Infrared small target detection method and system based on attribute guide feature weighted fusion

Abstract

The invention discloses an infrared small target detection method and system based on attribute guide feature weighted fusion. The method comprises the steps that (1) a small target size attribute rule is designed, attribute filtering is introduced to obtain attribute difference, rapid candidate target extraction is achieved, and meanwhile the attribute difference can guide parameter setting of improved density peak value clustering and can also deduce the weight of feature fusion; and (2) an improved density peak clustering algorithm is designed, a local feature extraction model based on improved density peak clustering is constructed, and local features of multiple targets of the cluster can be accurately represented. And (3) constructing a gray contribution model to extract global features of the small target based on the gray statistical characteristics of the small target. Through weighted fusion of local and global features, cluster targets are effectively enhanced and background clutters are suppressed. The method overcomes the defects of an existing method, and has a better cluster multi-target enhancement effect and higher multi-target detection performance.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing and target detection, and in particular to an infrared small target detection method and system based on attribute-guided feature weighted fusion. Background Art

[0002] Infrared small target detection is crucial for infrared early warning and low-speed small target monitoring. With the routine deployment of technologies such as multiple warhead missiles and drone swarms, missed targets may pose a major security threat. Therefore, robust detection of clustered multiple small targets is of great significance. Due to factors such as long-range detection or target size, the area occupied by the target in the infrared image is very small, with a pixel area ranging from a few pixels to dozens of pixels. In addition, considering the dense state of multiple targets and complex background interference, achieving accurate detection of infrared clustered small targets remains a challenging problem.

[0003] Existing infrared small target detection methods can be divided into three categories: the first category is methods based on background estimation filtering, such as wavelet Hough transform and top-hat filtering, which are simple and fast but have poor detection performance in complex scenes. The second category is low-rank sparse methods, which transform the target detection task into a matrix or tensor recovery problem by decomposing low-rank background and sparse targets, such as infrared block image model (IPI) and partial sum model of tensor nuclear norm (PSTNN), but the algorithm is sensitive to some corners and prominent edges. The third category is local feature-based methods, which characterize the difference between the target and the local background by designing different local window structures and local contrast measurement operators, such as multi-scale block contrast measurement (MPCM) and three-layer local contrast measurement (TLLCM). In order to fully extract the local features of small targets, researchers have also explored different methods, such as pixel-level local contrast measurement (PLLCM) using random walks and local intensity gradient model (LIG) using local gradient features. When faced with clutter whose local intensity is similar to that of the target, the local feature method has limitations. For this reason, some researchers have also tried to make up for the shortcomings of local features by fusing global features. For example:

[0004] The patent document with application number (CN202410107284.8) proposes an infrared single small target detection method and system based on local differences and global features. The local difference features are calculated through a local window, candidate targets are obtained through the features, and a global unique feature extraction model of the candidate target points is constructed. The detection performance is improved by fusing local difference features with global features.

[0005] Although the above local feature-based methods improve the accuracy of local features in different ways, they usually assume that there is only a single target in the local window. When faced with clustered multiple targets, the accuracy of local feature extraction is reduced. The above method that adds global unique features is designed for single target detection and cannot effectively extract the features of clustered multiple targets, resulting in reduced cluster target detection performance. Summary of the invention

[0006] In view of this, the purpose of the present invention is to provide an infrared small target detection method and system based on attribute-guided feature weighted fusion. On the one hand, the present invention designs small target size attribute rules, introduces attribute filtering to obtain attribute differences, and realizes rapid candidate target extraction. At the same time, the attribute difference can not only guide the parameter setting of improved density peak clustering, but also derive the weight of feature fusion. On the other hand, the present invention designs an improved density peak clustering algorithm, constructs a local feature extraction model based on improved density peak clustering, and can accurately represent the local features of clustered multi-targets. On the other hand, based on the grayscale statistical characteristics of small targets, the present invention constructs a grayscale contribution model to extract the global features of small targets. Through the weighted fusion of local and global features, the cluster targets are effectively enhanced and background clutter is suppressed. The present invention overcomes the shortcomings of existing methods and has better cluster multi-target enhancement effect and higher multi-target detection performance.

[0007] The present invention solves the technical problem by adopting a specific technical solution: an infrared small target detection method based on attribute-guided feature weighted fusion, comprising the following steps:

[0008] Step 1, obtaining a single frame original infrared image to be detected;

[0009] Step 2, obtaining a bright target attribute difference map corresponding to the original infrared image;

[0010] Step 3, perform threshold segmentation on the attribute difference map to obtain a set of candidate targets;

[0011] Step 4, calculating the local grayscale difference feature of each pixel in the candidate target set to obtain a local grayscale difference feature map;

[0012] Step 5, calculating the grayscale contribution feature of each pixel in the candidate target set to obtain a grayscale contribution feature map;

[0013] Step 6, constructing a weighted fusion graph based on the attribute difference graph obtained in step 2, the local grayscale difference feature graph obtained in step 4, and the contribution feature graph obtained in step 5;

[0014] Step 7: Perform maximum inter-class variance threshold segmentation on the weighted fusion image to obtain infrared small target detection results.

[0015] Furthermore, the specific implementation of step 2 is as follows;

[0016] First set the attribute rules It means that the value of the δ attribute of each area of ​​the original infrared image is compared with the reference value k, the area less than or equal to the reference value k is retained, the filtering operation is performed, and then the attribute profile is calculated by the following formula:

[0017]

[0018] Where γ(O) is the attribute refinement operation, and O is the original infrared image; finally, the attribute difference map of the bright target is calculated:

[0019] S b =min|O-AP(O)|.

[0020] Furthermore, the specific implementation of step 3 is as follows;

[0021] According to the attribute difference map, calculate the segmentation threshold:

[0022] Th c =max(α×V 1 ,V k )

[0023] Where V 1 and V k Represents attribute difference graph S b The maximum and kth largest pixel values ​​are obtained, α and k are empirical constants; then the pixels in the attribute difference map that are greater than the threshold are retained to obtain the candidate target set D ct .

[0024] Furthermore, in step 4, each pixel D in the candidate target set is z ∈D ct The following processing is performed: In the original infrared image O, D z The coordinates (x z ,y z ) as the center, construct L P1 ×L P1 The local block P1 of pixel size is then classified and re-labeled using the improved density peak clustering method to obtain the classification and labeling results, and the local block P1 is classified as the central target area T c , non-central target area T k , noise point N j , the nearest neighbor background area aB i and non-neighboring background area naB i ; According to the classification results, calculate the candidate target pixel D z After calculating all candidate target pixels, the local grayscale difference feature map PLD is obtained; where Dct represents the candidate target set.

[0025] Furthermore, the specific implementation of step 4 includes the following sub-steps:

[0026] Step 4.1, first define the dataset P composed of all pixels in the local window as follows:

[0027] P = {pix i =(I i ,x i ,y i ),i=1,2,3,…,N}

[0028] Among them, pix i Refers to the i-th pixel in the local window, I i is the pixel grayscale, x i and i is pixel pix i The coordinates in the local block P1, N is the total number of all pixels in the local block; since the small target has a spatial position in the image, in the clustering of the local area of ​​the image, the coordinate factor is added to define any two data points pix i and pix j The distance between them is:

[0029]

[0030] Step 4.2, define the cutoff distance d c for:

[0031] d c (x i ,y i )=σ×S b (x i ,y i )

[0032] Where σ is the gain factor, S b is the attribute difference graph;

[0033] Step 4.3, calculate each pixel pix in the block one by one according to the truncation distance i The local density ρ i and the local distance δ i ;

[0034] Step 4.4: In order to determine the cluster center more quickly, the δ of all pixels is i The values ​​are sorted in descending order, and the first n pixels are marked as cluster centers, which are marked with numbers C1, C2, C3, ..., Cn respectively;

[0035] Step 4.5, classify the remaining points one by one into the class of the data point with the nearest neighbor and greater density than it, define array C as the clustering result, assuming that after clustering, C contains Cn categories;

[0036] Step 4.6, initial labeling of clustering results: Initially label each category pixel set according to the following scale conditions:

[0037]

[0038] Among them, h and w are the maximum coordinate distances of a certain type of pixel set in the vertical and horizontal directions, respectively, B i 、N j and T represent the background area, noise, and target area, respectively;

[0039] Step 4.7, clustering result re-labeling: define the target area in T located at the center of block P1 as the central target area T of the current block c , the rest are non-central target areas T k ; Definition B i Center and center target T c The background area with no other background pixels between the nearest pixel distance is the nearest neighbor background area aB i , otherwise it is the non-neighboring background area naB i ;

[0040] Step 4.8, define the current candidate target D z (x z ,y z )∈D ct The local grayscale difference feature PLD is as follows:

[0041]

[0042] Among them, (x, y) is the coordinate of the pixel in the local grayscale difference feature map PLD, m Tc is the target area T c The average pixel gray value, m aBi is the nearest neighbor background area aB i The average pixel gray value, nb refers to the nearest background area aB i After calculating the global unique feature values ​​of all candidate target points, the local grayscale difference feature map PLD can be obtained.

[0043] Furthermore, in step 4.3, ρ i The calculation formula is as follows:

[0044]

[0045] Local distance δ i The formula is as follows:

[0046]

[0047] Among them, ρ refers to the set of local densities of all pixels, j:ρ i >ρ j Refers to the local density being less than the pixel pix i The local density of pixels.

[0048] Furthermore, the specific implementation of step 5 includes the following sub-steps:

[0049] Step 5.1: For each pixel D in the candidate target set, z ∈D ct The following processing is performed: In the original infrared image O, D z The coordinates (x z ,y z ) as the center, construct a local block P2 of 9×9 pixels, and then extract the grayscale histogram H of the local pixel block P2 z :

[0050] H z (r L )=n L ,L=0,1,2,3,…,255

[0051] where r L Refers to the Lth grayscale, n L Refers to the gray level r in the local block P2 L The number of pixel values;

[0052] Step 5.2: All candidate target pixels D z ∈D ct The histograms of are accumulated to form the total candidate area histogram. The mathematical formula is as follows:

[0053]

[0054] Where n refers to the number of candidate target pixels;

[0055] Step 5.3, in the total candidate area histogram, use the following formula to calculate the proportion of the histogram of each candidate target:

[0056]

[0057] Step 5.4, construct an all-0 matrix GLC with the same size as the original infrared image to store the grayscale contribution eigenvalues, where the calculation formula of the element value GLC(x,y) of the candidate target coordinates is as follows:

[0058]

[0059] Among them, (x, y) is the coordinate of the pixel in the grayscale contribution feature map GLC, n refers to the number of candidate target pixels, x z and z Refers to the coordinates of the zth candidate target point; after calculating the grayscale contribution eigenvalues ​​of all candidate target points, the grayscale contribution feature map GLC can be obtained.

[0060] Furthermore, the specific implementation of step 6 includes the following sub-steps:

[0061] Step 6.1, calculate the fusion weight map w according to the following formula:

[0062] w=S b / O

[0063] Among them, O is the original infrared image, S b is the attribute difference graph;

[0064] Step 6.2, multiply the fusion weight map, local grayscale difference feature map and contribution feature map, and calculate the fusion map Fu according to the following formula:

[0065] Fu=(w×PLD×GLC) *

[0066] in(·) * Refers to the normalization operation, PLD is the local grayscale difference feature map, and GLC is the grayscale contribution feature map.

[0067] Furthermore, the specific implementation of step 7 is as follows:

[0068] Step 7.1, take out the values ​​greater than 0 in the fusion image Fu and store them into the array R, R = {0, Fu(x, y)}, Fu(x, y)>0, where x and y refer to the coordinates of the pixels in the fusion image; then calculate the maximum inter-class variance threshold T of the fusion image Fu;

[0069] The calculation formula of the maximum inter-class variance threshold T is as follows:

[0070]

[0071] Among them, R i The index value of the ith element in R, and N refers to the number of elements less than T and the number of elements greater than or equal to T in the array R, respectively. R is the total number of elements in array R, The average value of the elements in index group R that are less than T. Refers to the average value of elements greater than or equal to T;

[0072] Step 7.2, segment the fusion image Fu according to the threshold T. After the threshold segmentation, the pixel values ​​in the fusion matrix that are greater than or equal to the threshold are set to 1, and the pixel values ​​that are less than the threshold are set to 0, and a binary image is obtained. The pixels that are set to 1 will be used as the target output.

[0073] The present invention also provides an infrared single small target detection system based on attribute-guided feature weighted fusion, comprising the following modules:

[0074] The original infrared image acquisition module is used to acquire a single frame of original infrared image to be detected;

[0075] The attribute difference map acquisition module is used to obtain the bright target attribute difference map corresponding to the original infrared image;

[0076] The threshold calculation module of the attribute difference map is used to perform threshold segmentation on the attribute difference map to obtain a set of candidate targets;

[0077] A local grayscale difference feature map acquisition module is used to calculate the local grayscale difference feature of each pixel in the candidate target set to obtain a local grayscale difference feature map;

[0078] A grayscale contribution feature map acquisition module is used to calculate the grayscale contribution feature of each pixel in the candidate target set to obtain a grayscale contribution feature map;

[0079] A fusion image acquisition module is used to construct a weighted fusion image based on the attribute difference image, the local grayscale difference feature image and the contribution feature image;

[0080] The infrared small target acquisition module is used to perform maximum inter-class variance threshold segmentation on the weighted fusion image to obtain the infrared small target detection results.

[0081] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0082] (1) This paper considers the size characteristics of small targets in infrared images and proposes a candidate target extraction method based on attribute filtering guidance. Compared with the square window calculation of local contrast, it can extract cluster targets more efficiently and greatly reduce the running time of the detection algorithm. In addition, this attribute can not only guide the parameter setting of improved density peak clustering (IDPC), but also derive the weight of feature fusion.

[0083] (2) In order to more accurately extract the local grayscale features of clustered multi-targets, the DPC algorithm was improved and a local grayscale difference model based on IDPC was established to avoid the interference of dense target states.

[0084] (3) Based on the grayscale statistical characteristics of small targets, a global feature representation scheme based on grayscale contribution is designed to extract the grayscale statistical characteristics of multiple targets and suppress highlight clutter.

[0085] Through the weighted fusion of local grayscale difference features and grayscale contribution features, the cluster targets are effectively enhanced, the cluster multi-target detection performance is further improved, and the problem of missed detection of multiple small targets in clusters by existing detection methods is compensated. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 A flowchart provided for an embodiment of the present invention;

[0087] Figure 2 A schematic diagram of local block classification provided by an embodiment of the present invention;

[0088] Figure 3 A schematic diagram of grayscale contribution feature calculation provided by an embodiment of the present invention;

[0089] Figure 4 Intuitive detection results of the method provided by the embodiment of the present invention and the existing algorithm;

[0090] Figure 5 The ROC curve diagram of the method provided in the embodiment of the present invention and the existing algorithm. DETAILED DESCRIPTION

[0091] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0092] like Figure 1 As shown, the target detection method process is represented in the form of a flow chart. The embodiment of the present invention provides an infrared small target detection method based on attribute-guided feature weighted fusion, and the specific process steps are as follows:

[0093] Step 1, obtain a single frame original infrared image O to be detected, whose size is M×N, and the grayscale value of the pixel of the image O is represented by O(x, y), where x and y represent pixel coordinates, x=1, 2, ..., M, y=1, 2, ..., N;

[0094] Step 2: Obtain bright target attribute difference map S b Specifically, the design attribute rules Indicates that the value of the δ attribute of each area of ​​the infrared original image is compared with the reference value k, and the area less than or equal to the reference value k is retained for filtering. In this embodiment, the area attribute and the diagonal length attribute of the circumscribed rectangle diag are selected to set the attribute rules of the attribute filtering. in Refers to an area that is less than or equal to 81 pixels. The diagonal of the minimum bounding rectangle of the region is less than or equal to 14 pixels. The attribute profile is then calculated using the following formula:

[0095]

[0096] Where γ(O) is the attribute refinement operation. Since the general detection task is for bright targets, this embodiment only uses the attribute refinement operation. Then calculate the bright target attribute difference map:

[0097] S b =min|O-AP(O)|

[0098] Step 3: attribute difference graph S b Perform threshold segmentation to obtain a set of candidate targets. Specifically, calculate the segmentation threshold based on the attribute difference graph:

[0099] Th c =max(α×V 1 ,V k )

[0100] Where V 1 and V k Represents attribute difference graph S b The maximum and kth largest pixel values ​​are obtained, and α and k are empirical constants. In an optional embodiment, the optimal value of α is 0.1, and the optimal value of k is 40. Then, the pixels in the attribute difference map that are greater than the threshold are retained to obtain the candidate target set D ct

[0101] Step 4: One by one, for each pixel D in the candidate target set z ∈D ct The following processing is performed: In the original infrared image O, D z The coordinates (x z ,y z ) as the center, construct L P1 ×L P1 The local block P1 of pixel size is then classified and re-labeled using the improved density peak clustering (IDPC) method, and the following can be obtained: Figure 2 The classification and labeling results shown in the figure classify the local block P1 as the central target area T c , non-central target area T k , noise point N j , the nearest neighbor background area aB i and non-neighboring background area naB i ; According to the classification results, calculate the candidate target pixel D z After calculating all candidate target pixels, the local grayscale difference feature map PLD is obtained. The specific steps include:

[0102] Step 4.1, first define the dataset P composed of all pixels in the local window as follows:

[0103] P = {pix i =(I i ,x i ,y i ),i=1,2,3,…,N}

[0104] Among them, pix i Refers to the i-th pixel in the local window, I i is the pixel grayscale, x i and i is pixel pix i The coordinates in the local block P1, N is the total number of all pixels in the local block. Since the small target has a spatial position in the image, in the clustering of the local area of ​​the image, the coordinate factor is added to define any two data points pix i and pix j The distance between them is:

[0105]

[0106] Step 4.2, define the cutoff distance d c for:

[0107] d c (x i ,y i )=σ×S b (x i ,y i )

[0108] Wherein σ is a gain factor. In an optional embodiment, the optimal value of σ is 0.65.

[0109] Step 4.3, calculate each pixel pix in the block one by one according to the truncation distance i The local density ρ i and the local distance δ i ρ i The calculation formula is as follows:

[0110]

[0111] Local distance δ i The formula is as follows:

[0112]

[0113] Among them, ρ refers to the set of local densities of all pixels, j:ρ i >ρ jRefers to the local density being less than the pixel pix i The local density of pixels.

[0114] Step 4.4: In order to determine the cluster center more quickly, the δ of all pixels is i The values ​​are sorted in descending order, and the first n pixel points are all marked as cluster centers, which are marked with numbers C1, C2, C3, ..., Cn respectively. In an optional embodiment, the best value of n is 5.

[0115] Step 4.5, classify the remaining points one by one into the class of the data point with the nearest neighbor and greater density than it, define array C as the clustering result, and assume that after clustering, C contains Cn categories.

[0116] Step 4.6, initial labeling of clustering results. Specifically, since noise points usually exist in isolation, there are no other isolated points in the neighborhood of the isolated point. When a small target has only a few pixels, the grayscale difference of its internal pixels may be large, resulting in over-segmentation of the target and multiple adjacent isolated points. When the center of a small target becomes an isolated point, it is necessary to re-label the isolated pixels in the neighborhood of the center of the small target to the same category as the center pixel. At the same time, since the size of small targets is between 2×1 and 9×9 pixels, in order to accurately divide the background area, noise points and target area, each category pixel set is initially labeled according to the following scale conditions:

[0117]

[0118] Among them, h and w are the maximum coordinate distances of a certain type of pixel set in the vertical and horizontal directions, respectively, B i 、N j and T represent the background area, noise and target area respectively. There may be multiple background areas, noise areas and target areas at the same time.

[0119] Step 4.7: Re-label the clustering results. In order to determine the target area of ​​the current block, define the target area located at the center of block P1 in T as the central target area T of the current block. c , the rest are non-target center areas T k ; In order to eliminate the influence of the highlight clutter edge, define B i Center and center target T c The background area with no other background pixels between the nearest pixel distance is the nearest neighbor background area aB i , otherwise it is the non-neighboring background area naB i The re-labeling result is shown in the figure below. Figure 2 shown.

[0120] Step 4.8, define the current candidate target D z (x z ,y z )∈Dct The local grayscale difference feature PLD is as follows:

[0121]

[0122] Where (x, y) is the coordinate of the pixel in the local grayscale difference feature map PLD, m Tc is the target area T c The average pixel gray value, m aBi is the nearest neighbor background area aB i The average pixel gray value, nb refers to the nearest background area aB i After calculating the global unique feature values ​​of all candidate target points, the local grayscale difference feature map PLD can be obtained.

[0123] Step 5: Calculate the grayscale contribution feature of each candidate target point. The calculation process is as follows: Figure 3 As shown, the grayscale contribution feature map GLC is obtained. Specifically, it includes the following sub-steps:

[0124] Step 5.1: For each pixel D in the candidate target set, z ∈D ct The following processing is performed: In the original infrared image O, D z The coordinates (x z ,y z ) as the center, construct a local block P2 of 9×9 pixels, and then extract the grayscale histogram H of the local pixel block P2 z :

[0125] H z (r L )=n L ,L=0,1,2,3,…,255

[0126] where r L Refers to the Lth grayscale, n L Refers to the gray level r in the local block P2 L The number of pixel values.

[0127] Step 5.2: All candidate target pixels D z ∈D ct The histograms of are accumulated to form the total candidate area histogram. The mathematical formula is as follows:

[0128]

[0129] Where n refers to the number of candidate target pixels.

[0130] Step 5.3, in the total candidate region histogram, most of the pixels corresponding to the grayscale of small targets come from small targets. In order to measure the grayscale contribution of the candidate region, the proportion of the histogram of each candidate target is calculated using the following formula:

[0131]

[0132] Step 5.4, construct a full zero matrix GLC with the same size as the original image to store the grayscale contribution eigenvalues, where the calculation formula of the element value GLC(x,y) of the candidate target coordinates is as follows:

[0133]

[0134] Among them, (x, y) is the coordinate of the pixel in the grayscale contribution feature map GLC, n refers to the number of candidate target pixels, x z and z Refers to the coordinates of the zth candidate target point. After calculating the grayscale contribution eigenvalues ​​of all candidate target points, the grayscale contribution feature map GLC can be obtained.

[0135] Step 6, constructing a weighted fusion graph Fu according to the attribute difference graph obtained in step 2, the local grayscale difference feature graph obtained in step 4, and the contribution feature graph obtained in step 5; specifically, the following sub-steps are included:

[0136] Step 6.1, calculate the fusion weight map w according to the following formula:

[0137] w=S b / O

[0138] Step 6.2, multiply the fusion weight map, local grayscale difference feature map and contribution feature map, and calculate the fusion map Fu according to the following formula:

[0139] Fu=(w×PLD×GLC) *

[0140] in(·) * Refers to the normalization operation.

[0141] Step 7, performing maximum inter-class variance threshold segmentation on the weighted fusion image to obtain infrared small target detection results, specifically including the following sub-steps:

[0142] Step 7.1, take out the values ​​greater than 0 in the fused image Fu and store them in the array R, R = {0, Fu(x, y)}, Fu(x, y)>0, where x and y refer to the coordinates of the pixels in the fused image. Then calculate the maximum inter-class variance threshold T of the Fu fused image, and the calculation formula is as follows:

[0143]

[0144] Among them, R i The index value of the ith element in R, and N refers to the number of elements less than T and the number of elements greater than or equal to T in the array R, respectively. R is the total number of elements in array R, The average value of the elements in index group R that are less than T. Refers to the average value of elements greater than or equal to T.

[0145] Step 7.2, segment the fused image Fu according to the threshold T. After the threshold segmentation, the pixel values ​​in the fusion matrix that are greater than or equal to the threshold are set to 1, and the pixel values ​​that are less than the threshold are set to 0, and a binary image is obtained. The pixels that are set to 1 will be used as the target output.

[0146] In order to prove the effectiveness and detection ability of this method, infrared multi-small target images in four complex scenes were selected for verification and compared with other commonly used target detection methods. The other compared algorithms are: non-deep learning algorithms (LCM, AMWLCM, ASPCM, FKRW, PLLCM, PSTNN, LogTFNN) and deep learning algorithms (ACM-U-Net and RDIAN). Figure 4 The intuitive detection results of each algorithm are shown, where the red rectangle refers to the detected target and the blue ellipse refers to the residual clutter. Figure 4 The comparison algorithm either cannot effectively detect all targets (target overlap or target missed detection), or detects targets but has a large number of false alarms, while the method of the present invention detects multiple targets without residual background clutter. It is proved that the method of the present invention has a better enhancement effect on clustered small targets and a stronger background clutter suppression ability. Figure 5 The ROC curves of various algorithms are shown. The closer the curve is to the upper left corner, the better the detection effect. AUC refers to the area under the curve. The larger the value, the better the detection performance. Figure 5 It can be seen that the curve of the method of the present invention is at the upper left corner of the other methods, and the AUC value of the method of the present invention is the largest, which proves that the method of the present invention has higher detection performance for cluster multi-targets.

[0147] The present invention also discloses an infrared single small target detection system based on attribute-guided feature weighted fusion, comprising the following modules:

[0148] The original infrared image acquisition module is used to acquire a single frame of original infrared image to be detected;

[0149] The attribute difference map acquisition module is used to obtain the bright target attribute difference map corresponding to the original infrared image;

[0150] The threshold calculation module of the attribute difference map is used to perform threshold segmentation on the attribute difference map to obtain a set of candidate targets;

[0151] A local grayscale difference feature map acquisition module is used to calculate the local grayscale difference feature of each pixel in the candidate target set to obtain a local grayscale difference feature map;

[0152] A grayscale contribution feature map acquisition module is used to calculate the grayscale contribution feature of each pixel in the candidate target set to obtain a grayscale contribution feature map;

[0153] A fusion image acquisition module is used to construct a weighted fusion image based on the attribute difference image, the local grayscale difference feature image and the contribution feature image;

[0154] The infrared small target acquisition module is used to perform maximum inter-class variance threshold segmentation on the weighted fusion image to obtain the infrared small target detection results.

[0155] The specific implementation method of each module is the same as each step and will not be described in detail in the present invention.

[0156] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.

[0157] This patent is not limited to the above-mentioned optimal implementation mode. Anyone can derive other various forms of infrared small target detection methods based on attribute-guided feature weighted fusion under the inspiration of this patent. All equal changes and modifications made according to the scope of the patent application of the present invention should be covered by this patent.

Claims

1. A method for infrared small target detection based on attribute-guided feature weighted fusion, characterized in that: The steps include: Step 1, obtaining a single frame original infrared image to be detected; Step 2, obtaining a bright target attribute difference map corresponding to the original infrared image; Step 3, perform threshold segmentation on the attribute difference map to obtain a set of candidate targets; Step 4, calculating the local grayscale difference feature of each pixel in the candidate target set to obtain a local grayscale difference feature map; Step 5, calculating the grayscale contribution feature of each pixel in the candidate target set to obtain a grayscale contribution feature map; Step 6, constructing a weighted fusion graph based on the attribute difference graph obtained in step 2, the local grayscale difference feature graph obtained in step 4, and the contribution feature graph obtained in step 5; Step 7: Perform maximum inter-class variance threshold segmentation on the weighted fusion image to obtain infrared small target detection results.

2. The infrared small target detection method based on attribute-guided feature weighted fusion as claimed in claim 1, characterized in that: The specific implementation of step 2 is as follows; First set the attribute rules It means that the value of the δ attribute of each area of ​​the original infrared image is compared with the reference value k, the area less than or equal to the reference value k is retained, the filtering operation is performed, and then the attribute profile is calculated by the following formula: Where γ(O) is the attribute refinement operation, and O is the original infrared image; finally, the attribute difference map of the bright target is calculated: S b =min|O-AP(O)|.

3. The infrared small target detection method based on attribute-guided feature weighted fusion as claimed in claim 1, characterized in that: The specific implementation of step 3 is as follows; According to the attribute difference map, calculate the segmentation threshold: Th c =max(α×V1,V k ) Where V1 and V k Represents attribute difference graph S b The maximum and kth largest pixel values ​​are obtained, α and k are empirical constants; then the pixels in the attribute difference map that are greater than the threshold are retained to obtain the candidate target set D ct .

4. The infrared small target detection method based on attribute-guided feature weighted fusion as claimed in claim 1, characterized in that: In step 4, each pixel D in the candidate target set is z ∈D ct The following processing is performed: In the original infrared image O, D z The coordinates (x z ,y z ) as the center, construct L P1 ×L P1 The local block P1 of pixel size is then classified and re-labeled using the improved density peak clustering method to obtain the classification and labeling results, and the local block P1 is classified as the central target area T c , non-central target area T k , noise point N j , the nearest neighbor background area aB i and non-neighboring background area naB i ; According to the classification results, calculate the candidate target pixel D z After calculating all candidate target pixels, the local grayscale difference feature map PLD is obtained; where D ct represents the candidate target set.

5. The infrared small target detection method based on attribute-guided feature weighted fusion as claimed in claim 4, characterized in that: The specific implementation of step 4 includes the following sub-steps: Step 4.1, first define the dataset P composed of all pixels in the local window as follows: P={pix i =(I i ,x i ,y i ),i=1,2,3,…,N} Among them, pix i Refers to the i-th pixel in the local window, I i is the pixel grayscale, x i and i is pixel pix i The coordinates in the local block P1, N is the total number of all pixels in the local block; since the small target has a spatial position in the image, in the clustering of the local area of ​​the image, the coordinate factor is added to define any two data points pix i and pix j The distance between them is: Step 4.2, define the cutoff distance d c for: d c (x i ,y i )=σ×S b (x i ,y i ) Where σ is the gain factor, S b is the attribute difference graph; Step 4.3, calculate each pixel pix in the block one by one according to the truncation distance i The local density ρ i and the local distance δ i ; Step 4.4: In order to determine the cluster center more quickly, the δ i The values ​​are sorted in descending order, and the first n pixels are marked as cluster centers, which are marked with numbers C1, C2, C3, ..., Cn respectively; Step 4.5, classify the remaining points one by one into the class of the data point with the nearest neighbor and greater density than it, define array C as the clustering result, assuming that after clustering, C contains Cn categories; Step 4.6, initial labeling of clustering results: Initially label each category pixel set according to the following scale conditions: Among them, h and w are the maximum coordinate distances of a certain type of pixel set in the vertical and horizontal directions, respectively, B i 、N j and T represent the background area, noise, and target area, respectively; Step 4.7, clustering result re-labeling: define the target area in T located at the center of block P1 as the central target area T of the current block c , the rest are non-target center areas T k ; Definition B i Middle and center target area T c The background area with no other background pixels between the nearest pixel distance is the nearest neighbor background area aB i , otherwise it is the non-neighboring background area naB i ; Step 4.8, define the current candidate target D z (x z ,y z )∈D ct The local grayscale difference feature PLD is as follows: Among them, (x, y) is the coordinate of the pixel in the local grayscale difference feature map PLD, m Tc is the target area T c The average pixel gray value, m aB i is the nearest neighbor background area aB i The average pixel gray value, nb refers to the nearest background area aB i After calculating the global unique feature values ​​of all candidate target points, the local grayscale difference feature map PLD can be obtained.

6. The infrared small target detection method based on attribute-guided feature weighted fusion as claimed in claim 5, characterized in that: In step 4.3, ρ i The calculation formula is as follows: Local distance δ i The formula is as follows: Among them, ρ refers to the set of local densities of all pixels, j:ρ i >ρ j Refers to the local density being less than the pixel pix i The local density of pixels.

7. The infrared small target detection method based on attribute-guided feature weighted fusion as claimed in claim 1, characterized in that: The specific implementation of step 5 includes the following sub-steps: Step 5.1: For each pixel D in the candidate target set, z ∈D ct The following processing is performed: In the original infrared image O, D z The coordinates (x z ,y z ) as the center, construct a local block P2 of 9×9 pixels, and then extract the grayscale histogram H of the local pixel block P2 z : H z (r L )=n L ,L=0,1,2,3,…,255 where r L Refers to the Lth grayscale, n L Refers to the gray level r in the local block P2 L The number of pixel values; Step 5.2: All candidate target pixels D z ∈D ct The histograms of are accumulated to form the total candidate area histogram. The mathematical formula is as follows: Where n refers to the number of candidate target pixels; Step 5.3, in the total candidate area histogram, use the following formula to calculate the proportion of the histogram of each candidate target: Step 5.4, construct an all-0 matrix GLC with the same size as the original infrared image to store the grayscale contribution eigenvalues, where the calculation formula of the element value GLC(x,y) of the candidate target coordinates is as follows: Among them, (x, y) is the coordinate of the pixel in the grayscale contribution feature map GLC, n refers to the number of candidate target pixels, x z and z Refers to the coordinates of the zth candidate target point; after calculating the grayscale contribution eigenvalues ​​of all candidate target points, the grayscale contribution feature map GLC can be obtained.

8. The infrared small target detection method based on attribute-guided feature weighted fusion as claimed in claim 1, characterized in that: The specific implementation of step 6 includes the following sub-steps: Step 6.1, calculate the fusion weight map w according to the following formula: w=S b / O Where O is the original infrared image, S b is the attribute difference graph; Step 6.2, multiply the fusion weight map, local grayscale difference feature map and contribution feature map, and calculate the fusion map Fu according to the following formula: It was=(w×PLD×GLC) * in(·) * Refers to the normalization operation, PLD is the local grayscale difference feature map, and GLC is the grayscale contribution feature map.

9. The infrared small target detection method based on attribute-guided feature weighted fusion as claimed in claim 1, characterized in that: The specific implementation of step 7 is as follows: Step 7.1, take out the values ​​greater than 0 in the fusion image Fu and store them into the array R, R = {0, Fu(x, y)}, Fu(x, y)>0, where x and y refer to the coordinates of the pixels in the fusion image; then calculate the maximum inter-class variance threshold T of the fusion image Fu; The calculation formula of the maximum inter-class variance threshold T is as follows: Among them, R i The index value of the ith element in R, and N refers to the number of elements less than T and the number of elements greater than or equal to T in the array R, respectively. R is the total number of elements in array R, The average value of the elements in index group R that are less than T. Refers to the average value of elements greater than or equal to T; Step 7.2, segment the fusion image Fu according to the threshold T. After the threshold segmentation, the pixel values ​​in the fusion matrix that are greater than or equal to the threshold are set to 1, and the pixel values ​​that are less than the threshold are set to 0, and a binary image is obtained. The pixels that are set to 1 will be used as the target output.

10. An infrared single small target detection system based on attribute-guided feature weighted fusion, characterized in that: Includes the following modules: The original infrared image acquisition module is used to acquire a single frame of original infrared image to be detected; The attribute difference map acquisition module is used to obtain the bright target attribute difference map corresponding to the original infrared image; The threshold calculation module of the attribute difference map is used to perform threshold segmentation on the attribute difference map to obtain a set of candidate targets; A local grayscale difference feature map acquisition module is used to calculate the local grayscale difference feature of each pixel in the candidate target set to obtain a local grayscale difference feature map; A grayscale contribution feature map acquisition module is used to calculate the grayscale contribution feature of each pixel in the candidate target set to obtain a grayscale contribution feature map; A fusion image acquisition module is used to construct a weighted fusion image based on the attribute difference image, the local grayscale difference feature image and the contribution feature image; The infrared small target acquisition module is used to perform maximum inter-class variance threshold segmentation on the weighted fusion image to obtain the infrared small target detection results.

Images