Intelligent detection method for low-altitude moving target

Abstract

The application discloses a low-altitude moving target intelligent detection method and belongs to the technical scheme of image processing. The application utilizes the difference technology of adjacent satellite remote sensing image frames to capture a moving target region; subsequently, the target region is divided into two sub-regions by extracting wing poles and performing region segmentation, so that a symmetry description image and a contour skeleton image are constructed, and these feature maps are input into a multi-scale fusion neural network, and finally, the type of the moving target is accurately recognized. The application solves the problem of low target recognition accuracy in the prior art.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and particularly relates to a low-altitude moving target intelligent detection method. BACKGROUND

[0002] In the modern aviation field, fixed-wing aircraft has become the core tool in many fields such as civil aviation and environmental monitoring due to its efficient transportation capacity and wide application scenarios. However, with the substantial increase in the frequency of low-altitude airspace use, the types of aerial targets are various, and the flight environment is complex and changeable. The traditional target detection technology has many limitations when facing moving targets such as fixed-wing aircraft in low-altitude airspace. The existing image recognition method mainly relies on the static shape features of the aircraft, such as the fuselage contour and wing shape, and then uses convolutional neural network (CNN) and support vector machine (SVM) for recognition. The existing technology focuses on the extraction and analysis of single feature, lacks comprehensive consideration of multi-dimensional features such as target structure symmetry and contour skeleton, and has the problem of low target recognition accuracy. SUMMARY

[0003] In view of the above problems in the prior art, the low-altitude moving target intelligent detection method provided by the present application solves the problem of low target recognition accuracy in the prior art.

[0004] In order to achieve the above-mentioned application purposes, the technical scheme adopted by the present application is as follows: a low-altitude moving target intelligent detection method, comprising the following steps:

[0005] S1, subtracting adjacent frame satellite remote sensing images to obtain a moving target region;

[0006] S2, extracting two wing pole points of the moving target region, segmenting the moving target region based on a segmentation line to obtain a first moving target sub-region and a second moving target sub-region;

[0007] S3, constructing a symmetry description image according to the symmetry of the first moving target sub-region and the second moving target sub-region;

[0008] S4, extracting an external contour image and an internal contour image of the moving target region, adding the pixel values of the external contour image and the internal contour image according to the pixel points to obtain a contour skeleton image;

[0009] S5, inputting the symmetry description image and the contour skeleton image into a target multi-scale fusion neural network to obtain the type of the moving target.

[0010] Further, S1 comprises the following sub-steps:

[0011] S11, subtract pixel values of satellite remote sensing images of adjacent frames at same pixel point positions, and take absolute values of the subtraction results to obtain a pixel difference value image;

[0012] S12, extract pixel points with pixel values not being 0 in the pixel difference value image to obtain suspected target points;

[0013] S13, extract connected regions formed by the suspected target points;

[0014] S14, merge the connected regions into one moving target region.

[0015] Further, S2 comprises the following steps:

[0016] S21, extract two wing pole points from the moving target region;

[0017] S22, connect the two wing pole points to obtain a wing connection straight line, and construct a segmentation line equation perpendicular to the wing connection straight line and passing through the midpoint of the wing connection straight line;

[0018] S23, according to the segmentation line equation, cut the moving target region into two parts to obtain a first moving target sub-region and a second moving target sub-region.

[0019] Further, S21 comprises the following steps:

[0020] S211, extract four pixel points with any one of maximum abscissa value, minimum abscissa value, maximum ordinate value and minimum ordinate value in the moving target region, and take the four pixel points as pole points;

[0021] S212, in the moving target region, take the average of each pixel value of the pixel points in the neighborhood range of the pole points to obtain a pixel feature;

[0022] S213, extract edge pixel points of the moving target region, take the pole points as the center, obtain a plurality of edge pixel points on both sides, arrange the plurality of edge pixel points on one side in sequence to form a first pixel point set, and arrange the plurality of edge pixel points on the other side in sequence to form a second pixel point set;

[0023] S214, calculate the Euclidean distance between the i-th pixel point coordinate in the first pixel point set and the i-th pixel point coordinate in the second pixel point set to obtain a distance feature, i being a positive integer;

[0024] S215, splice the pixel feature and the distance feature to obtain a feature vector of the pole point;

[0025] S216, calculate the similarity of any two pole points in the feature vector, and select two pole points corresponding to the maximum similarity as the wing pole points.

[0026] Further, the split line equation in S22 is: wherein y is the vertical coordinate on the split line, x is the horizontal coordinate on the split line, x1 is the horizontal coordinate of the first wing pole, x2 is the horizontal coordinate of the second wing pole, y1 is the vertical coordinate of the first wing pole, y2 is the vertical coordinate of the second wing pole, k1 is the slope of the wing connecting straight line, and k2 is the slope of the split line.

[0027] Further, S3 comprises the following sub-steps:

[0028] S31, in the first mobile target sub-region, taking each pixel point as the center, pixel values in a 3*3 neighborhood range are extracted and combined to obtain a first 3*3 combination;

[0029] S32, the frequency of the first 3*3 combination appearing in the second mobile target sub-region is counted, and when the frequency is greater than or equal to 1, the pixel point at the center is marked as a symmetry point;

[0030] S33, in the second mobile target sub-region, taking each pixel point as the center, pixel values in a 3*3 neighborhood range are extracted and combined to obtain a second 3*3 combination;

[0031] S34, the frequency of the second 3*3 combination appearing in the first mobile target sub-region is counted, and when the frequency is greater than or equal to 1, the pixel point at the center is marked as a symmetry point;

[0032] S35, the symmetry of all symmetry points is valued as 1, and the symmetry of all non-symmetry points is valued as 0, to obtain a symmetry description image.

[0033] Further, S4 comprises the following sub-steps:

[0034] S41, on the mobile target region, taking each pixel point as the center, when the number of pixel points in the center 3*3 neighborhood range is less than 8, the pixel point at the center is marked as an outer contour pixel point, the pixel values of other pixel points are set to 0, the pixel values of the outer contour pixel points are reserved, to obtain an outer contour image;

[0035] S42, on the mobile target region, taking each pixel point as the center, when the pixel values of 8 pixel points in the center 3*3 neighborhood range are all the same as the pixel value of the pixel point at the center, the pixel point at the center is marked as a non-contour point, the pixel value of the non-contour point is set to 0, and the pixel values of other pixel points are reserved, to obtain a mobile target contour image;

[0036] ​​S43, in the moving target contour image, there is a pixel point with a pixel value of 0 in the central 3*3 neighborhood range of each pixel point, the pixel point at the center is marked as an inner contour point, the pixel value of other pixel points is set to 0, the pixel value of the inner contour point is retained, and an internal contour image is obtained;

[0037] S44, the pixel values of the external contour image and the internal contour image are added according to the pixel points, and an outline skeleton image is obtained.

[0038] Further, the target multi-scale fusion neural network in S5 comprises: a first scale feature extraction fusion unit, a second scale feature extraction fusion unit, a first residual unit, a second residual unit, a Concat layer, a scale feature enhancement unit, a CNN network and a full connection layer;

[0039] The input end of the first scale feature extraction fusion unit and the input end of the second scale feature extraction fusion unit are both used for inputting the outline skeleton image and the symmetry description image;

[0040] The input end of the first residual unit is connected with the output end of the first scale feature extraction fusion unit; the input end of the second residual unit is connected with the output end of the second scale feature extraction fusion unit; the input end of the Concat layer is connected with the output end of the first residual unit and the output end of the second residual unit respectively, and the output end thereof is connected with the input end of the scale feature enhancement unit; the input end of the CNN network is connected with the output end of the scale feature enhancement unit, and the output end thereof is connected with the input end of the full connection layer; and the output end of the full connection layer is used as the output end of the target multi-scale fusion neural network.

[0041] Further, the first scale feature extraction fusion unit and the second scale feature extraction fusion unit both comprise: a first convolution block, a second convolution block, a third convolution block, a fourth convolution block, a multiplier M1 and a spatial attention module;

[0042] The input end of the first convolution block is used for inputting the symmetry description image, and the output end thereof is connected with the input end of the second convolution block; the input end of the third convolution block is used for inputting the outline skeleton image, and the output end thereof is connected with the input end of the fourth convolution block; the input ends of the multiplier M1 are respectively connected with the output end of the second convolution block and the output end of the fourth convolution block, and the output end thereof is connected with the input end of the spatial attention module; and the output end of the spatial attention module is used as the output end of the first scale feature extraction fusion unit or the second scale feature extraction fusion unit.

[0043] In the first scale feature extraction fusion unit, the convolution kernel of the first convolution block and the third convolution block is 1*1, and the convolution kernel of the second convolution block and the fourth convolution block is 3*3.

[0044] In the second scale feature extraction fusion unit, the convolution kernel of the first convolution block and the third convolution block is 3*3, and the convolution kernel of the second convolution block and the fourth convolution block is 5*5.

[0045] Further, the scale feature enhancement unit comprises: an up-sampling layer, a first convolution layer, a second convolution layer and an adder A1.

[0046] The input end of the up-sampling layer is connected with the input end of the first convolution layer and serves as the input end of the scale feature enhancement unit; the output end of the up-sampling layer is connected with the input end of the second convolution layer; the input ends of the adder A1 are respectively connected with the output end of the first convolution layer and the output end of the second convolution layer, and the output end thereof serves as the output end of the scale feature enhancement unit.

[0047] The present application has the following advantages:

[0048] The present application can effectively exclude background interference and accurately locate the target by subtracting the adjacent frame satellite remote sensing image to obtain the moving target region; in the feature extraction link, the wing pole is extracted and the target region is segmented, and then a symmetry description image is constructed, and a contour skeleton image is obtained, so that the fixed-wing aircraft structure symmetry and the contour multi-dimensional features are fully mined; compared with the traditional method which only relies on the static shape features, the target feature information is greatly enriched. Finally, the multi-dimensional feature image is analyzed by using the target multi-scale fusion neural network, the target features can be captured from different scales, the adaptability to the target in complex environment is enhanced, the recognition deviation caused by single feature extraction and analysis is effectively avoided, and the precision of the fixed-wing aircraft target recognition in low altitude is significantly improved. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 It is a flowchart of a low-altitude moving target intelligent detection method;

[0050] Figure 2 It is a schematic view of a segmentation line on a fixed-wing aircraft;

[0051] Figure 3 It is a structural schematic view of a target multi-scale fusion neural network;

[0052] Figure 4 It is a structural schematic view of a first scale feature extraction fusion unit and a second scale feature extraction fusion unit;

[0053] Figure 5 It is a structural schematic view of a scale feature enhancement unit;

[0054] Figure 6 It is a structural schematic view of a residual unit. DETAILED DESCRIPTION

[0055] The specific embodiments of the present application are described below to facilitate the understanding of the present application for those skilled in the art, but it should be clear that the present application is not limited to the scope of the specific embodiments, and for those skilled in the art, it is obvious that various changes are within the spirit and scope of the present application defined and determined by the appended claims, and all the inventions utilizing the concept of the present application are within the scope of protection.

[0056] As shown in Figure 1 A low-altitude moving target intelligent detection method, comprising the following steps:

[0057] S1, subtracting adjacent frames of satellite remote sensing images to obtain a moving target region;

[0058] S2, extracting two wing extreme points of the moving target region, segmenting the moving target region based on a segmentation line to obtain a first moving target sub-region and a second moving target sub-region;

[0059] S3, constructing a symmetry description image according to the symmetry of the first moving target sub-region and the second moving target sub-region;

[0060] S4, extracting an external contour image and an internal contour image of the moving target region, adding the pixel values of the external contour image and the internal contour image according to the pixel points to obtain a contour skeleton image;

[0061] S5, inputting the symmetry description image and the contour skeleton image into a target multi-scale fusion neural network to obtain the type of the moving target.

[0062] In this embodiment, S1 includes the following sub-steps:

[0063] S11, subtracting the pixel values of the satellite remote sensing images of adjacent frames according to the same pixel point position, and taking the absolute value of the subtraction result to obtain a pixel difference value image;

[0064] S12, extracting the pixel points with non-zero pixel values in the pixel difference value image to obtain suspected target points;

[0065] S13, extracting the connected regions formed by each suspected target point;

[0066] S14, merging the connected regions into one moving target region.

[0067] The present application can sensitively capture the differences caused by target movement by subtracting the pixel values of adjacent frames of satellite remote sensing images and taking the absolute value, effectively suppresses background interference, and highlights the target from the complex background. Then, the suspected target points with non-zero pixel values are extracted to accurately narrow the search range, remove a large amount of irrelevant information, and reduce the calculation amount. The extraction and merging operation of the connected regions of the suspected target points can completely preserve the target morphological structure.

[0068] In the embodiment, S2 comprises the following steps:

[0069] S21, extracting two wing pole points from the moving target region;

[0070] S22, connecting the two wing pole points to obtain a wing connecting straight line, and constructing a segmentation line equation perpendicular to the wing connecting straight line and passing through the midpoint of the wing connecting straight line;

[0071] S23, according to the segmentation line equation, cutting the moving target region into two parts to obtain a first moving target sub-region and a second moving target sub-region, as shown in Figure 2 .

[0072] Connecting the wing pole points to obtain a straight line, based on which a segmentation line perpendicular to the straight line and passing through the midpoint is constructed, which can symmetrically cut the target region into two parts. This operation is consistent with the structural characteristics of the left-right symmetry of the fixed-wing aircraft, so that the two sub-regions after segmentation can more evenly retain the structural information of each part of the aircraft.

[0073] In the embodiment, S21 comprises the following steps:

[0074] S211, extracting four pixel points with any one of the maximum horizontal coordinate value, the minimum horizontal coordinate value, the maximum vertical coordinate value and the minimum vertical coordinate value in the moving target region as pole points;

[0075] S212, in the moving target region, taking the average of each pixel value of the pixel points in the neighborhood range of the pole points to obtain a pixel feature;

[0076] In S212, the neighborhood range size is 5x5;

[0077] S213, extracting edge pixel points from the moving target region, taking the pole points as the center, obtaining a plurality of edge pixel points on both sides, arranging the plurality of edge pixel points on one side in sequence to form a first pixel point set, and arranging the plurality of edge pixel points on the other side in sequence to form a second pixel point set;

[0078] S214, calculating the Euclidean distance between the i-th pixel point coordinate in the first pixel point set and the i-th pixel point coordinate in the second pixel point set to obtain a distance feature, i being a positive integer;

[0079] S215, concatenating the pixel feature and the distance feature to obtain a feature vector of the pole point;

[0080] S216, calculating the similarity of any two pole points in the feature vector, and selecting two pole points corresponding to the maximum similarity as the wing pole points.

[0081] In this embodiment, cosine similarity is selected in S216.

[0082] In this embodiment, edge pixels are extracted from the moving target area. When the number of pixels in the 3×3 neighborhood of a pixel in the moving target area is less than 8 (excluding the background), the center pixel is the edge pixel. With the pole as the center, M consecutive pixels on each side of the pole are obtained and arranged in order of distance from the pole to form a pixel set, where M is a positive integer greater than 2.

[0083] This invention extracts the extreme points of the boundary of the moving target region as initial extreme points to quickly locate key parts of the target and narrow the search range. The average value of the pixels in the neighborhood of the extreme points is taken to effectively filter image noise, making the pixel features more representative of the real target region. Edge pixels are extracted and a set of pixels on both sides is constructed. The Euclidean distance between the corresponding coordinates of the pixels on both sides is calculated to quantify the symmetry around the extreme points from a geometric perspective, matching the left-right symmetry of an aircraft wing and facilitating the distinction between the wing and other parts of the fuselage. Pixel features and distance features are fused into a feature vector to comprehensively describe the extreme point attributes from both pixel and geometric dimensions. Finally, extreme point similarity is calculated based on the feature vector, and the similarity of the wing feature distribution is used to accurately select wing extreme points, eliminating interference from other fuselage structures.

[0084] In this embodiment, the equation of the dividing line in S22 is: , , In this equation, y is the ordinate of the dividing line, x is the abscissa of the dividing line, x1 is the abscissa of the first wing pole, x2 is the abscissa of the second wing pole, y1 is the ordinate of the first wing pole, y2 is the ordinate of the second wing pole, k1 is the slope of the line connecting the wings, and k2 is the slope of the dividing line. When the abscissas of the two wing poles are the same, k1 does not exist, and the slope of the dividing line is 0.

[0085] The specific process of S23 is as follows: For each pixel in the moving target area... ,exist Then the pixel point Assigned to the first moving target sub-region; in Then the pixel point It was assigned to the second moving target sub-region.

[0086] In this embodiment, S3 includes the following sub-steps:

[0087] S31. In the first moving target sub-region, with each pixel as the center, extract the pixel values ​​of a 3×3 neighborhood range and combine them to obtain the first 3×3 combination.

[0088] S32, count the frequency of the first 3x3 combination appearing in the second moving target sub-region, and mark the pixel point at the center as a symmetry point when the frequency is greater than or equal to 1;

[0089] S33, in the second moving target sub-region, take each pixel point as the center, extract the pixel values in the 3x3 neighborhood range for combination, and obtain the second 3x3 combination;

[0090] S34, count the frequency of the second 3x3 combination appearing in the first moving target sub-region, and mark the pixel point at the center as a symmetry point when the frequency is greater than or equal to 1;

[0091] S35, assign the symmetry of all symmetry points to 1, and assign the symmetry of all non-symmetry points to 0, to obtain a symmetry description image.

[0092] When there are less than 8 pixel values in the 3x3 neighborhood range of the pixel point in S31 and S33, that is, the pixel point is at the edge, the missing position is filled with 0.

[0093] From the structural characteristics of the target, the fixed-wing aircraft has the characteristics of left-right symmetry in structure. By extracting the neighborhood pixel value combination in two sub-regions with each pixel point as the center, the local feature distribution of the target on the image is fully considered. The frequency of the combination appearing in the other sub-region is counted, and the symmetry points are marked accordingly. The similarity of the symmetry structure in the pixel distribution is utilized, which can accurately capture the symmetry information of the target. The symmetry points and non-symmetry points are assigned respectively, and finally a symmetry description image is obtained. This quantitative method can highlight the symmetry characteristics of the target, so that the symmetry information of the target can be more accurately input and analyzed in the subsequent target multi-scale fusion neural network, and the recognition ability of the fixed-wing aircraft target is enhanced.

[0094] 3x3 combination refers to a 3x3 pixel point region.

[0095] In this embodiment, S4 includes the following sub-steps:

[0096] S41, on the moving target region, take each pixel point as the center, and mark the pixel point at the center as an outer contour pixel point when the number of pixel points in the center 3x3 neighborhood range is less than 8. Set the pixel value of other pixel points to 0, and retain the pixel value of the outer contour pixel point to obtain an outer contour image;

[0097] S42, on the moving target region, take each pixel point as the center, and mark the pixel point at the center as a non-contour point when the pixel values of the 8 pixel points in the center 3x3 neighborhood range are the same as the pixel value of the pixel point at the center. Set the pixel value of the non-contour point to 0, and retain the other pixel values to obtain a moving target contour image;

[0098] S43. In the moving target contour image, with each pixel as the center, there are pixels with a pixel value of 0 in the 3×3 neighborhood of the center. Mark the pixel at the center as the inner contour point, set the pixel value of other pixels to 0, and retain the pixel value of the inner contour point to obtain the inner contour image.

[0099] S44. Add the pixel values ​​of the outer contour image and the inner contour image to obtain the contour skeleton image.

[0100] In the moving target region, each pixel is centered on a neighborhood of fewer than 8 pixels and is marked as an outer contour pixel. This operation utilizes the spatial distribution characteristics of the target's outer contour pixels. Because the target's outer contour is usually located at the edge of the pixel distribution in the image, with fewer pixels in its neighborhood, this method can accurately identify the target's outer contour pixels.

[0101] For processing the contour image of a moving target, a pixel is used as the center. When the pixel values ​​of eight neighboring pixels are all the same as the pixel value at the center, it is marked as a non-contour point and its pixel value is set to 0. Other pixel values ​​are retained, and pixel values ​​in non-contour regions are discarded. Considering that the inner contour point is located within the target's internal contour, the presence of a pixel with a value of 0 in its neighborhood means that the point is at the contour edge. In this way, the inner contour features of the target can be accurately extracted, and the target's internal contour information can be completely preserved.

[0102] Finally, the outer contour image and the inner contour image are summed pixel by pixel to obtain the contour skeleton image. This fusion operation integrates the outer contour and inner contour information, comprehensively reflecting the contour structure of the target from multiple angles. This makes the contour skeleton image contain complete information about the target contour and eliminates redundant texture transition areas.

[0103] like Figure 3 As shown, the target multi-scale fusion neural network in S5 includes: a first-scale feature extraction and fusion unit, a second-scale feature extraction and fusion unit, a first residual unit, a second residual unit, a Concat layer, a scale feature enhancement unit, a CNN network, and a fully connected layer;

[0104] The input terminals of the first-scale feature extraction and fusion unit and the second-scale feature extraction and fusion unit are both used to input the contour skeleton image and the symmetry description image.

[0105] The input end of the first residual unit is connected with the output end of the first scale feature extraction fusion unit; the input end of the second residual unit is connected with the output end of the second scale feature extraction fusion unit; the input end of the Concat layer is connected with the output end of the first residual unit and the output end of the second residual unit respectively, and the output end thereof is connected with the input end of the scale feature enhancement unit; the input end of the CNN network is connected with the output end of the scale feature enhancement unit, and the output end thereof is connected with the input end of the full connection layer; and the output end of the full connection layer is the output end of the target multi-scale fusion neural network.

[0106] The first scale feature extraction fusion unit and the second scale feature extraction fusion unit both process the symmetry description image and the contour skeleton image, extract target features of different scales, and the residual unit introduces a function of relieving gradient disappearance, enables the network to learn more complex feature transformation, optimizes feature representation, and enhances feature discrimination. The Concat layer splices and fuses different scale features, comprehensively integrates multi-scale information, enables the network to obtain more rich and comprehensive target information, and helps more accurate target recognition; and the scale feature enhancement unit further strengthens the fused features, highlights key features, and suppresses useless information, so that the subsequent CNN network can better focus on effective features and improve detection precision.

[0107] As shown in Figure 4 The first scale feature extraction fusion unit and the second scale feature extraction fusion unit both include a first convolution block, a second convolution block, a third convolution block, a fourth convolution block, a multiplier M1 and a spatial attention module.

[0108] The input end of the first convolution block is used for inputting the symmetry description image, the output end thereof is connected with the input end of the second convolution block; the input end of the third convolution block is used for inputting the contour skeleton image, the output end thereof is connected with the input end of the fourth convolution block; the input ends of the multiplier M1 are respectively connected with the output end of the second convolution block and the output end of the fourth convolution block, and the output end thereof is connected with the input end of the spatial attention module; and the output end of the spatial attention module is the output end of the first scale feature extraction fusion unit or the second scale feature extraction fusion unit.

[0109] In the first scale feature extraction fusion unit, the convolution kernel of the first convolution block and the third convolution block is 1x1, and the convolution kernel of the second convolution block and the fourth convolution block is 3x3.

[0110] In the second scale feature extraction fusion unit, the convolution kernel of the first convolution block and the third convolution block is 3x3, and the convolution kernel of the second convolution block and the fourth convolution block is 5x5.

[0111] The step length of the first convolution block, the second convolution block, the third convolution block and the fourth convolution block is set to 1. The convolution block includes a convolution layer, an activation function layer and a normalization layer.

[0112] In the two-scale feature extraction and fusion unit, different sizes of convolution kernels are set to extract features of different scales.

[0113] The spatial attention module is a Spatial Attention Module.

[0114] In the scale feature extraction and fusion unit, the symmetry description image and the contour skeleton image are respectively subjected to feature extraction, and then a multiplier M1 is used to perform Hadamard product operation, so that the features of the symmetry description image and the features of the contour skeleton image are mutually enhanced, and the fixed-wing aircraft recognition accuracy is improved.

[0115] As shown in Figure 5 , the scale feature enhancement unit includes an up-sampling layer, a first convolution layer, a second convolution layer and an adder A1.

[0116] The input end of the up-sampling layer is connected with the input end of the first convolution layer and serves as the input end of the scale feature enhancement unit; the output end of the up-sampling layer is connected with the input end of the second convolution layer; the input ends of the adder A1 are respectively connected with the output end of the first convolution layer and the output end of the second convolution layer, and the output end thereof serves as the output end of the scale feature enhancement unit.

[0117] In this embodiment, the convolution kernel size of the first convolution layer and the second convolution layer is 3x3, the step length of the first convolution layer is set to 1, and the step length of the second convolution layer is set to 2.

[0118] The up-sampling layer enlarges the size of the input feature map, so that the feature map contains more spatial detail information. After up-sampling, the feature map size is controlled by the second convolution layer (step length is 2) to make it fused with the output of the first convolution layer, and the feature is further enhanced.

[0119] In this embodiment, the structures of the first residual unit and the second residual unit are as shown in Figure 6 .

[0120] In this embodiment, the classification types include: the target is a fixed-wing aircraft and the target is not a fixed-wing aircraft.

[0121] The present application obtains a moving target region by subtracting adjacent frame satellite remote sensing images, effectively eliminates background interference, and accurately locates the target; in the feature extraction link, wing poles are extracted and the target region is segmented, and then a symmetry description image is constructed, and a contour skeleton image is obtained at the same time, the fixed-wing aircraft structure symmetry and contour multi-dimensional features are fully tapped, compared with the traditional method which only relies on static shape features, the target feature information is greatly enriched. Finally, the multi-scale fusion neural network is used for analyzing the multi-dimensional feature image, the target features can be captured from different scales, the adaptability to the target in complex environment is enhanced, the recognition deviation caused by single feature extraction and analysis is effectively avoided, and the precision of the fixed-wing aircraft target recognition in low altitude is significantly improved.

[0122] The above is only the preferred embodiment of the present application and is not used to limit the present application. For those skilled in the art, the present application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A low-altitude moving target intelligent detection method, characterized in that, The method comprises the following steps: S1, subtracting satellite remote sensing images of adjacent frames to obtain a moving target region; S2, extracting two wing extreme points of the moving target region, segmenting the moving target region based on a segmentation line to obtain a first moving target sub-region and a second moving target sub-region; S3, constructing a symmetry description image according to the symmetry of the first moving target sub-region and the second moving target sub-region; S4, extracting an external contour image and an internal contour image of the moving target region, adding pixel values of the external contour image and the internal contour image according to pixel points to obtain a contour skeleton image; S5, inputting the symmetry description image and the contour skeleton image into a target multi-scale fusion neural network to obtain a type of the moving target; The S3 comprises the following steps: S31, in the first moving target sub-region, extracting pixel values in a 3*3 neighborhood range to obtain a first 3*3 combination; S32, counting a frequency of the first 3*3 combination in the second moving target sub-region, and marking a pixel point at the center as a symmetry point when the frequency is greater than or equal to 1; S33, in the second moving target sub-region, extracting pixel values in a 3*3 neighborhood range to obtain a second 3*3 combination; S34, counting a frequency of the second 3*3 combination in the first moving target sub-region, and marking a pixel point at the center as a symmetry point when the frequency is greater than or equal to 1; S35, assigning a symmetry value of 1 to all symmetry points and a symmetry value of 0 to all non-symmetry points to obtain the symmetry description image; The S4 comprises the following steps: S41, in the moving target region, marking a pixel point at the center as an external contour pixel point when the number of pixel points in a 3*3 neighborhood range at the center is less than 8, setting pixel values of other pixel points to 0, retaining pixel values of the external contour pixel point, and obtaining an external contour image; S42, in the moving target region, marking a pixel point at the center as a non-contour point when pixel values of 8 pixel points in a 3*3 neighborhood range at the center are all the same as a pixel value of the pixel point at the center, setting the pixel value of the non-contour point to 0, retaining other pixel values, and obtaining a moving target contour image; S43, in the moving target contour image, marking a pixel point at the center as an internal contour point when there is a pixel point with a pixel value of 0 in a 3*3 neighborhood range at the center, setting pixel values of other pixel points to 0, retaining the pixel value of the internal contour point, and obtaining an internal contour image; S44, adding pixel values of the external contour image and the internal contour image according to pixel points to obtain a contour skeleton image. 2.The low-altitude moving target intelligent detection method according to claim 1, characterized in that, The S1 comprises the following steps: S11, subtracting pixel values of satellite remote sensing images of adjacent frames according to the same pixel point position, and taking an absolute value of a subtraction result to obtain a pixel difference value image; S12, extracting pixel points with a pixel value of 0 in the pixel difference value image to obtain suspected target points; S13, extracting connected regions formed by the suspected target points; S14, merge the contacted communication areas into one moving target area. 3.The low-altitude moving target intelligent detection method according to claim 1, characterized in that, The S2 comprises the following steps: S21, extract two wing pole points from the moving target area; S22, connect the two wing pole points to obtain a wing connection straight line, and construct a segmentation line equation perpendicular to the wing connection straight line and passing through the midpoint of the wing connection straight line; S23, according to the segmentation line equation, the moving target area is divided into two parts to obtain a first moving target sub-area and a second moving target sub-area. 4.The low-altitude moving target intelligent detection method according to claim 3, characterized in that, The S21 comprises the following steps: S211, extract four pixel points with any one of the maximum horizontal coordinate value, the minimum horizontal coordinate value, the maximum vertical coordinate value and the minimum vertical coordinate value in the moving target area as the pole points; S212, in the moving target area, take the average of each pixel value of the pixel points in the neighborhood range of the pole points to obtain a pixel feature; S213, extract edge pixel points from the moving target area, take a plurality of edge pixel points on one side as the center to obtain a first pixel point set, and take a plurality of edge pixel points on the other side as the center to obtain a second pixel point set; S214, calculate the Euclidean distance between the i-th pixel point coordinates in the first pixel point set and the i-th pixel point coordinates in the second pixel point set to obtain a distance feature, i is a positive integer; S215, concatenate the pixel feature and the distance feature to obtain a feature vector of the pole point; S216, calculate the similarity of any two pole points in the feature vector, and select the two pole points corresponding to the maximum similarity as the wing pole points. 5.The low-altitude moving target intelligent detection method according to claim 3, characterized in that, The split line equation in S22 is: , , wherein y is the vertical coordinate on the split line, x is the horizontal coordinate on the split line, x1 is the horizontal coordinate of the first wing pole, x2 is the horizontal coordinate of the second wing pole, y1 is the vertical coordinate of the first wing pole, y2 is the vertical coordinate of the second wing pole, k1 is the slope of the wing connecting straight line, and k2 is the slope of the split line. 6.The low-altitude moving target intelligent detection method according to claim 1, characterized in that, The target multi-scale fusion neural network in the S5 comprises a first scale feature extraction fusion unit, a second scale feature extraction fusion unit, a first residual unit, a second residual unit, a Concat layer, a scale feature enhancement unit, a CNN network and a full connection layer; The input end of the first scale feature extraction fusion unit and the input end of the second scale feature extraction fusion unit are used to input the contour skeleton image and the symmetry description image; The input end of the first residual unit is connected with the output end of the first scale feature extraction fusion unit; the input end of the second residual unit is connected with the output end of the second scale feature extraction fusion unit; the input end of the Concat layer is connected with the output end of the first residual unit and the output end of the second residual unit respectively, and the output end thereof is connected with the input end of the scale feature enhancement unit; the input end of the CNN network is connected with the output end of the scale feature enhancement unit, and the output end thereof is connected with the input end of the full connection layer; the output end of the full connection layer is used as the output end of the target multi-scale fusion neural network. 7.The low-altitude moving target intelligent detection method according to claim 6, characterized in that, The first scale feature extraction fusion unit and the second scale feature extraction fusion unit both comprise a first convolution block, a second convolution block, a third convolution block, a fourth convolution block, a multiplier M1 and a spatial attention module; The input end of the first convolutional block is configured to input the symmetry description image, the output end of the first convolutional block is connected with the input end of the second convolutional block, the input end of the third convolutional block is configured to input the contour skeleton image, the output end of the third convolutional block is connected with the input end of the fourth convolutional block, the input ends of the multiplier M1 are respectively connected with the output end of the second convolutional block and the output end of the fourth convolutional block, and the output end of the multiplier M1 is connected with the input end of the spatial attention module; the output end of the spatial attention module is used as the output end of the first scale feature extraction and fusion unit or the second scale feature extraction and fusion unit; In the first scale feature extraction and fusion unit, the convolution kernel of the first convolutional block and the third convolutional block is 1×1, and the convolution kernel of the second convolutional block and the fourth convolutional block is 3×3; In the second scale feature extraction and fusion unit, the convolution kernel of the first convolutional block and the third convolutional block is 3×3, and the convolution kernel of the second convolutional block and the fourth convolutional block is 5×5. 8.The low-altitude moving target intelligent detection method according to claim 6, characterized in that, The scale feature enhancement unit comprises an up-sampling layer, a first convolutional layer, a second convolutional layer and an adder A1. The input end of the up-sampling layer is connected with the input end of the first convolutional layer and is used as the input end of the scale feature enhancement unit, the output end of the up-sampling layer is connected with the input end of the second convolutional layer, the input ends of the adder A1 are respectively connected with the output end of the first convolutional layer and the output end of the second convolutional layer, and the output end of the adder A1 is used as the output end of the scale feature enhancement unit.

Images