An infrared dim and small target detection method based on the clustering idea

Abstract

The present invention discloses an infrared dim and small target detection method based on the clustering idea. Based on the global and local characteristics of infrared dim and small targets, different types of clustering methods are adopted to achieve the detection of dim and small targets. First, the original infrared image is preprocessed using the morphological features of small targets to generate a new density feature map. Second, an improved density peak clustering algorithm is used to roughly locate potential candidate targets. Then, for the local candidate set of potential targets, a weighted fuzzy C-clustering algorithm is adopted to finely segment the target and background regions of the local candidate set, and the difference between the target and the background is used to enhance the target while suppressing false alarms. Finally, the real target is extracted by an adaptive threshold. This method accurately locates the target region through different levels of clustering methods, reduces the interference of background pixels, realizes the detection of infrared dim and small targets with size changes, and effectively improves the target detection rate.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and its main content is an infrared small and weak target detection method for unknown or variable target sizes under complex backgrounds, which is particularly suitable for high-precision detection of targets from far to near under complex backgrounds in an infrared search and tracking system. Background Art

[0002] Infrared detection systems have the advantages of strong concealment, all-weather operation, and the ability to penetrate clouds and fog to a certain extent, and are widely used in fields such as security monitoring and field rescue. However, due to the small size of the target itself or the long distance from the detector, in the images output by the infrared detection system, the target imaging area is small, lacking shape and texture information; at the same time, due to the influence of atmospheric scattering and absorption, the brightness of the target is also relatively weak. Such targets are called infrared small and weak targets. In real application scenarios, the backgrounds of infrared images are complex and variable, and sometimes various complex backgrounds, such as trees, buildings, clouds, and waves, may inevitably appear. They may have brightness exceeding that of real targets and relatively complex edge clutter information, which is likely to interfere with detection and make real targets extremely easy to be submerged in complex backgrounds. In addition, when the target approaches the detector from far to near, the size of the target also changes, which further increases the difficulty of target detection. To sum up, it is still a difficult task to achieve the detection of infrared small and weak targets with unknown sizes under complex backgrounds.

[0003] In the early stage, some traditional filtering methods were mainly used to suppress the background and enhance the target of infrared images through specially designed filters. The advantages of such methods are low computational cost, easy implementation, and suitability for real-time processing, such as Top-Hat transform, Max-median filter, Two-dimensional Least Mean Square (TDLMS) filter, etc. However, in the case of complex backgrounds and low target signal-to-noise ratios, there are a large number of false alarms in such algorithmic methods.

[0004] In recent years, methods based on low-rank sparse recovery have received attention from some scholars. The essence of such methods is to utilize the sparse characteristics of small targets and the non-local self-correlation (low-rank characteristics) of infrared background images to convert the small target detection problem into a mathematical optimization problem of recovering a low-rank sparse matrix, and to achieve the detection of small and weak targets by threshold segmentation of the sparse matrix. Although such algorithms are insensitive to target sizes, they rely on the fitting degree of the target background norm. Since strong edge clutter and high-brightness noise also have sparse characteristics, they are easily misdecomposed into sparse matrices, resulting in a relatively high false alarm rate; in addition, when the background is complex, the low-rank characteristic assumption of infrared background images does not hold, and the detection performance will drop significantly.

[0005] In 2014, inspired by the human visual system (HVS), some scholars introduced HVS into the detection of dim and small infrared targets and proposed the method of local contrast measure (LCM). This algorithm uses a two-layer nested window to capture the target and the background respectively. According to the gray difference between the two, it can enhance the target while suppressing the background. However, this method cannot suppress the highlight background and point noise, and the detection performance is restricted by the proximity degree of the target and the window shape and size. Since such algorithms have good detection performance and real-time performance, many subsequent scholars have proposed many improved algorithms on this basis. The relative local contrast measure (RLCM) constructs a combined contrast measure to reduce the interference of the highlight background, effectively enhances the target, and improves the detection performance. To adapt to targets of different sizes, many improved algorithms use multi-scale windows for adaptive detection. However, multi-scale algorithms will form an "expansion effect", which affects the detection performance. To solve the "expansion effect" problem, some scholars proposed the three-layer window local contrast measure method (TTLCM), which adopts a new three-layer nested window design. Through the difference between different windows, it can achieve detection ability equivalent to that of multi-scale at a single scale and overcome the "expansion effect". The above methods design different window templates and construct more complex contrast measures. These algorithms are sensitive to the target size, and the detection performance is restricted by the fitting degree of the target and the window.

[0006] In summary, in complex backgrounds, the current traditional infrared dim and small target detection algorithms still have the defects of high false alarm rate and low detection rate when facing targets with unknown sizes. The reason for the high false alarm rate is that the complex background cannot be effectively suppressed, and the reason for the low detection rate is that the features of the contrast measure cannot be well extracted. Therefore, the detection algorithm proposed in this invention based on the clustering idea is of great significance for realizing the detection of infrared dim and small targets with unknown sizes in complex backgrounds. Summary of the Invention

[0007] To overcome the deficiencies of the prior art, the present invention provides an infrared dim and small target detection method based on the clustering idea that is robust to target size and has a low false alarm rate, which is used to solve the engineering application problems of unknown target size and high detection false alarm rate in the prior art. This method mainly uses the density peak clustering algorithm (DPC) insensitive to size to perform global search and quickly and roughly extract the coordinates of candidate targets; then, it uses an improved unsupervised fuzzy C-means clustering algorithm to finely segment the neighborhood of the candidate targets, divide target pixels and background pixels, construct an accurate contrast measure, enhance the target pixels, and obtain the final saliency map; finally, an adaptive threshold segmentation is adopted to obtain the target position.

[0008] The above object of the present invention is achieved by the following technical solutions:

[0009] An infrared dim small target detection method based on the clustering idea, characterized in that the global density peak metric of the target is combined with the local contrast metric after local clustering segmentation, and the method includes the following steps:

[0010] 1. Use a new annular structural element to perform morphological filtering on the saliency map. Equations (1) and (2) represent first using annular structure dilation, that is, covering the target area pixel values with the pixel values of the target neighborhood, and then using solid structure erosion, that is, taking the minimum value of the target neighborhood to replace the target area, and subtracting the original image from the image after dilation and erosion to suppress the background and noise.

[0011]

[0012] Among them, f represents the original infrared image; B i and B o are the inner ring radius and the outer ring radius respectively; B b is a solid structure, and the radius is between B i and B o ; B oi represents the combination of two structural elements; f·B oi is a closing operation, first dilating and then eroding; (x, y) is the coordinate of the pixel; ΔB is an annular structure, and the width ΔB = B o -B i ; represents the dilation operation;! represents the erosion operation;

[0013] M(x, y) = f(x, y) - f·B oi (x, y), (2)

[0014] f(x, y) is the gray value of the original infrared image, and M(x, y) represents the saliency map after morphological processing.

[0015] 2. Calculate the pixel density peak value pixel by pixel according to formulas (3) to (6), perform global search and sorting, and take the first N maximum values as candidate targets according to the situation to obtain the coordinates of the candidate targets.

[0016] The density peak γ i is a joint concept of local maximum density and the closest distance.

[0017] γ i = ρ i ×δ i , (3)

[0018] ρ i = M i , (4)

[0019] M i represents the gray value of pixel i in the image after morphological preprocessing.

[0020]

[0021] The nearest relevant distance δ i is calculated by the minimum Euclidean distance between pixel i and a pixel j with a higher density than it.

[0022]

[0023] d ij is the Euclidean distance between two pixels. (x i , y i ) are the coordinates of pixel i, and (x j , y j ) are the coordinates of pixel j.

[0024] 3. With the coordinates of the candidate target as the center, extract N local image patches of size 11×11 centered on the candidate target points.

[0025] Ω = {(x i , y i ) | (|x i - x 0 | <= 5), (|y i - y 0 | <= 5)}, (7)

[0026] Ω represents the set of pixels of the local image patch; (x 0 , y 0 ) are the central pixel coordinates, i.e., the coordinates of the extracted candidate target.

[0027] 4. The target has compactness and heterogeneity. According to formulas (8) to (10), a fuzzy C - means clustering method combined with spatial - information weighting is used for fine segmentation of the local image patches of the candidate target, and the pixels in the region are divided into two categories: target and background.

[0028]

[0029] Among them, c represents the number of clustering centers; I j represents the gray - scale value of the pixel; q is the fuzzy value; v k represents the corresponding clustering center. μ jk represents the membership - degree function of the belonging class and should satisfy the constraint of formula (9).

[0030]

[0031] Among them, w jk is the spatial weight factor; (x 0 , y 0 ) are the coordinates of the central pixel, i.e., the coordinate values of the extracted candidate target; (xj ,y j ) is the coordinate of pixel j in the image block; S is a constant, which is set to twice the size of the local block; 1 / k is used to adjust the weight of the pixel belonging to the kth category pixel. In the present invention, k=2, which is divided into target class and background class pixels.

[0032] 5. Count the two types of pixels, calculate the contrast according to formula (11), and enhance the target type pixels.

[0033] d(T,B)=g T -max(g B ), (11)

[0034] Among them, g T Indicates the target gray value; g B Represents the background gray value; d(T,B) is the difference between the target gray value and the maximum gray value in the background.

[0035] 6. Due to the target compactness, the feature descriptor constraints are generated according to formula (12).

[0036]

[0037] Among them, Area_num represents the number of areas classified as target categories after clustering. F is the feature descriptor.

[0038] 7. According to formula (13), the results of each part are fused to obtain the final saliency map SM.

[0039] SM=F×d(T,B)×g T , (13)

[0040] 8. Calculate the adaptive segmentation threshold T according to formula (14) to perform binary segmentation on SM. If it is greater than the threshold, it is set to 1, otherwise it is set to 0, and the position set to 1 is the target position.

[0041] T=u+λ×δ, (14)

[0042] Among them, u is the mean of all enhanced pixels; δ is the standard deviation of all enhanced pixels; λ is a hyperparameter, the value range of λ is 0.5~1.5, and T is the segmentation threshold.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] 1) Through density peak clustering and global search and calculation of density peaks, the interference of highlight background and structural clutter is effectively suppressed;

[0045] 2) Through weighted fuzzy C-means clustering segmentation, the target size and intensity are adaptively acquired, local contrast is generated in a targeted manner, and the false alarm rate is effectively reduced;

[0046] 3) By combining the global and local attributes of the target, the present invention is robust to the target size and effectively improves the detection rate. Brief Description of the Drawings

[0047] Figure 1 It is a flowchart of the implementation process of the present invention.

[0048] Figure 2 It is the morphological filtering structure element mentioned in the present invention.

[0049] Figure 3 It is the original infrared image used as the input in the present invention. The object to be detected is within the square frame in the figure.

[0050] Figure 4 The candidate objects extracted after density peak clustering are within the white circular frames, including the real objects.

[0051] Figure 5 It is the density peak graph in the present invention. The real object is within the square frame in the figure.

[0052] Figure 6 On the right is the local area of the candidate object in the present invention. On the left is the segmentation result graph of the local area. The white pixels are the object class pixels, and the black area is the background class pixels.

[0053] Figure 7 It is a schematic diagram of the contrast enhancement template in the present invention. The internal area is the object class pixels after segmentation. The area within the white is the possible area where the object class pixels may appear, and the external black area is the background area.

[0054] Figure 8 It is the result graph after threshold segmentation in the present invention. Detailed Embodiment

[0055] Next, the technical solutions in the embodiments of the present invention will be described in detail with reference to the accompanying drawings in the embodiments of the present invention. Several parameters are involved, and these parameters need to be adjusted according to the specific processing environment to achieve good performance. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0056] The test pictures used in the present invention come from the dataset for detecting and tracking small and weak aircraft targets in the infrared sequence images of National University of Defense Technology.

[0057] Simulation environment: Matlab2018b;

[0058] Test images: medium-wave infrared images, resolution 256×256, background is near-earth background;

[0059] Target information: unmanned aerial vehicle.

[0060] Test steps:

[0061] (1) As shown in Figure 1 , first, input the original infrared image Figure 3 , and obtain Figure 2 after filtering with a morphological template. Figure 4 .

[0062] (2) Use Figure 4 as the input to globally search for density peaks, and the density peak sorting is shown in Figure 5 . Extract the first N maximum values of the density peaks as candidate targets, and extract local blocks of size 11×11 centered on the coordinates of the candidate targets.

[0063] (3) Perform fuzzy C-means clustering segmentation with a weighted spatial weight factor on the extracted local image blocks.

[0064] The effect diagram is shown in Figure 6 .

[0065] (4) Obtain the specific pixel category information within the local block after segmentation. Use a window template of size 11×11, and the template is shown in Figure 7 , and directly perform enhancement.

[0066] (5) Finally, adopt adaptive threshold segmentation to obtain the final result. See Figure 8 .

Claims

1. An infrared dim and small target detection method based on clustering idea, characterized in that it includes the following steps: 1) Define a circular structural element for morphological filtering and use the circular structural element for filtering: Among them, f represents the original infrared image; B i and B o are the inner ring radius and the outer ring radius respectively; B b is a solid structure, with a radius between B i and B o ; B oi represents the combination of two structural elements; f·B oi is a closing operation, first dilating and then eroding; (x, y) are the coordinates of the pixel; ΔB is an annular structure, with a width ΔB = B o - B i ; represents the dilation operation;! represents the erosion operation; M(x,y) = f(x,y) - f·B oi (x,y), (2) f(x, y) is the gray value of the original infrared image, and M(x, y) represents the saliency map after morphological processing; 2) Global search, calculate the pixel density peak for each pixel, sort, and take the first N maximum values as candidate targets according to the situation to obtain the coordinates of the candidate targets; Density peak γ i is a combined concept of local maximum density and the nearest neighbor distance, γ i = ρ i × δ i , (3) ρ i = M i , (4) M i represents the gray value of pixel i in the image after morphological preprocessing, The nearest neighbor correlation distance δ i is calculated by the minimum Euclidean distance between pixel i and a pixel j with a higher density than it: d ij is the Euclidean distance between two pixels, where (x i , y i ) are the coordinates of pixel i, and (x j , y j ) are the coordinates of pixel j; 3) Take the coordinates of the candidate targets as the center and extract N local image blocks of size 11×11 centered on the candidate target points; Ω={(x i ,y i )|(|x i -x 0 |≤5),(|y i -y 0 |≤5)}, (7) Ω represents the set of pixels of a local image patch; (x 0 , y 0 ) is the coordinate of the central pixel, i.e., the coordinate value of the extracted candidate target; 4) According to the target point diffusion characteristics, use the fuzzy C-means clustering method combined with spatial information weighting to finely segment the pixel categories in the candidate target local block; Among them, c represents the number of clustering centers; I j represents the gray value of the pixel; q is the fuzzy value; v k represents the corresponding clustering center; μ jk represents the membership function of the belonging category, which should satisfy the constraint of Equation (9); where, w jk is the spatial weight factor; (x 0 , y 0 ) is the coordinate of the central pixel, i.e., the coordinate value of the extracted candidate target; (x j , y j ) is the coordinate of the pixel j in the image block; S is a constant, set to be twice the local block size; 1 / k is used to adjust the weight of the pixel belonging to the k-th category of pixels, k = 2, divided into target class and background class pixels; 5) Locally count the two types of pixels, calculate the contrast, and enhance the target class pixels; d(T,B) = g T -max(g B ), (11) where g T represents the target gray value; g B represents the background gray value; d(T,B) is the difference between the target gray value and the maximum gray value in the background; 6) Generate a feature descriptor constraint according to the prior information of target compactness and heterogeneity; where Area_num represents the number of regions classified as the target class after clustering, and F is the feature descriptor; 7) Fuse the results after each part of the processing to obtain the final saliency map SM; SM = F × d(T,B) × g T , (13) 8) Calculate the adaptive segmentation threshold T according to formula (14) to segment SM and determine the target position; T = u + λ×δ, (14) In formula (14), u is the mean value of all enhanced pixels; δ is the standard deviation of all enhanced pixels; λ is a hyperparameter, and the value range of λ is 0.5 to 1.5, and T is the segmentation threshold.

Images