UAV cruise detection method based on adaptive edge feature semantic segmentation

Abstract

This invention discloses a UAV patrol detection method based on adaptive edge feature semantic segmentation. It utilizes a predefined multi-scale, multi-shape edge detection operator group and a multi-shape edge feature fusion method based on edge probability to obtain multi-shape edge feature maps. A multi-scale, multi-shape edge feature weighting method based on target matching degree is then used to obtain more multi-scale, multi-shape edge feature maps. An improved DeepLabV3+ semantic segmentation network is constructed, using a receptive field fusion and hollow spatial pyramid pooling module in the encoder to supplement the multi-scale, multi-shape edge feature map group into the network decoder. The improved network is then trained. Finally, the trained network is used to segment and detect targets in remote sensing images obtained from UAV patrols. This invention extracts generalized feature representations of patrol targets through training on a UAV patrol dataset, highlighting the discriminability between different target features, improving detection accuracy, and reducing the workload of manual patrols.

Description

Technical Field

[0001] This invention relates to a drone cruise detection method based on adaptive edge feature semantic segmentation, belonging to the fields of semantic segmentation and machine vision. Background Technology

[0002] In recent years, the Internet of Things (IoT) has developed rapidly. As an important component of new digital infrastructure, it is accelerating the digitalization and intelligentization of industries, profoundly changing the technology and industrial system. This has led to a further increase in the demand for unmanned and automated processes in many industries, with drone inspection playing a crucial role. Traditional inspections rely on manual labor, requiring significant investment of manpower, resources, and capital, and also presenting challenges in terms of efficiency and safety. Drones, with their advantages of simple operation, minimal site limitations, and high efficiency, make it possible to solve these problems through drone inspections. Unlike traditional inspections, drone inspections typically feature a wide overhead view, long distance, and high dynamic range, resulting in images that differ from ordinary inspection images. Remote sensing images obtained from drone inspections generally have complex backgrounds, large differences in object scale, and high resolution. Furthermore, the dynamic changes in illumination of the inspected targets and the semantic ambiguity of multiple target edges make existing detection methods ineffective in this scenario.

[0003] In conclusion, how to accurately identify multiple targets in drone inspection images using existing technologies has become a pressing issue for industry professionals. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a UAV cruise detection method based on adaptive edge feature semantic segmentation, which solves the problems of inaccurate edge detection and low target segmentation accuracy in UAV inspection images with large target scale differences in the existing technology.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0006] A method for detecting drone cruise based on adaptive edge feature semantic segmentation includes the following steps:

[0007] Step 1: Define a set of multi-scale, multi-shape edge detection operators;

[0008] Step 2: Using the multi-shape edge feature fusion method based on edge probability, and the multi-scale multi-shape edge detection operator group from Step 1, a multi-shape edge feature map group is obtained.

[0009] Step 3: Based on the target matching degree, the multi-scale edge feature weighting method is used to weight the multi-shape edge feature map obtained in Step 2 according to the target matching degree to obtain a multi-scale multi-shape edge feature map group.

[0010] Step 4: Build a semantic segmentation network and train the semantic segmentation network;

[0011] Step 5: Use the trained semantic segmentation network to detect multiple targets in the UAV inspection images.

[0012] Furthermore, the multi-scale, multi-shape edge detection operator group mentioned in step 1 is defined as follows: the operator group is divided into n categories according to the extracted edge feature shape and m categories according to the operator scale, for a total of n×m edge detection operators.

[0013] Furthermore, the steps of the multi-shape edge feature fusion method based on edge probability described in step 2 are as follows:

[0014] Step 2.1: Convert the image to be detected from an RGB image to a grayscale image, and then obtain n×m edge feature maps according to the multi-scale multi-shape edge detection operator group proposed in Step 1.

[0015] Step 2.2: Define the shape edge strength as the arithmetic mean of the sum of the gray values ​​of adjacent pixels along the edge shape direction, and the relative edge gradient as the arithmetic mean of the sum of the gradient values ​​of pixels along the edge shape direction.

[0016] Step 2.3: Define pixels with non-zero gray values ​​in the edge feature map as edge pixels to be determined, and calculate the average gray value, average shape edge intensity, and average relative edge gradient of the edge pixels to be determined.

[0017] Step 2.4: Calculate the probability that the current pixel is an edge pixel based on its own gray value, shape edge strength, and relative edge gradient.

[0018] Step 2.5: Binarize the edge feature map according to γ(p) to obtain n binarized shape edge feature maps corresponding to each operator size;

[0019] Step 2.6: Perform an OR operation on the n binarized shape edge feature maps corresponding to each operator size, and fuse them into a single multi-shape edge feature map E. S , to obtain a set of multi-shape edge feature maps including m multi-shape edge feature maps.

[0020] Furthermore, the binarization operation specifically involves treating pixels with γ(p) higher than a set threshold as edge pixels and setting their grayscale value to 255; and treating pixels with γ(p) lower than the threshold as non-edge pixels and setting their grayscale value to 0.

[0021] Furthermore, the shape edge intensity of pixel p Relative edge gradient Where Gray(pu) represents the gray value of the u-th pixel in the negative direction of the edge shape of pixel p, Gray(p+u) represents the gray value of the u-th pixel in the positive direction of the edge shape of pixel p, and Gray(p) represents the gray value of pixel p.

[0022] Furthermore, the probability that the current pixel p is an edge pixel. Gray(p), These represent the current pixel's own grayscale value, shape edge intensity, and relative edge gradient, respectively. These represent the average grayscale value, average shape edge intensity, and average relative edge gradient of the edge pixels to be determined, respectively.

[0023] Furthermore, the steps of the multi-scale edge feature weighting method based on target matching degree described in step 3 are as follows:

[0024] Step 3.1, define the high impact factor, the specific calculation method is as follows:

[0025]

[0026] Among them, [Alt] min Alt max ] represents the drone's cruising altitude range, Alt cur This represents the altitude at which the drone acquires the inspection images. k represents the average cruising altitude of the drone. i k j These represent the operator scales corresponding to the i-th and j-th multi-shape edge feature maps within the multi-shape edge feature map group, respectively. The height influence factor represents the i-th multi-shape edge feature map within the multi-shape edge feature map group;

[0027] Define the edge quantity factor The specific calculation method is as follows:

[0028]

[0029] in, Represents an intermediate variable. This represents the number of edge points in the i-th and j-th multi-shape edge feature maps within the multi-shape edge feature map group;

[0030] Step 3.2: Based on the high influence factor and marginal quantity factor obtained in Step 3.1, calculate the target matching degree. The specific calculation method is as follows:

[0031]

[0032] in, This represents the target matching degree corresponding to the i-th multi-shape edge feature map within the multi-shape edge feature map group;

[0033] Step 3.3: Weight the multi-shape edge feature map group according to the target matching degree to obtain a multi-scale multi-shape edge feature map group. The specific calculation method is as follows:

[0034]

[0035] in, This is a weighted multi-scale multi-shape edge feature map of the i-th multi-shape edge feature map within a multi-shape edge feature map group. Let i be the i-th multi-shape edge feature map in the multi-shape edge feature map group.

[0036] Furthermore, in step 4, an improved DeepLabV3+ semantic segmentation network is built. The specific steps are as follows:

[0037] Step 4.1: In the encoder part of the DeepLabV3+ semantic segmentation network, the receptive field fusion hollow space pyramid pooling module is used to adaptively perform weighted fusion on branches with different receptive fields. The construction process of the receptive field fusion hollow space pyramid pooling module is as follows:

[0038] Step 4.1.1: Perform a convolutional layer P1 and the three dilated convolutional layers P2, P3, and P4 in the channel direction to obtain the fused features. in, H represents the three-dimensional features of the feature map, W represents the feature map height, and C represents the feature map depth.

[0039] Step 4.1.2: Pool the fused feature P in the height and width directions to obtain... and

[0040] Step 4.1.3, for Z h and Z w Perform channel concatenation, then adjust the number of channels using a 1×1 convolution operation, then pass it through a BN layer and activate it using the ReLU function to obtain the desired result. Where r is the scaling factor;

[0041] Step 4.1.4: Split f into two independent features. and Then, a 1×1 convolution operation is used for dimensionality increase, and the Sigmoid function is used for activation to obtain... and

[0042] Step 4.1.5, using gh and g w Calculate the weighted output P′, and perform channel concatenation operation between the weighted output P′ and the global average pooling layer output P5 to obtain the overall output P” of the receptive field fusion void space pyramid pooling module;

[0043] Step 4.2: In the decoder part of the DeepLabV3+ semantic segmentation network, the multi-scale and multi-shape edge feature map group is downsampled by four times and then concatenated with the high-level semantic features obtained from the encoder part in the channel direction.

[0044] This invention also provides a drone cruise detection system based on adaptive edge feature semantic segmentation. This system performs drone cruise detection based on the method described above, specifically including:

[0045] The multi-shape edge feature fusion module is used to obtain a set of multi-shape edge feature maps based on a predefined set of multi-scale multi-shape edge detection operators;

[0046] The multi-scale edge feature weighting module is used to weight the edges based on the target matching degree to obtain a multi-scale, multi-shape edge feature map set.

[0047] The semantic segmentation network module is used to build and train a semantic segmentation network.

[0048] The present invention also provides a computer-readable storage medium for storing one or more programs, said one or more programs including instructions that, when executed by a computing device, cause the computing device to perform the method described above.

[0049] The present invention also provides a drone cruise detection device based on adaptive edge feature semantic segmentation, comprising one or more processors, one or more memories, and one or more programs, wherein the one or more programs are stored in the one or more memories and configured to be executed by the one or more processors, and the one or more programs include instructions for performing the method described above.

[0050] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0051] 1. This invention proposes a multi-shape edge feature fusion method based on edge probability and a multi-scale edge feature weighting method based on target matching degree, which can extract the edges of objects more accurately;

[0052] 2. This invention proposes a receptive field fusion hollow space pyramid pooling module, which is suitable for UAV remote sensing images with large differences in target object scale, and improves the accuracy of image semantic segmentation;

[0053] 3. The method of the present invention can accurately segment and identify target objects in the video and image information captured by the drone during drone inspection, providing accurate information for inspection decision-making. Attached Figure Description

[0054] Figure 1 Here is a flowchart of a multi-target semantic segmentation and detection algorithm;

[0055] Figure 2 To experience the integrated hollow space pyramid pooling module structure diagram;

[0056] Figure 3 To improve the overall structure of the DeepLabV3+ network. Detailed Implementation

[0057] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0058] The present invention describes a UAV cruise detection method based on adaptive edge feature semantic segmentation, such as... Figure 1 As shown, it includes the following steps:

[0059] Step 1: Define a set of multi-scale, multi-shape edge detection operators.

[0060] Based on the extracted edge feature shape, they can be divided into n categories, such as vertical / horizontal / 45° left tilt / 45° right tilt, etc. Based on the operator template scale, they can be divided into m categories, such as 3×3, 5×5, 7×7, etc., for a total of n×m edge detection operators.

[0061] Step 2: Using a multi-shape edge feature fusion method based on edge probability, and employing the multi-scale multi-shape edge detection operator group from Step 1, m multi-shape edge feature maps are obtained according to the operator scale, as follows:

[0062] (1) Convert the image to be detected from an RGB image to a grayscale image. Then, based on the multi-scale multi-shape edge detection operator group proposed in step 1, a total of n×m edge feature maps can be obtained from n edge shapes and m operator scales. Among them, each scale operator can obtain a set of n edge feature maps reflecting different edge shapes, denoted as E. v ,v∈[1,n].

[0063] (2) Let the scale of the edge detection operator be k, and the gray value of pixel p be Gray(p), where p represents the pixel position. Define the shape edge intensity of pixel p according to the scale of the operator and the direction of the edge shape. The specific calculation method is as follows:

[0064]

[0065] Where Gray(pu) represents the gray value of the u-th pixel in the negative direction of the edge shape, and Gray(p+u) represents the gray value of the u-th pixel in the positive direction of the edge shape.

[0066] The relative edge gradient of pixel p is defined based on the operator's scale and the direction of the edge shape. In reality, it is the sum of the gradient values ​​in the positive and negative directions of the edge shape of the current pixel p. The specific calculation method is as follows:

[0067]

[0068] (3) Pixels with non-zero gray values ​​in the edge image are called undetermined edge pixels, and their number is denoted as Num. The average gray value of the undetermined edge pixels is then calculated. Average shape edge strength and average relative edge gradient The specific calculation method is as follows:

[0069]

[0070] Then, based on the current pixel p's own grayscale value Gray(p) and the shape edge intensity... Relative edge gradient The probability γ(p) that the current pixel is an edge pixel is calculated as follows:

[0071]

[0072] Then, the edge feature map is binarized according to the probability γ(p). The threshold is set to 0.5. Pixels above the threshold are regarded as edge pixels and their gray value is set to 255. Pixels below the threshold are regarded as non-edge pixels and their gray value is set to 0. The specific calculation method is as follows:

[0073]

[0074] Through the above operations, the edge detection operator at each scale can obtain n binarized shape edge feature maps, denoted as E. i ′, i∈[1,n].

[0075] (4) The binarized shape edge feature map E i A pixel-wise OR operation is performed on ′,i∈[1,n] to fuse n binarized shape edge feature maps into a single multi-shape edge feature map E. S The specific calculation method is as follows:

[0076] E S=E1′⊙E2′......⊙E n ′

[0077] Here, ⊙ represents a pixel-by-pixel OR operation.

[0078] Through the above operations, each scale of edge detection operator can obtain a multi-shape edge feature map E. S Finally, m multi-shape edge feature maps are obtained.

[0079] Step 3: Using a multi-scale edge feature weighting method based on target matching degree, the m multi-shape edge feature maps obtained in Step 2 are weighted according to the target matching degree to obtain a multi-scale multi-shape edge feature map group, as follows:

[0080] (1) Based on step 2, a set of multi-shape edge feature maps is obtained, which consists of m multi-shape edge feature maps, denoted as m.

[0081] (2) As the UAV's cruising altitude increases, the size of objects in the UAV remote sensing image will decrease accordingly. Therefore, the weight assigned to the feature map obtained by the small-scale edge detection operator should be larger, and the weight assigned to the feature map obtained by the large-scale edge detection operator should be smaller. The UAV cruising altitude range is denoted as [Alt]. min Alt max To measure the impact of the drone's cruising altitude on the weights of multi-shape edge feature maps obtained by operators of different scales, an altitude influence factor is defined. The specific calculation method is as follows:

[0082]

[0083] Among them, Alt cur This represents the altitude at which the drone remote sensing image was acquired. k represents the average cruising altitude of the drone. i This represents the operator scale corresponding to the i-th multi-shape edge feature map.

[0084] The number of edge points in each multi-shape edge feature map is denoted as The more edge points there are, the more suitable the operator at that scale is for the image. To measure the impact of the number of edges on the weights of multi-shape edge feature maps obtained by operators at different scales, an edge number factor is defined. The specific calculation method is as follows:

[0085]

[0086] in, It represents an intermediate variable.

[0087] (3) Based on the high impact factor and marginal quantity factor Calculate the target matching degree The specific calculation method is as follows:

[0088]

[0089] Finally, based on the target matching degree For multi-shape edge feature map group E S Weighting is performed to obtain a set of multi-scale, multi-shape edge feature maps. The specific calculation method is as follows:

[0090]

[0091] Step 4: Build an improved DeepLabV3+ semantic segmentation network, such as... Figure 3 The details are as follows:

[0092] (1) Construct a receptive field fusion void space pyramid pooling module, such as Figure 2 As shown, the ordinary convolutional layer P1 and the three dilated convolutional layers P2, P3, and P4 are first spliced ​​along the channel direction to obtain the fused features. in, H represents the three-dimensional features of the feature map, W represents the feature map height, and C represents the feature map depth.

[0093] Then, pooling is performed on the fused feature P in the height and width directions to obtain... and The calculation process is as follows:

[0094]

[0095] Then for Z h and Z w Perform channel concatenation, then adjust the number of channels using a 1×1 convolution operation, pass through a BN layer, and activate using the ReLU function to obtain... The calculation process is as follows:

[0096]

[0097] in, Represents channel splicing operation, kernel 1×1 (·) represents a 1×1 convolution, and r is the scaling factor.

[0098] Then, f is split into two independent features. and Then, a 1×1 convolution operation is used for dimensionality increase, and the Sigmoid function is used for activation to obtain... and The calculation process is as follows:

[0099] g h =Sigmoid(kernel) 1×1 (f h ))

[0100] g w =Sigmoid(kernel) 1×1 (f w )).

[0101] Finally, using g h and g w The weighted output P′ is calculated, and then channel concatenation is performed between the weighted output P′ and the pooling layer output P5 to obtain the overall output P” of the receptive field fusion void space pyramid pooling module. The calculation process is as follows:

[0102]

[0103] (2) In the encoder part, the original hollow space pyramid pooling module is replaced by the receptive field fusion hollow space pyramid pooling module. In the decoder part, the multi-scale multi-shape edge feature map obtained in step 3 is added to the network. After four times downsampling, it is spliced ​​with the high-level semantic features obtained in the encoder part in the channel direction, and the network model is trained.

[0104] This invention discloses a UAV patrol detection method based on adaptive edge feature semantic segmentation. First, a multi-scale, multi-shape edge detection operator group is defined. Second, edge feature maps are extracted using the multi-scale, multi-shape edge detection operator group, and a multi-shape edge feature fusion method based on edge probability is used to obtain multi-shape edge feature maps. Further, a multi-scale edge feature weighting method based on target matching degree is used to weight the multi-shape edge feature maps according to the target matching degree, resulting in a multi-scale, multi-shape edge feature map group. Then, an improved DeepLabV3+ semantic segmentation network is built. In the encoder, a receptive field fusion hollow spatial pyramid pooling module is used to supplement the multi-scale, multi-shape edge feature map group into the network decoder, and the improved DeepLabV3+ semantic segmentation network is trained. Finally, the trained DeepLabV3+ semantic segmentation network is used to segment and detect targets in remote sensing images obtained from UAV patrols. This invention extracts generalized feature representations of patrol targets by training on a UAV patrol dataset, highlighting the distinguishability between different target features, improving detection accuracy, and reducing the workload of manual patrols.

[0105] This invention also proposes a UAV cruise detection system based on adaptive edge feature semantic segmentation, specifically including:

[0106] The multi-shape edge feature fusion module is used to obtain a set of multi-shape edge feature maps based on a predefined set of multi-scale multi-shape edge detection operators;

[0107] The multi-scale edge feature weighting module is used to weight the edges based on the target matching degree to obtain a multi-scale, multi-shape edge feature map set.

[0108] The semantic segmentation network module is used to build and train a semantic segmentation network.

[0109] The technical solution of the above-mentioned UAV cruise detection system is similar to the aforementioned method, and will not be repeated here.

[0110] Based on the same technical solution, the present invention also discloses a computer-readable storage medium for storing one or more programs, the one or more programs including instructions, which, when executed by a computing device, cause the computing device to perform the above-described UAV cruise detection method based on adaptive edge feature semantic segmentation.

[0111] Based on the same technical solution, the present invention also discloses a computing device, including one or more processors, one or more memories, and one or more programs, wherein the one or more programs are stored in the one or more memories and configured to be executed by the one or more processors, and the one or more programs include instructions for executing the above-described UAV cruise detection method based on adaptive edge feature semantic segmentation.

[0112] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0113] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1A device that provides the functions specified in one or more boxes.

[0114] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0115] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0116] The above embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solutions based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

Claims

1. A UAV cruise detection method based on adaptive edge feature semantic segmentation, characterized in that, Includes the following steps: Step 1: Define a set of multi-scale, multi-shape edge detection operators; Step 2: Using the multi-shape edge feature fusion method based on edge probability, and the multi-scale multi-shape edge detection operator group from Step 1, a multi-shape edge feature map group is obtained. Step 3: Based on the target matching degree, the multi-scale edge feature weighting method is used to weight the multi-shape edge feature map obtained in Step 2 according to the target matching degree to obtain a multi-scale multi-shape edge feature map group. Step 4: Build a semantic segmentation network and train the semantic segmentation network; Step 5: Use the trained semantic segmentation network to detect multiple targets in the UAV inspection images; The steps of the multi-scale edge feature weighting method based on target matching degree described in step 3 are as follows: Step 3.1, define the high impact factor, the specific calculation method is as follows: Among them, [Alt] min Alt max ] represents the drone's cruising altitude range, Alt cur This represents the altitude at which the drone acquires the inspection images. k represents the average cruising altitude of the drone. i k j These represent the operator scales corresponding to the i-th and j-th multi-shape edge feature maps within the multi-shape edge feature map group, respectively. The height influence factor represents the i-th multi-shape edge feature map within the multi-shape edge feature map group; Define the edge quantity factor The specific calculation method is as follows: in, Represents an intermediate variable. This represents the number of edge points in the i-th and j-th multi-shape edge feature maps within the multi-shape edge feature map group; Step 3.2: Based on the high influence factor and marginal quantity factor obtained in Step 3.1, calculate the target matching degree. The specific calculation method is as follows: in, This represents the target matching degree corresponding to the i-th multi-shape edge feature map within the multi-shape edge feature map group; Step 3.3: Weight the multi-shape edge feature map group according to the target matching degree to obtain a multi-scale multi-shape edge feature map group. The specific calculation method is as follows: in, This is a weighted multi-scale multi-shape edge feature map of the i-th multi-shape edge feature map within a multi-shape edge feature map group. Let i be the i-th multi-shape edge feature map in the multi-shape edge feature map group.

2. The UAV cruise detection method based on adaptive edge feature semantic segmentation according to claim 1, characterized in that, The multi-scale, multi-shape edge detection operator group mentioned in step 1 is defined as follows: the operator group is divided into n categories according to the extracted edge feature shape and m categories according to the operator scale, for a total of n×m edge detection operators.

3. The UAV cruise detection method based on adaptive edge feature semantic segmentation according to claim 2, characterized in that, The steps of the multi-shape edge feature fusion method based on edge probability described in step 2 are as follows: Step 2.1: Convert the image to be detected from an RGB image to a grayscale image, and then obtain n×m edge feature maps according to the multi-scale multi-shape edge detection operator group proposed in Step 1. Step 2.2: Define the shape edge strength as the arithmetic mean of the sum of the gray values ​​of adjacent pixels along the edge shape direction, and the relative edge gradient as the arithmetic mean of the sum of the gradient values ​​of pixels along the edge shape direction. Step 2.3: Define pixels with non-zero gray values ​​in the edge feature map as edge pixels to be determined, and calculate the average gray value, average shape edge intensity, and average relative edge gradient of the edge pixels to be determined. Step 2.4: Calculate the probability that the current pixel is an edge pixel based on its own gray value, shape edge strength, and relative edge gradient. Step 2.5: Binarize the edge feature map according to γ(p) to obtain n binarized shape edge feature maps corresponding to each operator size; Step 2.6: Perform an OR operation on the n binarized shape edge feature maps corresponding to each operator size, and fuse them into a single multi-shape edge feature map E. S , to obtain a set of multi-shape edge feature maps including m multi-shape edge feature maps.

4. The UAV cruise detection method based on adaptive edge feature semantic segmentation according to claim 3, characterized in that, The binarization operation specifically involves treating pixels with γ(p) higher than a set threshold as edge pixels and setting their grayscale value to 255; and treating pixels with γ(p) lower than the threshold as non-edge pixels and setting their grayscale value to 0.

5. The UAV cruise detection method based on adaptive edge feature semantic segmentation according to claim 3, characterized in that, Shape edge strength of pixel p Relative edge gradient Where Gray(pu) represents the gray value of the u-th pixel in the negative direction of the edge shape of pixel p, Gray(p+u) represents the gray value of the u-th pixel in the positive direction of the edge shape of pixel p, and Gray(p) represents the gray value of pixel p.

6. The UAV cruise detection method based on adaptive edge feature semantic segmentation according to claim 3, characterized in that, The probability that the current pixel p is an edge pixel Gray(p), These represent the current pixel's own grayscale value, shape edge intensity, and relative edge gradient, respectively. These represent the average grayscale value, average shape edge intensity, and average relative edge gradient of the edge pixels to be determined, respectively.

7. The UAV cruise detection method based on adaptive edge feature semantic segmentation according to claim 1, characterized in that, Step 4 involves building an improved DeepLabV3+ semantic segmentation network. The specific steps are as follows: Step 4.1: In the encoder part of the DeepLabV3+ semantic segmentation network, the receptive field fusion hollow space pyramid pooling module is used to adaptively perform weighted fusion on branches with different receptive fields. The construction process of the receptive field fusion hollow space pyramid pooling module is as follows: Step 4.1.1: Perform a convolutional layer P1 and the three dilated convolutional layers P2, P3, and P4 in the channel direction to obtain the fused features. in, H represents the three-dimensional features of the feature map, W represents the feature map height, and C represents the feature map depth. Step 4.1.2: Pool the fused feature P in the height and width directions to obtain... and Step 4.1.3, for Z h and Z w Perform channel concatenation, then adjust the number of channels using a 1×1 convolution operation, then pass it through a BN layer and activate it using the ReLU function to obtain the desired result. Where r is the scaling factor; Step 4.1.4: Split f into two independent features. and Then, a 1×1 convolution operation is used for dimensionality increase, and the Sigmoid function is used for activation to obtain... and Step 4.1.5, using g h and g w Calculate the weighted output P′, and perform channel concatenation operation between the weighted output P′ and the global average pooling layer output P5 to obtain the overall output P” of the receptive field fusion void space pyramid pooling module; Step 4.2: In the decoder part of the DeepLabV3+ semantic segmentation network, the multi-scale and multi-shape edge feature map group is downsampled by four times and then concatenated with the high-level semantic features obtained from the encoder part in the channel direction.

8. A UAV cruise detection system based on adaptive edge feature semantic segmentation, characterized in that, The system performs drone patrol detection based on the method described in any one of claims 1 to 7, specifically including: The multi-shape edge feature fusion module is used to obtain a set of multi-shape edge feature maps based on a predefined set of multi-scale multi-shape edge detection operators; The multi-scale edge feature weighting module is used to weight the edges based on the target matching degree to obtain a multi-scale, multi-shape edge feature map set. The semantic segmentation network module is used to build and train a semantic segmentation network.

9. A computer-readable storage medium storing one or more programs, said one or more programs comprising instructions, characterized in that, When the instruction is executed by the computing device, it causes the computing device to perform the method as described in any one of claims 1 to 7.

Images