Unmanned aerial vehicle detection method, device, equipment and medium

Abstract

The invention relates to the technical field of image recognition, and discloses an unmanned aerial vehicle detection method and device, equipment and a medium, and the method comprises the steps: obtaining a multi-scale feature image containing various types of unmanned aerial vehicles; by taking the multi-scale feature image as input, constructing a lightweight convolution model comprising a first detection head corresponding to conventional target identification and a second detection head corresponding to unmanned aerial vehicle identification; training the lightweight convolution model, and determining an identification model; and performing unmanned aerial vehicle detection based on the identification model. According to the scheme, the multi-scale feature image is used as the training data, so that the convolution model can adapt to different environments and backgrounds, the robustness of the convolution model is improved, and the convolution model effectively executes multi-scale detection on the premise of not sacrificing the response speed, thereby realizing consideration of the detection precision and the response speed.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, specifically to a method, apparatus, equipment, and medium for detecting unmanned aerial vehicles (UAVs). Background Technology

[0002] Drones are increasingly prevalent in military, transportation, and logistics fields. With advancements in drone technology, their designs are becoming smaller. This makes detecting small or long-range drones in image processing more challenging.

[0003] In related technologies, image detection mostly utilizes two-stage object detection models and single-stage object detection. Two-stage object detection models, such as Faster R-CNN, while performing well in object detection accuracy, typically have slow response times, making them unsuitable for applications requiring real-time response, such as drone surveillance. Single-stage object detection models, such as the YOLO (You Only Look Once) framework, can guarantee fast response times, but their accuracy cannot be guaranteed when detecting small or long-range drones. Summary of the Invention

[0004] In view of this, the present invention provides a method, apparatus, device, and medium for detecting unmanned aerial vehicles (UAVs) to solve the technical problem in the related art that it is difficult to balance detection accuracy and response speed when detecting conventional targets and UAVs.

[0005] In a first aspect, the present invention provides a method for detecting unmanned aerial vehicles (UAVs), the method comprising: acquiring multi-scale feature images containing multiple types of UAVs; using the multi-scale feature images as input, constructing a lightweight convolutional model containing a first detection head corresponding to conventional target recognition and a second detection head corresponding to UAV recognition; training the lightweight convolutional model using joint training or independent training as training strategies based on training efficiency requirements or model accuracy requirements, and determining a recognition model; and performing image detection and recognition based on the image to be tested and the recognition model.

[0006] In conjunction with the first aspect, in one possible implementation of the first aspect, the process of training the lightweight convolutional model includes: simultaneously using the image with the highest resolution among the multi-scale feature images as the input to the second detection head, and using images of other scales among the multi-scale feature images as the input to the first detection head; constructing the loss function of the lightweight convolutional model; and optimizing the lightweight convolutional model based on the loss function.

[0007] In conjunction with the first aspect, in one possible implementation of the first aspect, the loss function of the lightweight convolutional model is expressed by the following formula: L_total=α* L_normal+β* L_small Where L_total represents the loss function, L_normal represents the detection task corresponding to the normal target, L_small represents the detection task corresponding to the UAV recognition, and α and β represent the weight coefficients of the corresponding tasks.

[0008] In conjunction with the first aspect, in one possible implementation of the first aspect, the process of training the lightweight convolutional model includes: freezing the second detection head, using images of other scales in the multi-scale feature images as input to the first detection head, and optimizing the backbone network of the lightweight convolutional model and the first detection head; freezing the backbone network and the first detection head, using the image with the highest resolution in the multi-scale feature images as input to the second detection head, and optimizing the second detection head.

[0009] In conjunction with the first aspect, in one possible implementation of the first aspect, acquiring multi-scale feature images containing multiple types of UAVs includes: collecting historical images containing multiple types of UAVs; and determining multi-scale feature images based on the historical images.

[0010] In conjunction with the first aspect, in one possible implementation of the first aspect, determining a multi-scale feature image based on historical images includes: determining a stitched image based on historical images; and extracting features from the stitched image based on a deep convolutional network to form a multi-scale feature image.

[0011] In conjunction with the first aspect, in one possible implementation of the first aspect, determining the stitched image based on historical images includes: stitching multiple historical images together to form a corresponding stitched image.

[0012] Secondly, the present invention provides a drone detection device, comprising: an acquisition module for acquiring multi-scale feature images containing multiple types of drones; a construction module for constructing a lightweight convolutional model containing a first detection head corresponding to conventional target recognition and a second detection head corresponding to drone recognition, using the multi-scale feature images as input; a training module for training the lightweight convolutional model to determine a recognition model; and a recognition module for performing drone detection based on the model to be recognized.

[0013] Thirdly, the present invention provides a computer device, including: a memory and a processor, which are communicatively connected to each other. The memory stores computer instructions, and the processor executes the computer instructions to perform the unmanned aerial vehicle detection method of the first aspect or any corresponding embodiment described above.

[0014] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to execute the unmanned aerial vehicle detection method of the first aspect or any corresponding embodiment described above.

[0015] The technical solution of this invention has the following advantages: This invention provides a method, apparatus, device, and medium for detecting unmanned aerial vehicles (UAVs). The method includes: acquiring multi-scale feature images containing various types of UAVs; constructing a lightweight convolutional model containing a first detection head corresponding to conventional target recognition and a second detection head corresponding to UAV recognition; training the lightweight convolutional model based on training efficiency requirements or model accuracy requirements to determine a recognition model; and performing UAV detection based on the recognition model. In this process, by constructing a convolutional model containing a first detection head corresponding to conventional target recognition and a second detection head corresponding to UAV recognition, the detection capability of the convolutional model for UAVs is improved. Furthermore, by using multi-scale feature images as training data, the convolutional model can adapt to different environments and background changes. While maintaining the conventional target recognition capability, it enhances the detection capability of the convolutional model for UAVs in complex environments, improving the robustness of the convolutional model. Moreover, because the convolutional model used is lightweight, it effectively performs multi-scale detection without sacrificing response speed, thus achieving a balance between detection accuracy and response speed. Attached Figure Description

[0016] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating a drone detection method according to an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the use of joint training as a training strategy in the lightweight convolutional model provided in an embodiment of the present invention. Figure 3 This is a schematic diagram illustrating the use of independent training as a training strategy for the lightweight convolutional model provided in an embodiment of the present invention. Figure 4 This is a schematic diagram illustrating the formation of a multi-scale feature image according to an embodiment of the present invention; Figure 5 This is a schematic diagram of a stitched image provided according to an embodiment of the present invention; Figure 6This is a structural block diagram of a drone detection device according to an embodiment of the present invention; Figure 7 This is a schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] According to an embodiment of the present invention, an embodiment of a drone detection method is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0020] This embodiment provides a method for detecting unmanned aerial vehicles (UAVs), such as... Figure 1 As shown, the method includes the following steps: S101. Acquire multi-scale feature images containing various types of drones.

[0021] Specifically, multi-scale feature images are feature maps of multiple scales obtained by a deep convolutional network extracting features from an input image, such as... Figure 4 As shown, four feature images at different scales are illustrated. The deep convolutional network can be YOLO or other deep convolutional networks, and this embodiment does not specifically limit it. The input image refers to historical images obtained through processing of various types of drones, making the input image applicable to targets of varying sizes, thus providing a data foundation for ensuring that the convolutional model can adapt to different environments and background changes.

[0022] S102. Using multi-scale feature images as input, construct a lightweight convolutional model that includes a first detection head corresponding to conventional target recognition and a second detection head corresponding to UAV recognition.

[0023] Specifically, the detection head included in the lightweight convolutional model can be an image acquisition device such as a camera. The first detection head is used to identify regular targets, such as regular-sized drones and other objects in the image, while the second detection head is used to identify small drones and long-range drones.

[0024] Specifically, since the multi-scale feature images introduced cover different environmental conditions, such as lighting, weather, and different backgrounds, while improving the model's accuracy, they also bring more parameters and computational load. Therefore, in order to ensure the model's response speed, a lightweight convolutional model is selected to reduce the model's complexity.

[0025] The standard convolutional model employs a computationally intensive approach. Assuming the input feature dimension is c × h × w, where c represents the number of input channels, and h and w represent the height and width of the feature respectively, the desired output channel is n. The standard convolutional model uses n sets of convolutional kernels, each containing c k × k two-dimensional kernels. During computation, each set of kernels performs a comprehensive convolution operation with all channels of the input feature, then sums the results of the c channels to generate a single set of output features. This process is repeated n times to obtain n sets of output features. This can be represented as Y = X * f + b, where Y represents the output feature, X represents the input feature, f represents the convolutional kernel, and b represents the bias term.

[0026] In contrast, lightweight convolutional models decompose the feature generation process into two steps. **Generate Intrinsic Feature Maps:** First, a small-scale ordinary convolution, typically a 1×1 convolution, is used to generate some "intrinsic" or "core" features. Assuming we still need n output channels and set a hyperparameter s, usually 2, this step will generate m = n / s intrinsic features. The calculation formula is Y` = X * f`, where the number of output channels in f` is much smaller than f in an ordinary convolution. **Generate "Inexpensive" / Lightweight Feature Maps:** Next, for each of the m intrinsic features Y` generated in the first step, a series of computationally inexpensive linear transformations (CheapOperations) are performed to generate s-1 additional "inexpensive" / lightweight features.

[0027]

[0028] Concatenate: Finally, the m intrinsic features generated in the first step and the m×(s-1) inexpensive features generated in the second step are concatenated along the channel dimension to form the final n=m×s output feature maps. In this way, lightweight convolution cleverly avoids the expensive ordinary convolution calculation for all output feature maps, and instead achieves the same goal with a small number of core features and a large number of "inexpensive" transformations.

[0029] To more intuitively understand the advantages of lightweight convolution, we will analyze its theoretical compression ratio in terms of parameter size and computational cost (FLOPs). To simplify the analysis, we assume that the convolution kernel size is k×k, and the linear transformation in the second step also uses a convolution kernel of size d×d, where k=d.

[0030] The number of parameters in a regular convolution is:

[0031] The lightweight convolution parameter set consists of two parts: the parameters of the first step of ordinary convolution plus the parameters of the second step of depthwise convolution.

[0032]

[0033] The theoretical compression ratio is:

[0034] As can be seen from the formula, the number of parameters in lightweight convolution is only about 1 / s that of ordinary convolution.

[0035] The computational cost of a regular convolution is:

[0036] The computational cost of lightweight convolution is:

[0037] The theoretical compression ratio is:

[0038] Similarly, lightweight convolution requires only about 1 / s of the computation of ordinary convolution. This means that when s=2, the lightweight module can achieve similar performance to ordinary convolution with almost half the number of parameters and computation, thus ensuring the response speed of the model when multi-scale feature images are introduced.

[0039] S103. Train the lightweight convolutional model to determine the recognition model.

[0040] Specifically, in practical applications, image recognition and detection may face different requirements, such as needing higher training efficiency or higher accuracy of the trained model. Therefore, this technical solution provides two training strategies to address different practical application scenarios. One of these is joint training, such as... Figure 2 As shown, two detection heads are trained simultaneously, utilizing shared feature learning to optimize the entire network. Different loss weights are set for each detection head to ensure that the optimization of the smaller target detection head is not affected by the larger target detection head. Independent training is also performed, as shown below. Figure 3As shown, two detection heads are trained separately. First, the first detection head is trained for general target recognition, such as for recognizing normal-sized drones or other objects in the image. Then, the second detection head is trained specifically for drone recognition, such as for recognizing small drones and long-range drones. Independent training allows for more detailed adjustment of their respective parameters and optimization of their performance.

[0041] S104. Based on the recognition model, conduct drone detection.

[0042] Specifically, drone detection based on a recognition model refers to inputting the image to be detected into a trained recognition model and using the model's output prediction result as the drone detection result.

[0043] This invention provides a method, apparatus, device, and medium for detecting unmanned aerial vehicles (UAVs). The method includes: acquiring multi-scale feature images; constructing a lightweight convolutional model containing a first detection head corresponding to conventional target recognition and a second detection head corresponding to UAV recognition; training the lightweight convolutional model to determine a recognition model; and performing UAV detection based on the image to be tested and the recognition model. In this process, by constructing a convolutional model containing a first detection head corresponding to conventional target recognition and a second detection head corresponding to UAV recognition, the detection capability of the convolutional model for UAVs is improved. Furthermore, by using multi-scale feature images as training data, the convolutional model can adapt to different environments and background changes. While maintaining the conventional target recognition capability, it enhances the detection capability of the convolutional model for UAVs in complex environments, improving the robustness of the convolutional model. Moreover, because the convolutional model used is lightweight, it effectively performs multi-scale detection without sacrificing response speed, thus achieving a balance between detection accuracy and response speed.

[0044] In one alternative implementation, the process of training a lightweight convolutional model includes: Simultaneously, the image with the highest resolution among the multi-scale feature images is used as the input to the second detection head, and images of other scales among the multi-scale feature images are used as the input to the first detection head; a loss function for a lightweight convolutional model is constructed; and the lightweight convolutional model is optimized based on the loss function.

[0045] Specifically, such as Figure 2 As shown, joint training is a holistic, synchronous optimization method that treats the entire network, including a shared feature extraction backbone and two independent detection heads, as a single model for end-to-end training. Its core idea is to utilize the general features learned by the shared network to simultaneously serve the detection tasks of both normal-sized and micro-sized UAVs. During training, images are processed by the shared backbone network to extract multi-scale feature maps. The highest-resolution feature map is fed into a detection head specifically designed for micro-targets, while the rest are fed into a detection head for normal-sized targets.

[0046] Specifically, to address the issue that an excessive number of normal target samples in the dataset may lead to model optimization bias, this strategy introduces a weighted loss function and uses this loss function to optimize the lightweight convolutional model, thus achieving a joint training strategy.

[0047] In one alternative implementation, the loss function of the lightweight convolutional model is represented by the following formula: L_total=α* L_normal+β* L_small Where L_total represents the loss function, L_normal represents the detection task corresponding to the normal target, L_small represents the detection task corresponding to the UAV recognition, and α and β represent the weight coefficients of the corresponding tasks.

[0048] Specifically, by setting a higher weight β, the proportion of the small target detection task in the total loss can be artificially increased, ensuring that the network gives sufficient attention to this challenging task when updating parameters, and preventing its optimization process from being "overwhelmed" by the dominant normal target loss. The advantage of joint training is that it has high training efficiency and can promote the network to learn shared features with stronger generalization ability.

[0049] In one alternative implementation, the process of training a lightweight convolutional model includes: The second detection head is frozen, and images of other scales in the multi-scale feature images are used as input to the first detection head to optimize the backbone network of the lightweight convolutional model and the first detection head; the backbone network and the first detection head are frozen, and the image with the highest resolution in the multi-scale feature images is used as input to the second detection head to optimize the second detection head.

[0050] Specifically, such as Figure 3As shown, independent training employs a "divide and conquer" strategy, breaking down the training process into two independent stages for more targeted optimization. In the first stage, the system freezes the second detection head corresponding to UAV recognition and concentrates all computing resources to train the shared backbone network and the first detection head corresponding to regular target recognition. The goal of this stage is to enable the network to learn a robust and efficient set of general feature extraction capabilities, laying a solid foundation for subsequent tasks. After completing the first stage, the system enters the second stage. Here, the system reverses this process, freezing the already trained shared backbone and the first detection head corresponding to regular target recognition, and dedicating all computing resources to training the second detection head corresponding to UAV recognition. Because the backbone network provides high-quality features, this stage allows for fine-tuning of the detection capabilities for small-target or long-range UAVs. The advantage of independent training lies in its clear optimization objectives, with each stage operating independently, especially capable of pushing higher performance limits for challenging tasks involving small targets, and with a simpler parameter tuning process.

[0051] In one alternative implementation, acquiring multi-scale feature images containing multiple types of UAVs includes: Collect historical images containing various types of drones; based on the historical images, determine multi-scale feature images.

[0052] Specifically, collecting historical images of various types of drones refers to collecting a large number of images of drones of various types and sizes. These images should cover different environmental conditions, such as lighting, weather, and different backgrounds. At the same time, the collected images need to be accurately labeled, including the drone's outline and category.

[0053] In one alternative implementation, determining multi-scale feature images based on historical images includes: Based on historical images, a stitched image is determined; features are extracted from the stitched image using a deep convolutional network to form a multi-scale feature image.

[0054] Specifically, extracting features from stitched images using deep convolutional networks to form multi-scale feature images refers to using deep convolutional networks to extract features from stitched images and form multi-scale feature images, such as... Figure 4 As shown, four different scales of feature images are illustrated, applicable to various targets from large to small. This provides the data foundation for subsequently using the highest resolution image from the multi-scale feature images as input to the second detection head, and for using images of other scales from the multi-scale feature images as input to the first detection head. The deep convolutional network used in this example can be YOLO.

[0055] In one alternative implementation, determining the stitched image based on historical images includes: stitching multiple historical images together to form a corresponding stitched image.

[0056] Specifically, equipment used to detect drones typically possesses high-resolution visual information, while publicly available image data online usually has only low resolution. To enable the model to adapt well to high-resolution information, multiple historical images are stitched together to form a corresponding stitched image used to acquire high-resolution images, such as... Figure 5 As shown, this example illustrates the process of stitching together four historical images to obtain the corresponding stitched image.

[0057] This embodiment also provides a drone detection device for implementing the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0058] This embodiment provides a drone detection device, such as Figure 6 As shown, it includes: The acquisition module 201 is used to acquire multi-scale feature images containing various types of UAVs. For details, please refer to the description of step S101 in the above embodiments, which will not be repeated here.

[0059] The construction module 202 is used to construct a lightweight convolutional model that includes a first detection head corresponding to conventional target recognition and a second detection head corresponding to UAV recognition, taking multi-scale feature images as input. For details, please refer to the relevant description of step S102 in the above embodiments, which will not be repeated here.

[0060] Training module 203 is used to train the lightweight convolutional model and determine the recognition model. For details, please refer to the description of step S103 in the above embodiments, which will not be repeated here.

[0061] The identification module 204 is used to detect drones based on the identification model. For details, please refer to the description of step S104 in the above embodiments, which will not be repeated here.

[0062] In this embodiment, the UAV detection device is presented in the form of a functional unit. Here, a unit refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0063] This invention also provides a computer device having the above-described features. Figure 6The drone detection device shown.

[0064] Please see Figure 7 , Figure 7 This is a schematic diagram of the structure of a computer device provided in an optional embodiment of the present invention, such as... Figure 7 As shown, the computer device includes one or more processors 301, memory 302, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 7 Take processor 301 as an example.

[0065] Processor 301 may be a central processing unit, a network processor, or a combination thereof. Processor 301 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0066] The memory 302 stores instructions executable by at least one processor 301 to cause the at least one processor 301 to perform the method shown in the above embodiments.

[0067] Memory 302 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device. Furthermore, memory 302 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, memory 302 may optionally include memory remotely located relative to processor 301, and this remote memory may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0068] The memory 302 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 302 may also include combinations of the above types of memory. The computer device also includes a communication interface 303 for communicating with other devices or communication networks.

[0069] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods shown in the above embodiments.

[0070] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A method for detecting unmanned aerial vehicles (UAVs), characterized in that, The method includes: Acquire multi-scale feature images containing various types of drones; Using the multi-scale feature image as input, a lightweight convolutional model is constructed that includes a first detection head corresponding to conventional target recognition and a second detection head corresponding to UAV recognition; The lightweight convolutional model is trained to determine the recognition model; Based on the aforementioned recognition model, drone detection is performed.

2. The method according to claim 1, characterized in that, The process of training the lightweight convolutional model includes: Simultaneously, the image with the highest resolution among the multi-scale feature images is used as the input of the second detection head, and images of other scales among the multi-scale feature images are used as the input of the first detection head; Construct the loss function for the lightweight convolutional model; Based on the loss function, the lightweight convolutional model is optimized.

3. The method according to claim 2, characterized in that, The loss function of a lightweight convolutional model can be expressed by the following formula: L_total=α* L_normal+β* L_small Where L_total represents the loss function, L_normal represents the detection task corresponding to the normal target, L_small represents the detection task corresponding to the UAV recognition, and α and β represent the weight coefficients of the corresponding tasks.

4. The method according to claim 1, characterized in that, The process of training the lightweight convolutional model includes: Freeze the second detection head, and use images of other scales in the multi-scale feature image as input to the first detection head to complete the optimization of the backbone network of the lightweight convolutional model and the first detection head; The backbone network and the first detection head are frozen, and the image with the highest resolution among the multi-scale feature images is used as the input of the second detection head to complete the optimization of the second detection head.

5. The method according to claim 1, characterized in that, The acquisition of multi-scale feature images containing multiple types of UAVs includes: Collect historical images containing various types of drones; Based on the historical images, multi-scale feature images are determined.

6. The method according to claim 5, characterized in that, The step of determining multi-scale feature images based on the historical images includes: Based on the historical images, determine the stitched image; Feature extraction is performed on the stitched image using a deep convolutional network to form a multi-scale feature image.

7. The method according to claim 6, characterized in that, The step of determining the stitched image based on the historical image includes: Multiple historical images are stitched together to form a corresponding stitched image.

8. A drone detection device, characterized in that, The device includes: The acquisition module is used to acquire multi-scale feature images containing various types of drones; The construction module is used to construct a lightweight convolutional model that includes a first detection head corresponding to conventional target recognition and a second detection head corresponding to UAV recognition, using the multi-scale feature image as input. The training module is used to train the lightweight convolutional model and determine the recognition model; The identification module is used to detect drones based on the identification model.

9. A computer device, characterized in that, include: The system includes a memory and a processor, which are interconnected and the memory stores computer instructions. The processor executes the computer instructions to perform the UAV detection method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to perform the unmanned aerial vehicle (UAV) detection method according to any one of claims 1 to 7.

Images