to the prior art are as follows:

[0017] This invention elevates the original image domain to a high-dimensional gradient vector domain through data domain transformation operations, enhancing the feature differences between the target and background clutter and solving the problem of insufficient information on weak targets in the original image. Simultaneously, it employs multi-scale closed contour template convolution to achieve background suppression and target region contrast enhancement. This addresses the issue of unsatisfactory performance of fixed-parameter template-based detection algorithms for targets of different shapes and scales, while also improving the difference between the target and background clutter. This invention can effectively suppress strong clutter interference and improve the detection rate of weak targets. Attached Figure Description

[0018] Figure 1 This is an overall flowchart of the method for detecting weak targets in complex scenarios according to the present invention.

[0019] Figure 2 This is the Gaussian filter template for this invention.

[0020] Figure 3 This is the horizontal template in the closed contour template of the present invention.

[0021] Figure 4 This is the vertical template in the closed contour template of the present invention.

[0022] Figure 5 This is the original image of a small target in an embodiment of the present invention.

[0023] Figure 6 This is a multi-scale background suppression mapping diagram according to an embodiment of the present invention. Detailed Implementation

[0024] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are implemented based on the technical solution of the present invention, providing detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0025] like Figure 1 As shown, this embodiment provides a method for detecting weak targets in complex scenes, including the following steps:

[0026] Step 1: Image noise removal;

[0027] Step 2: Calculate the gradient vector domain distribution of the image;

[0028] Step 3: Calculate the multi-scale background suppression map;

[0029] Step 4: Target area pre-detection;

[0030] Step 5: Target time-domain filtering.

[0031] Specifically, step 1 includes: performing noise filtering on the image. Images of weak targets are often contaminated by random noise. This invention uses Gaussian filtering in spatial domain filtering algorithms to remove noise from the image and smooth it. The Gaussian filtering uses discrete 5*5 Gaussian filter templates (e.g., ...). Figure 2 As shown, a convolution operation is performed on the image. Noise interference in the image can be suppressed through Gaussian convolution.

[0032] Step 2 includes: Calculating the image gradient vector domain distribution requires traversing every pixel in the image. The image gradient vector domain includes horizontal and vertical gradient distributions. The specific calculation steps are as follows:

[0033] Step (2.1) Calculate the sum of gray levels in the first direction: C1(x,y)=I(xk,yk)+I(xk,y)+I(xk,y+k)

[0034] Step (2.2) Calculate the gray sum in the second direction: C2(x,y)=I(x+k,yk)+I(x+k,y)+I(x+k,y+k)

[0035] Step (2.3) Calculate the gray sum of the third direction: C3(x,y)=I(xk,yk)+I(x,yk)+I(x+k,yk)

[0036] Step (2.4) Calculate the gray sum in the fourth direction: C4(x,y)=I(xk,y+k)+I(x,y+k)+I(x+k,y+k)

[0037] The horizontal gradient distribution described in step (2.5) is: G x (x,y)=C2(x,y)-C1(x,y)

[0038] The vertical gradient distribution described in step (2.6) is: G y (x,y)=C4(x,y)-C3(x,y)

[0039] The image gradient vector domain in step (2.7) is: M(x,y)=[G x (x,y)G y (x,y)]

[0040] Where I(x,y) is the gray value of the image after noise filtering; k is the gradient calculation step size, preferably between 2 and 3; x is the horizontal coordinate of the image pixel; and y is the vertical coordinate of the image pixel. The gradient vector domain distribution of the image can be calculated through the above seven steps.

[0041] Step 3 includes: calculating a multi-scale background suppression map, wherein the multi-scale background suppression map is achieved by convolution of the image gradient vector domain and a multi-scale closed contour template, wherein the closed contour template is composed of a horizontal template. and vertical template Composition, such as Figure 3 , Figure 4 As shown, where s is the template scale parameter, the calculation method of the multi-scale background suppression map includes:

[0042] Step (3.1) involves convolving the horizontal gradient distribution in the image gradient vector domain using a closed contour horizontal template:

[0043]

[0044] The It is the background suppression map in the horizontal direction when the template scale parameter is s; x0 and y0 represent the horizontal and vertical traversal parameters during the convolution operation, respectively.

[0045] Step (3.2) involves convolving the vertical gradient distribution in the image gradient vector domain using a closed contour vertical template:

[0046]

[0047] The It is the vertical background suppression map when the template scale parameter is s.

[0048] Step (3.3) involves summing the background suppression maps from steps (3.1) and (3.2):

[0049]

[0050] The This is the background suppression map with template scale parameter s.

[0051] The multi-scale background suppression mapping (ME) described in step (3.4) o (x,y) is represented by a minimum-scale background suppression map. The minimum value of the background suppression map among the multi-scale parameters is obtained by subtracting it. The specific calculation formula is as follows:

[0052]

[0053] Wherein, min represents the minimum value of the multi-scale background suppression map; s1 and s2 represent the minimum and maximum scales of the multi-scale closed contour template, respectively, where s1 is preferably 2 and s2 is preferably 4.

[0054] Through steps 1-3, a multi-scale background suppression map can be obtained. In this image, the target region is enhanced, and background clutter, including cloud edges, is well suppressed. Figure 5 , Figure 6 These are comparison images before and after background suppression. It can be seen that the background clutter suppression is very obvious, and the target contrast is enhanced, which facilitates the next step of target detection.

[0055] Step 4 includes: extracting candidate target regions through threshold segmentation to achieve target region pre-detection. The segmentation threshold Th is calculated using the following formula:

[0056]

[0057] in, Let f be the mean of the multi-scale background suppression map, σ be the variance of the multi-scale background suppression map, and f be the mean of the multi-scale background suppression map. k It is the threshold segmentation adjustment coefficient, and the specific calculation method is as follows:

[0058] Step (4.1) Based on the size of the multi-scale background suppression map, nine rectangular regions are collected at equal intervals, each rectangular region being 32*32 in size;

[0059] Step (4.2) Calculate the mean m for each rectangular region. i and variance σ i where i = 1 to 9;

[0060] Step (4.3) involves applying the mean m to the mean m. i and variance σ i Sort the data from largest to smallest and find the value m. o and σ o ;

[0061] Step (4.4) Mean of the multi-scale background suppression map Using the median m o Variance σ is expressed as median σ. o ;

[0062] In step (4.5), regions exceeding the segmentation threshold Th in the multi-scale background suppression mapping are candidate target regions.

[0063] Step 5 includes: constructing a pipeline filter for target motion trajectory and feature fusion, and outputting candidate targets that successfully pass through the pipeline filter as real targets. The specific calculation method is as follows:

[0064] Step (5.1) Parameter initialization. Assuming the pipeline can hold N image frames, each candidate target in the image at time t... Effective counter C i (i = 1, 2, ..., n) are set to 0, where i represents the index of the i-th candidate target, and the total number of candidate targets is n. The radius of each candidate target is set to r. i t (i = 1, 2, ..., n), the pipe filter radius size for each candidate target. Choose a size slightly larger than the candidate target, preferably

[0065] Step (5.2) takes the time of the first frame image as the current time t, determines all candidate targets in the first frame image, and records their position information in the first frame image. and eigenvector f i t(i = 1, 2, ..., n), where the feature vector is preferably the target gray-level direction histogram vector;

[0066] Step (5.3) involves all candidate targets. Search for the pipe radius in the image at the next time step t+1. The center is Within a circular pipe region, detect whether there are candidate targets. If so, extract all candidate targets within the circular pipe region. Then, the correlation and matching degree of the attribute features of these candidate targets and the candidate targets in the pipeline area at time t are further compared, where m represents the number of all candidate targets extracted in the circular pipeline area at time t+1.

[0067] Step (5.4) introduces an attribute feature pipeline filter for a second association matching because the uncertainty of building the pipeline filter solely based on motion information is high.

[0068] First, we consider the feature vector f of the candidate target inside the pipe at time t+1. i t+1 Normalize (i = 1, 2, ..., m) such that ||f i t+1 ||=1 (i=1,2,...,m), where || represents the computation vector f i t+1 The L2 norm; then the feature vector of the candidate target to be matched at the center of the pipeline region at time t. Normalization is also performed, making Finally, the minimum cosine distance between the candidate target feature vectors within the pipeline region at time t+1 and the feature vectors at time t is calculated. The calculation formula is:

[0069]

[0070] If the minimum cosine distance is less than the specified threshold M T If the association is successful, then the effective counter C of the corresponding candidate target will be... j Add 1, where the threshold M T It can be obtained from a separate training set or by manually setting the scene for verification; if the association is unsuccessful, skip the frame image and move to the next frame image to continue the search until all N frames in the pipeline region have been searched.

[0071] Step (5.5): After processing N frames of images, determine the effective counter C for each candidate target. iThe output value of (i = 1, 2, ..., n). If the value of the effective counter is greater than or equal to K, the candidate target corresponding to the effective counter is determined to be a real target and its position is marked; otherwise, it is regarded as a false target and eliminated, where K is preferably N-1.

Claims

1. A method for detecting weak targets in complex scenes, characterized in that, Includes the following steps: Step 1) Filter the image using spatial filtering methods to reduce noise interference from the imaging system; Step 2) Calculate the image gradient vector domain distribution, elevating the image from a two-dimensional planar domain to a high-dimensional vector domain; the image gradient vector domain includes horizontal and vertical gradient distributions, and the calculation steps are as follows: Step (2.1) First calculation direction grayscale and : Step (2.2) Second calculation direction grayscale and : Step (2.3) Third calculation direction gray level and : Step (2.4) Fourth calculation of grayscale in the direction : The horizontal gradient distribution described in step (2.5) for: The vertical gradient distribution described in step (2.6) for: The image gradient vector domain described in step (2.7) for: in, The grayscale value of the image after noise filtering; Calculate the step size for the gradient; The horizontal coordinates of the image pixels; The vertical coordinates of the image pixels; Step 3) Calculate the multi-scale background suppression map, which is achieved by convolution of the image gradient vector domain and a multi-scale closed contour template. The multi-scale closed contour template consists of a horizontal template and a vertical template. The calculation steps for the multi-scale background suppression map are as follows: Step (3.1) involves applying the horizontal gradient distribution in the image gradient vector domain using the horizontal template of the multi-scale closed contour template. Perform convolution; Step (3.2) involves applying the vertical direction template of the multi-scale closed contour template to the vertical gradient distribution in the image gradient vector domain. Perform convolution; Step (3.3) accumulates the convolution results from steps (3.1) and (3.2) to obtain a background suppression map at a certain scale; The multi-scale background suppression map in step (3.4) is obtained by subtracting the minimum value of the background suppression map in the multi-scale range from the minimum-scale background suppression map; Step 4) Pre-detect the target region using a threshold segmentation method to extract each candidate target region in the image; Step 5) Extract candidate targets using the target temporal filtering method. Construct the target motion trajectory and feature fusion pipeline filter based on the position of the candidate targets in the image in the previous few frames. Output the candidate targets that successfully pass through the feature fusion pipeline filter as the real targets.

2. The method for detecting weak targets in complex scenes according to claim 1, characterized in that, Step 4) includes: Step (4.1) Based on the size of the multi-scale background suppression map, nine rectangular regions are collected at equal intervals, each rectangular region being 32 pixels in size. 32; Step (4.2) Calculate the mean of each rectangular region. and variance ,in ; Step (4.3) involves calculating the mean separately. and variance Sort by size from largest to smallest and calculate the mean. median and variance median ; Step (4.4) Mean of the multi-scale background suppression map Using median ,variance Using median ; Step (4.5) The multi-scale background suppression map exceeds the segmentation threshold The region that is selected is the candidate target region.

3. The method for detecting weak targets in complex scenes according to claim 2, characterized in that, Step 5) includes: Step (5.1) Parameter initialization: Assume the number of image frames the pipeline can hold is... ,Will Each candidate target in the time-lapse image effective counter Set to 0, where Representing the The index of the candidate target, and the total number of candidate targets. The radius of each candidate target is set to... Pipeline filter radius size for each candidate target for i = 1, 2, ..., n; Step (5.2) takes the time of the first frame image as the current time. The algorithm identifies all candidate targets in the first frame image and records their position information within the first frame image. and eigenvectors , where the feature vector The target gray-level direction histogram vector; Step (5.3) involves all candidate targets. , respectively in The search pipeline radius in the image at time t is The center of the circle is Within a circular pipe region, detect whether there are candidate targets; if so, extract all candidate targets within the circular pipe region. Then, further compare candidate targets. and pipeline area The attribute feature association matching degree of the candidate target at time step m, where j = 1, 2, ..., m. express The number of all candidate targets extracted within the circular pipe region at any given time; Step (5.4) introduces an attribute feature pipeline filter for a second association matching: First of all Feature vector of candidate targets within the pipeline region at time t. Normalization is performed so that ,in Indicates the calculation of eigenvectors of Norm; then for The feature vector of the candidate target to be matched at the center of the pipeline region at time t. Normalization is also performed, making ; final calculation Feature vectors of candidate targets within the pipeline region at time step and eigenvectors at time step minimum cosine distance between The calculation formula is: If the minimum cosine distance Less than the specified threshold If the association is successful, then the effective counter of the corresponding candidate target will be updated. Increment by 1; if the association fails, skip the current frame and proceed to the next frame to continue the search until the pipeline region is reached. The search continues until all frames have been searched. Step (5.5) when After the frame image is processed, determine the effective counter for each candidate target. The output value; if the valid counter The value is greater than or equal to If a target is identified as a real target, its location is marked; otherwise, the corresponding candidate target is discarded as a false target. for .

Images