A UAV target detection method and system based on multispectral information fusion
By using a dual-stream network structure convolutional neural network model with multi-spectral information fusion in drone target detection, the problem of strong concealment of drone targets in complex scenarios is solved, and high-precision drone target detection is achieved.
Patent Information
- Application Number
- CN202311759016.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-20
- Publication Date
- 2025-02-14
- Estimated Expiration
- 2043-12-20
AI Technical Summary
In complex scenarios, drone targets are highly concealed, traditional detection methods have low detection accuracy, making it difficult to effectively identify and locate drone targets.
Using a drone object detection method based on multi-spectral information fusion, the original channel image and spectral index fusion image of multi-spectral images are extracted and fused through a convolutional neural network model with a dual-stream network structure to achieve object detection.
In a complex background, the accuracy and robustness of drone target detection are significantly improved, and the drone target can be effectively identified and positioned, improving the detection difficulty in the prior art.
Smart Images
Figure CN117789062B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of target detection, and in particular to a method and system for detecting unmanned aerial vehicle targets based on multispectral information fusion. Background Art
[0002] With the rapid development of drone technology, the potential safety hazards brought by drones have become increasingly prominent. For air safety, drone detection systems have become an important technical tool. At present, domestic and foreign scholars have carried out a lot of research in the field of drone detection and countermeasure technology. The main detection methods include optical images, radar, spectrum, audio information and multi-sensor fusion, and the various detection methods have different effects. Among them, the optical image detection scheme shoots the optical image of the scene and detects and locates the target based on the detection algorithm. It has the advantages of low cost and high detection accuracy. It is currently an important anti-unmanned monitoring method.
[0003] In recent years, with the rapid development of deep learning technology, target detection algorithms have also shifted from traditional algorithms based on manual features to detection technologies based on deep neural networks. Using convolutional neural networks, image features can be automatically learned, and the accuracy and robustness of target detection are significantly improved. According to different detection ideas, target detection algorithms can be divided into two-stage target detection algorithms and single-stage target detection algorithms. Among them, the two-stage target detection algorithm is also called the target detection algorithm based on region proposal. It converts the detection problem into the classification problem of the local image in the generated proposal area through explicit region proposal. Representative two-stage target detection algorithms include R-CNN, Mask R-CNN, FPN, etc. Single-stage target detection algorithms are also called regression-based target detection algorithms. This type of algorithm does not directly generate the region of interest but regards the target detection task as a regression task for the entire image. Representative algorithms include YOLO, SSD, RetinaNet series, etc.
[0004] Multispectral detection is a passive detection method that can reflect the response of light at different wavelengths, just like discovering the optical "genes" of different substances, revealing the essential physical properties of the detection target, and can break through the interference of complex environments, lighting changes, and camouflage materials, effectively improving the accuracy of target detection. In drone detection tasks, due to the complex operating environment of drones (for example: vegetation, water bodies, sky, low light, etc.) and the small area of small drone targets in the image, the detection accuracy of traditional target detection methods based on visible light sensors is low. The method based on multispectral detection has the advantages of both imaging and spectral detection, which is conducive to the recognition and detection of drone targets in complex backgrounds. However, there is relatively little research on multispectral images in target detection scenarios, especially for drone detection, and a deep learning network architecture based on multispectral deep target detection has not yet been formed. Summary of the invention
[0005] In view of the problem that UAV targets are highly concealed and difficult to detect in complex scenes, the present invention provides a UAV target detection method based on multi-spectral information fusion, in order to improve the existing technology.
[0006] The technical solution provided by the present invention is as follows:
[0007] A method for detecting unmanned aerial vehicle targets based on multispectral information fusion, characterized in that it comprises the following steps:
[0008] S1. Collect multispectral images to build a dataset
[0009] Acquire multispectral images containing UAV targets collected by a multispectral camera, perform channel registration and target annotation on the collected multispectral images, and construct a dataset for training and testing the UAV target detection network model;
[0010] S2. Build a UAV target detection network model based on multi-spectral information fusion
[0011] The UAV target detection network model is a convolutional neural network model with a dual-stream network structure, which performs feature fusion processing on the original channel image and the spectral index fusion image of the input multispectral image after feature extraction respectively;
[0012] S3. Training of UAV target detection network model based on multi-spectral information fusion
[0013] Using the data set constructed in step S1 to train the drone target detection network model constructed in step S2, to obtain a trained drone target detection network model;
[0014] S4. Implement UAV target detection based on multi-spectral information fusion
[0015] Channel registration is performed on the multispectral image to be detected collected by the multispectral camera, and the spectral index fusion image is calculated. The original channel image of the registered multispectral image and the calculated spectral index fusion image are input into the UAV target detection network model trained in step S3 to obtain the target information output by the model.
[0016] On the basis of the above scheme, further improved or preferred schemes also include:
[0017] Furthermore, the multispectral image captured by the multispectral camera includes blue, green, red and near-infrared spectral bands.
[0018] Furthermore, the process of processing the input image data by the drone target detection network model built in step S2 includes:
[0019] S21, feature extraction stage: using two independent and identical feature extraction backbone networks to extract features from the original channel image and the spectral index fusion image respectively;
[0020] S22, feature fusion stage: based on the feature fusion module, the features extracted by the two feature extraction backbone networks are fused, and the fusion features are extracted, and the fusion feature map is output;
[0021] S23, target detection stage: perform target detection on the fused feature map output by step S22, and output the detected target information.
[0022] Furthermore, the method for generating the spectral index fusion image comprises the following steps:
[0023] 1) For the registered multispectral image, the normalized vegetation index and normalized water index of each pixel are calculated to obtain the normalized vegetation index map and the normalized water index map. The calculation formula is:
[0024]
[0025]
[0026] Wherein, NIR, Red, and Green represent the reflectance of the near infrared spectral band, the red spectral band, and the green spectral band of the pixel point, respectively; the normalized vegetation index map refers to an image in which the pixel value is represented by the NDVI value, and the normalized water index map refers to an image in which the pixel value is represented by the NDWI value;
[0027] 2) For the normalized vegetation index map obtained in step 1), a segmentation threshold d is set to obtain pixel points in the normalized vegetation index map whose pixel values are not less than the segmentation threshold, so as to extract the vegetation area and generate a mask map, which is expressed as:
[0028]
[0029] Among them, p NDVI Represents the pixel value of any pixel point p in the normalized vegetation index map, I Mask represents the pixel value of pixel p on the mask image; the above formula indicates that if the pixel value of pixel p in the normalized vegetation index image is not less than the preset segmentation threshold, the pixel value of pixel p in the mask image I Mask Set to 1, otherwise set to 0;
[0030] 3) Generate a spectral index fusion image using the mask image obtained in step 2) by:
[0031] For the pixel points with a pixel value of 1 in the mask image, fill in the pixel value of the pixel point in the normalized vegetation index map; for the pixel points with a pixel value of 0 in the mask image, fill in the pixel value of the pixel point in the normalized water index map. The expression formula is:
[0032] I Fusion =I NDVI *I Mask +I NDWI *(1-I Mask )
[0033] Among them, I Fusion Represents the pixel value of a pixel in the spectral index fusion image, I NDVI ,I NDWI ,I Mask They are respectively the pixel values corresponding to the same pixel position on the normalized vegetation index map, the normalized water index map, and the mask map, and * is the multiplication symbol.
[0034] Furthermore, in the step 2), the segmentation threshold d set for the normalized vegetation index map obtained in the step 1) has a value range of d≥0.
[0035] Furthermore, in the process of training the model in step S3, a loss function is used to measure the distance between the output and the data label, and the network model weight is updated to reduce the loss function value, and finally the network model weight with the minimum loss function value is saved.
[0036] A detection system for implementing the above-mentioned unmanned aerial vehicle target detection method based on multispectral information fusion, characterized in that it includes:
[0037] A data set construction module is used to store multispectral images containing UAV targets collected by a multispectral camera, and to perform channel registration and target labeling on the collected multispectral images to obtain a data set consisting of sample images with labels;
[0038] UAV target detection network model building module, used to build a convolutional neural network model for identifying UAV targets;
[0039] A UAV target detection network model training module is used to train the convolutional neural network model using a data set and output a trained UAV target detection network model;
[0040] The target prediction module is used to detect the multispectral image to be processed using the trained UAV target detection network model and output the detected target information.
[0041] Further, the convolutional neural network model includes a backbone part, a neck connection layer and a detection head;
[0042] The backbone part is composed of two parallel feature extraction backbone networks, wherein the input of the first feature extraction backbone network is the original channel image of the multispectral image, and the input of the second feature extraction backbone network is the spectral index fusion image;
[0043] The neck connection layer fuses the features extracted by the two feature extraction backbone networks based on the feature fusion module, and further extracts features from the fused feature map;
[0044] The detection head is used to perform target detection on the fused feature map and output the detected target information.
[0045] A computer system, comprising a processor and a memory, characterized in that:
[0046] The memory is used to store computer programs and instructions;
[0047] When the processor runs the computer program according to the instructions, the UAV target detection method based on multi-spectral information fusion as described above is implemented.
[0048] A computer-readable storage medium having a computer program stored thereon, characterized in that:
[0049] When the computer program is executed by the processor, the UAV target detection method based on multispectral information fusion as described above is implemented.
[0050] The beneficial effects of the present invention are:
[0051] The method and system of the present invention are implemented based on multispectral images, analyze and design the spectral characteristics of the background environment and the UAV target in the detection task, and propose a UAV target detection network model with multispectral information fusion. The design of the dual-feature extraction backbone network in the network model framework can be obtained by transformation on the basis of most current single-stage target detection algorithm network models, and has certain universality. The dual-feature extraction backbone network in the present invention takes the spectral channel image and spectral index fusion image of the multispectral image as input. After experimental verification, after the fusion of the input features of these two items, the performance of the method of the present invention for UAV target detection is significantly improved compared with the existing detection methods. The present invention aims at the problem that UAV targets are highly concealed and difficult to detect in complex scenes, and realizes high-precision UAV target detection. BRIEF DESCRIPTION OF THE DRAWINGS
[0052] Figure 1 is a flow chart of the drone target detection method of the present invention;
[0053] Figure 2 It is a schematic diagram of the network structure of the drone target detection in the present invention;
[0054] Figure 3 It is a schematic diagram of the process of generating an exponential fusion image in the present invention;
[0055] Figure 4 It is a comparison chart of spectral channel image and spectral index image of multi-spectrum;
[0056] Figure 5 It is an experimental effect diagram of the method of the present invention for detecting UAV targets in various environments. DETAILED DESCRIPTION
[0057] The present invention will be described in more detail below in conjunction with specific embodiments and accompanying drawings.
[0058] Embodiment 1:
[0059] The multispectral images used in the experiment of this embodiment are collected based on the RedEdge five-channel multispectral camera, and the collected scenes include vegetation, water, sky, buildings and other complex scenes to verify the advantages of the method proposed in the present invention. The processing of the multispectral image data and the training and testing of the convolutional neural network model for drone target detection are carried out in the Windows 10 environment, based on the pytorch open source deep framework, and accelerated using the NVDIARTX 3060 12G graphics card.
[0060] This embodiment provides a method for detecting unmanned aerial vehicle targets based on multispectral information fusion, which specifically includes the following steps:
[0061] S1. Collect multispectral images to build a dataset
[0062] Obtain a multispectral image containing a drone target acquired using a RedEdge five-channel multispectral camera. The spectral channels of the image include at least red, blue, green, and near-infrared spectral bands. Then, perform preprocessing such as channel registration (alignment between spectral bands is achieved through image registration algorithm) on the acquired multispectral image, and annotate the drone target in the image to obtain a sample set of drone images with labels, thus completing the construction of the dataset.
[0063] The multispectral image registration method adopted in this embodiment is SIFT+RANSAC, and the data of the registered image is annotated by the Labelimg annotation tool.
[0064] S2. Build a UAV target detection network model based on multi-spectral information fusion
[0065] The UAV target detection network model is a convolutional neural network model with a dual-stream network structure. After extracting features from the original channel image and the spectral index fusion image of the input multispectral image, feature fusion is performed, and then target classification and target prediction are performed based on the feature fusion image.
[0066] The network structure of the UAV target detection network model is as follows Figure 2 As shown. This embodiment is modified on the basis of the YOLOv5 network structure, and the structure of the YOLOv5l includes a feature extraction backbone network (CSPDarkNet53), a Neck connection layer (FPN+PAN) and a Precision detection head. This embodiment uses two independent feature extraction backbone networks in the backbone part, wherein the input of the first feature extraction backbone network is a multispectral image composed of spectral images of the three original channels of red, blue and green, and the input of the second feature extraction backbone network is a spectral index fusion image. Afterwards, the two feature extraction backbone networks each send their outputs to the Neck neck connection layer, in which the features extracted by the two feature extraction backbone networks are spliced and fused, and finally the Precision detection head outputs the detection results of the drone target.
[0067] The process of the UAV target detection network model processing the input image data includes:
[0068] S21, feature extraction stage: using two independent and identical feature extraction backbone networks to extract features from the original channel image and the spectral index fusion image respectively;
[0069] S22, feature fusion stage: based on the feature fusion module, the features extracted by the two feature extraction backbone networks are fused, and the fusion features are extracted, and the fusion feature map is output;
[0070] S23, target detection stage: perform target detection on the fused feature map output by step S22, and output the detected target information.
[0071] In this embodiment, the spectral index fusion image is a fusion image of the Normalized Difference Vegetation Index (NDVI) and the Normalized Difference Water Index (NDWI), and the generation method thereof is as follows:
[0072] 1) For the registered multispectral image, the normalized vegetation index and water index of each pixel are calculated to obtain the normalized vegetation index map and the normalized water index map. The calculation formula is:
[0073]
[0074]
[0075] Among them, NIR, Red, and Green respectively represent the reflectance of the near-infrared spectral band, red spectral band, and green spectral band of the pixel point; the normalized vegetation index map refers to an image in which the pixel value of each pixel point is represented by the NDVI value, and the normalized water index map refers to an image in which the pixel value of each pixel point is represented by the NDWI value.
[0076] 2) For the normalized vegetation index map obtained in step 1), a segmentation threshold d is set to obtain all pixel points in the normalized vegetation index map whose pixel values are not less than the segmentation threshold, so as to extract the vegetation area in the image and generate a mask map, which is expressed as follows:
[0077]
[0078] Among them, p NDVI Represents the pixel value (NDVI value) of pixel point p in the normalized vegetation index map, I Mask represents the pixel value of pixel p on the mask image; the above formula indicates that if the pixel value of pixel p in the normalized vegetation index image is not less than the preset segmentation threshold, the pixel value of pixel p in the mask image I Mask Set to 1 (all white), otherwise I Mask Set to 0 (completely black);
[0079] In the index value range of [-1, 1], generally, the normalized vegetation index pixel value of the vegetation area is ≥ 0, so the value of the segmentation threshold d should not be less than 0. In this embodiment, the segmentation threshold d is set to 0.2.
[0080] 3) Using the mask image obtained in step 2), generate a spectral index fusion image. The specific operation method is as follows:
[0081] For the pixel points with a pixel value of 1 in the mask map, fill in the pixel value (NDVI value) of the pixel point in the normalized vegetation index map; for the pixel points with a pixel value of 0 in the mask map, fill in the pixel value (NDWI value) of the pixel point in the normalized water index map. The process can be expressed as:
[0082] I Fusion =I NDVI *I Mask +I NDWI *(1-I Mask )
[0083] Among them, I Fusion Represents the pixel value of a pixel in the spectral index fusion image, I NDVI ,I NDWI ,I Mask They are the pixel values at the same pixel position on the normalized vegetation index map, normalized water index map, and mask map, respectively, and * is the multiplication symbol.
[0084] S3. Training of UAV target detection network model based on multi-spectral information fusion
[0085] Use the data set constructed in step S1 to train the convolutional neural network model built in step S2. During the training process, use the loss function to measure the distance between the output and the data label, update the network model weights to reduce the loss function value, and finally save the network model weights with the smallest loss function value.
[0086] S4. Implement UAV target detection based on multi-spectral information fusion
[0087] Channel registration is performed on the multispectral image to be detected collected by the multispectral camera, and the corresponding spectral index fusion image is calculated. The red, blue and green original channel images of the registered multispectral image and the spectral index fusion image calculated according to the above method are input into the UAV target detection network model trained in step S3 to obtain the target detection information output by the model.
[0088] In this example, we conducted a comparative experiment to demonstrate the advancedness of the method of the present invention. Figure 3 As shown in the figure: in the pseudo-color image composed of the red, green and blue three-channel spectral images in the multispectral image, the UAV target is difficult to distinguish with the naked eye; in the normalized difference vegetation index (NDVI) image, when the background of the UAV is vegetation, the vegetation area is highlighted and the UAV target is highlighted, and when the background of the UAV is water, the UAV target is less obvious; in the normalized water index (NDWI) image, when the background of the UAV is water, the UAV target is relatively obvious, and when the background is vegetation, the highlighting effect of the UAV in the normalized vegetation index (NDVI) image is less obvious; when the method of the present invention is used for UAV target detection, in the spectral index fusion image, no matter the background is vegetation or water, the UAV target can be effectively highlighted, and the method of the present invention has the advantages of both normalized vegetation index (NDVI) and normalized water index (NDWI) imaging.
[0089] Figure 4 These are the detection results obtained in some experiments, involving a variety of environments, which can prove that the method of the present invention can achieve stable and high-precision UAV target detection.
[0090] At the same time, although the embodiment of the present invention is modified on the basis of the YOLOv5 network structure, the structural framework of the dual feature extraction backbone network constructed by the present invention can also be used to transform most of the current network models that use a single-stage target detection algorithm, so as to achieve higher-precision drone target detection.
[0091] Embodiment 2:
[0092] This embodiment provides a detection system for implementing the UAV target detection method using multispectral information fusion as described in Embodiment 1, which specifically includes the following components:
[0093] The data set construction module is used to store the multispectral images containing the UAV targets collected by the multispectral camera, and to perform channel registration and target labeling on the collected multispectral images to obtain a set of sample images with labels, thereby forming a data set;
[0094] UAV target detection network model building module, used to build a convolutional neural network model for detecting UAV targets;
[0095] A UAV target detection network model training module is used to train the convolutional neural network model using the multispectral image of the data set to obtain a trained UAV target detection network model;
[0096] The target prediction module is used to detect the multispectral image to be processed using the trained UAV target detection network model and output the detected target information.
[0097] In the detection system, the convolutional neural network model includes a backbone part, a neck connection layer and a detection head, wherein:
[0098] The backbone part is composed of two parallel feature extraction backbone networks, wherein the input of the first feature extraction backbone network is an image containing three original channels of red, blue and green, and the input of the second feature extraction backbone network is a spectral index fusion image fused with a normalized vegetation index and a normalized water index;
[0099] The neck connection layer fuses the features extracted by the two feature extraction backbone networks based on the feature fusion module, performs further feature extraction on the fused features, and outputs a fused feature map;
[0100] The detection head performs target detection based on the fused feature map and outputs the detected target information.
[0101] Embodiment 3:
[0102] A computer system, comprising a processor and a memory, characterized in that:
[0103] The memory is used to store computer programs and instructions;
[0104] When the processor runs the computer program according to the instructions, the UAV target detection method based on multispectral information fusion as described in Example 1 is implemented.
[0105] Embodiment 4:
[0106] A computer-readable storage medium having a computer program stored thereon, characterized in that:
[0107] When the computer program is executed by the processor, the UAV target detection method based on multispectral information fusion as described in Example 1 is implemented.
[0108] The above are only preferred implementations of the present invention. The protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that those skilled in the art can realize that the algorithm steps described in Example 1 can be implemented by electronic hardware or a combination of computer software and electronic hardware. For those of ordinary skill in the art, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.
Claims
1. A method for detecting unmanned aerial vehicle targets based on multispectral information fusion, characterized in that: The following steps are involved: S1. Collect multispectral images to build a dataset Acquire multispectral images containing UAV targets collected by a multispectral camera, perform channel registration and target annotation on the collected multispectral images, and construct a dataset for training and testing the UAV target detection network model; S2. Build a UAV target detection network model based on multi-spectral information fusion The UAV target detection network model is a convolutional neural network model with a dual-stream network structure, which extracts features from the original channel image and the spectral index fusion image of the multispectral image respectively, and then performs feature fusion processing; S3. Training of UAV target detection network model based on multi-spectral information fusion Using the data set constructed in step S1 to train the drone target detection network model constructed in step S2, to obtain a trained drone target detection network model; S4. Implement UAV target detection based on multi-spectral information fusion Perform channel registration on the multispectral image to be detected collected by the multispectral camera, and calculate the corresponding spectral index fusion image, input the original channel image of the registered multispectral image and the calculated spectral index fusion image into the trained UAV target detection network model in step S3, and obtain the target information output by the model; The method for generating the spectral index fusion image comprises the following steps: 1) For the registered multispectral image, the normalized vegetation index and normalized water index of each pixel are calculated to obtain the normalized vegetation index map and the normalized water index map. The calculation formula is: Wherein, NIR, Red, and Green represent the reflectance of the near infrared spectral band, the red spectral band, and the green spectral band of the pixel point, respectively; the normalized vegetation index map refers to an image in which the pixel value is represented by the NDVI value, and the normalized water index map refers to an image in which the pixel value is represented by the NDWI value; 2) For the normalized vegetation index map obtained in step 1), a segmentation threshold d is set to obtain pixel points whose pixel values in the normalized vegetation index map are not less than the segmentation threshold, so as to extract the vegetation area and generate a mask map, which is expressed as: Among them, p NDVI Represents the pixel value of any pixel point p in the normalized vegetation index map, I Mask represents the pixel value of pixel p on the mask image; the above formula indicates that if the pixel value of pixel p in the normalized vegetation index image is not less than the preset segmentation threshold, the pixel value of pixel p in the mask image I Mask Set to 1, otherwise set to 0; 3) Generate a spectral index fusion image using the mask image obtained in step 2) by: For the pixel points with a pixel value of 1 in the mask image, fill in the pixel value of the pixel point in the normalized vegetation index map; for the pixel points with a pixel value of 0 in the mask image, fill in the pixel value of the pixel point in the normalized water index map. The expression formula is: I Fusion =I NDVI *I Mask +I NDWI *(1-I Mask ) Among them, I Fusion Represents the pixel value of a pixel in the spectral index fusion image, I NDVI ,I NDWI ,I Mask They are respectively the pixel values corresponding to the same pixel position on the normalized vegetation index map, the normalized water index map, and the mask map, and * is the multiplication symbol.
2. The method for detecting unmanned aerial vehicle targets based on multispectral information fusion according to claim 1, characterized in that: The multispectral image collected by the multispectral camera includes blue, green, red and near-infrared spectral bands.
3. The method for detecting unmanned aerial vehicle targets based on multispectral information fusion according to claim 1, characterized in that: The process of processing the input image data by the drone target detection network model built in step S2 includes: S21, feature extraction stage: using two independent and identical feature extraction backbone networks to extract features from the original channel image and the spectral index fusion image respectively; S22, feature fusion stage: based on the feature fusion module, the features extracted by the two feature extraction backbone networks are fused, and the fusion features are extracted, and the fusion feature map is output; S23, target detection stage: perform target detection on the fused feature map output by step S22, and output the detected target information.
4. The method for detecting unmanned aerial vehicle targets based on multispectral information fusion according to claim 1, characterized in that: In the step 2), the segmentation threshold d set for the normalized vegetation index map obtained in the step 1) has a value range of d≥0.
5. The method for detecting unmanned aerial vehicle targets based on multispectral information fusion according to claim 1, characterized in that: In the process of training the model in step S3, the loss function is used to measure the distance between the output and the data label, and the network model weight is updated to reduce the loss function value, and finally the network model weight with the smallest loss function value is saved.
6. A detection system for implementing the unmanned aerial vehicle target detection method based on multispectral information fusion as described in any one of claims 1 to 5, characterized in that: include: A data set construction module is used to store multispectral images containing UAV targets collected by a multispectral camera, and to perform channel registration and target labeling on the collected multispectral images to obtain a data set consisting of sample images with labels; UAV target detection network model building module, used to build a convolutional neural network model for identifying UAV targets; A UAV target detection network model training module is used to train the convolutional neural network model using a data set to obtain a trained UAV target detection network model; The target prediction module is used to detect the multispectral image to be processed using the trained UAV target detection network model and output the detected target information.
7. The detection system according to claim 6, characterized in that: The convolutional neural network model includes a backbone part, a neck connection layer and a detection head; The backbone part is composed of two parallel feature extraction backbone networks, wherein the input of the first feature extraction backbone network is the original channel image of the multispectral image, and the input of the second feature extraction backbone network is the spectral index fusion image; The neck connection layer fuses the features extracted by the two feature extraction backbone networks based on the feature fusion module, extracts the fusion features, and outputs the fusion feature map; The detection head is used to perform target detection on the fused feature map and output the detected target information.
8. A computer system comprising a processor and a memory, characterized in that: The memory is used to store computer programs and instructions; When the processor runs the computer program according to the instructions, the UAV target detection method based on multispectral information fusion as described in any one of claims 1 to 5 is implemented.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for detecting unmanned aerial vehicle targets based on multispectral information fusion as described in any one of claims 1 to 5 is implemented.