Low-power-consumption multi-unmanned aerial vehicle small target detection method based on spiking neural network

Abstract

The invention discloses a low-power-consumption multi-unmanned aerial vehicle small target detection method based on a spiking neural network, and the method comprises the steps: firstly distributing a data collection task to an edge-end unmanned aerial vehicle, and enabling the edge-end unmanned aerial vehicle to carry out the data collection through the crowdsourcing perception technology; secondly, the central server distributes the initial weight of the global target detection model to the edge-end unmanned aerial vehicle, and the edge-end unmanned aerial vehicle locally obtains a local target detection model through training of collected data; and then the edge-end unmanned aerial vehicle transmits the trained local target detection model gradient back to the central server, and the central server updates the target detection model by using an aggregation algorithm. And finally, the updated target detection model of the central server is used for target detection to obtain a target detection result. According to the method, the problems of high power consumption, insufficient pulse distribution rate, poor noise environment robustness, privacy risk caused by centralized data acquisition and the like of an existing unmanned aerial vehicle small target detection method are solved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of artificial intelligence, computer vision and unmanned aerial vehicle target detection, and particularly relates to a low-power multi-unmanned aerial vehicle small target detection method based on a spiking neural network. BACKGROUND

[0002] In recent years, unmanned aerial vehicles have been widely used in many fields, especially in disaster relief, environmental monitoring and agricultural applications. Unmanned aerial vehicles have strong mobility and deployment flexibility, making them ideal tools for monitoring, data collection and target identification. By carrying high-definition cameras or sensors, unmanned aerial vehicles can obtain real-time image data of the ground or target at high altitude and perform real-time processing and feedback. The EMM-Det method can accurately detect targets in a noisy environment, and also provides a data collection method that can protect data privacy while collecting enough target detection model training data.

[0003] However, the prior art has the following significant deficiencies: insufficient pulse firing rate, spiking neural networks require neurons to accumulate membrane potential before firing, which makes it difficult for neurons to maintain a high firing rate in a short period of time, thereby limiting their feature extraction capabilities; insufficient anti-noise interference capability: environmental factors such as light and weather changes can significantly reduce the quality of unmanned aerial vehicle images, introducing noise that existing processing networks cannot manage; large data collection workload: target detection networks require a large amount of labeled data to achieve high accuracy. However, centralized collection of unmanned aerial vehicle training data is labor-intensive and resource-intensive. In sensitive areas such as disaster assessment and rescue, data aggregation can also pose a serious risk of privacy leakage. SUMMARY

[0004] To make up for the deficiencies of the prior art, the present application provides a low-power multi-unmanned aerial vehicle small target detection method based on a spiking neural network to address the many deficiencies of traditional methods in the field of invention.

[0005] The low-power multi-unmanned aerial vehicle small target detection method based on a spiking neural network is mainly divided into four stages, and the specific process is as follows: S1. Distribute the data collection task to the edge-end unmanned aerial vehicle, and the edge-end unmanned aerial vehicle uses crowd-sourced perception technology to collect data.

[0006] S2. The center server distributes the initial weights of the global target detection model to the edge-end unmanned aerial vehicle, and the edge-end unmanned aerial vehicle locally trains a local target detection model using the collected data.

[0007] S3. The edge-end unmanned aerial vehicle returns the gradient of the trained local target detection model to the center server, and the center server updates the target detection model using an aggregation algorithm.

[0008] S4. The updated target detection model of the center server is used for target detection, obtaining a target detection result.

[0009] Further, the global target detection method used in step S2 is as follows: Step S2.1. An impulse neural network is used to perform feature representation through sparse binary.

[0010] The impulse neuron is modeled as follows: Wherein represents the time The total current flowing through the neuron at time t, wherein represents the current flowing through the resistor, represents the current flowing through the capacitor. The current flowing through the capacitor can be expressed as wherein represents the capacitance, is a small time, is the voltage change in time is the rate of change of the voltage across the capacitor with respect to time. The above description can be solved to represent as: wherein, is the membrane potential at time is the initial membrane potential, is the membrane resistance, is the capacitance value. Set as the leakage time constant, control the leakage strength of the membrane potential, then the above formula can be expressed as: By setting to to reduce the computational complexity. Considering the difficulty of simulating continuous changes in neuron membrane potential using computing devices such as GPUs, the equation can be discretized as follows: wherein is the simulation time step, the above formula can be iteratively solved, wherein the th step is: Step S2.2. An improved impulse neural network is proposed by using a memory-enhanced dynamic leakage constant model Me-LiF to obtain corresponding charging functions and recovery functions.

[0011] ​A dynamic leakage constant is designed to reduce the time cost during inference. The spiking neuron of this invention can be represented as follows: definition This is an adjustable parameter of the capacitor. When When the capacitance regulation increases the leakage of the membrane potential, it suppresses pulsed discharge; when At this time, capacitive regulation reduces membrane potential leakage and promotes pulsed discharge. When At that time, the system behaves as a standard neuron model.

[0012] The charging and firing processes of neurons are approximated by simulating time steps. By discretizing these time steps, the dynamic behavior of neurons can be simulated by progressively updating the neuron's membrane potential. Let... This allows us to obtain the mathematical formula describing the dynamics of neurons: in, It represents Membrane potential at time step, time constant The membrane controls the potential rate of decay. This represents the weight of neuron i. It is the first The binary input value of a presynaptic neuron. This represents the membrane potential leakage behavior, which causes the membrane potential to decay towards zero in the absence of input. It is the weighted sum of all incoming spike inputs.

[0013] There are two ways to reset the membrane potential after a neuron fires. A hard reset directly sets the membrane potential to the resting potential. In contrast, a soft reset, as shown in the equation below, subtracts a threshold from the membrane potential, thus preserving the network's memory: Considering the need to further preserve information at each time step in the network, this invention proposes an exponential function reset (exp-reset) scheme, as follows: In summary, the memory-enhancing spiking neuron of this invention can be mathematically modeled as follows: .

[0014] Step S2.3. Construct an object detection network using the improved spiking neural network, integrate ME-LiF into the feature extraction module, and convert each convolution into a fully spiked convolutional layer. The object detection network is finally constructed using a quintuple. a set of predictions is outputted in the form of and denote the center coordinates of the prediction bounding box, and denote its width and height, respectively, is the confidence score.

[0015] Further, when the target detection method used in the step S2 faces the environment with poor visibility such as rain and snow, the visual information collected by the unmanned aerial vehicle will contain noise, which affects the final detection accuracy, the application selects to use Haar wavelet transform for image processing, and the Haar wavelet can be combined with convolution operation, and the Haar wavelet can effectively extract low-frequency and high-frequency information from each unmanned aerial vehicle image.

[0016] The application has the following beneficial effects: The application introduces a low-power target detection architecture based on a pulse neural network, combines a memory-enhanced dynamic leakage constant pulse neuron (Me-LiF), a Haar wavelet transform down-sampling module, and a federated learning and differential privacy mechanism, solves the problems of high power consumption, insufficient pulse firing rate, poor noise environment robustness, and privacy risks brought by centralized data collection in the existing unmanned aerial vehicle small target detection method, and brings the comprehensive effects of significantly improving detection accuracy and reasoning efficiency under low-power conditions, enhancing anti-noise ability in complex environments, reducing model training and deployment energy consumption, and realizing multi-unmanned aerial vehicle collaborative training and privacy protection. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 It is a schematic diagram of the framework of the application. DETAILED DESCRIPTION

[0018] As Figure 1 shown, the low-power multi-unmanned aerial vehicle small target detection method based on a pulse neural network mainly includes four stages, and the specific process is as follows: S1. Distribute the data collection task to the edge unmanned aerial vehicle, and the edge unmanned aerial vehicle uses crowd-sensing technology to collect data.

[0019] S2. The center server distributes the initial weight of the global target detection model to the edge unmanned aerial vehicle, and the edge unmanned aerial vehicle locally trains the local target detection model through the collected data.

[0020] S3. The edge unmanned aerial vehicle returns the gradient of the trained local target detection model to the center server, and the center server updates the target detection model using an aggregation algorithm.

[0021] S4. The updated target detection model of the center server is used for target detection, and the target detection result is obtained.

[0022] Furthermore, the crowdsourced data collection method adopted in step S1 is as follows: To improve the generalization ability of the target detection model under multiple scenarios and device conditions, this method uses a crowdsourced perception approach for dataset collection. First, a unified data collection platform is designed, opening collection tasks to the general public, volunteers, and users in specific industries. Participants can use local drone equipment to collect image or video data in natural environments. Subsequently, the collected image data is bounded to targets and labeled with categories to ensure the accuracy and consistency of the labeling. Finally, a task incentive mechanism (such as points, ranking, or rewards) is used to enhance participation enthusiasm and achieve continuous data collection and diversity expansion. This method effectively integrates distributed perception resources, reduces the cost of constructing large-scale target detection datasets, and enhances the scene diversity and realism of the data.

[0023] Furthermore, the global target detection method used in step S2 is as follows: In order to reduce the energy consumption of the target detection network and make it suitable for UAV systems, a spiking neural network is used in this invention. Compared with the large number of dense floating-point addition and multiplication operations in traditional neural networks, the spiking neural network uses sparse binary for feature representation, which significantly reduces the network power consumption.

[0024] Step S2.1. Employ a spiking neural network to represent features using sparse binary representation.

[0025] Spiking neurons can be modeled as follows: in Indicates time The total current flowing through the neuron, of which This represents the current flowing through the resistor. This represents the current flowing through the capacitor. It can be represented as ,in Indicates capacitance. It was a tiny moment. yes Voltage change over time, This is the rate of change of the voltage across the capacitor with time. The above description can be solved as follows: in, For a moment membrane potential, The initial membrane potential, For membrane resistance, Let be the capacitance value. Let be the leakage time constant, and control the membrane potential leakage intensity. Then the above formula can be expressed as: To optimize the performance of the equipment of the unmanned aerial vehicle, the calculation complexity can be reduced by setting to . Considering the difficulty of simulating the continuous change of the membrane potential of the neuron using a computing device such as a GPU, the equation can be discretized as follows: where is the simulation time step, and the above equation can be iteratively solved, where the first step is: Step S2.2. A memory-enhanced dynamic leak constant model Me-LiF improved pulse neural network is proposed, and the corresponding charging function and recovery function are obtained.

[0026] The existing pulse neuron model has the problem of insufficient memory and difficulty in firing in a short time step. When the membrane potential approaches the firing threshold without input current, the membrane potential decays. This leads to a low firing rate in a short time step, reducing the feature extraction capability of the network. In addition, increasing the number of time steps greatly increases the power consumption during network training, thereby increasing the training load of the unmanned aerial vehicle. To solve these problems, the present application proposes a memory-enhanced dynamic leak constant model (Me-LiF). Since the leak constant controls the decay of the membrane potential, the present application designs a dynamic leak constant to reduce the time cost during inference, thereby simultaneously reducing the energy consumption of the system. The pulse neuron of the present application can be represented as follows: where is defined as the adjustable parameter of the capacitance. When , the capacitance adjustment increases the leakage of the membrane potential, thereby suppressing the pulse firing; when , the capacitance adjustment reduces the leakage of the membrane potential, promoting the pulse firing. When , the system behaves as a standard neuron model.

[0027] In the field of computer vision, a pulse neural network usually approximates the charging and discharging process of a neuron by simulating time steps. By discretizing these time steps, the dynamic behavior of a neuron can be simulated by updating the membrane potential of the neuron step by step. In this framework, let , the mathematical formula describing the dynamics of the neuron can be obtained as follows: where represents the membrane potential at time step , the time constant is the rate at which the membrane controls the decay of the potential. represents the weight of neuron i, is the binary input value of the pre-synaptic neuron. represents the membrane potential leak behavior, which causes the membrane potential to decay towards zero without input, while is the weighted sum of all incoming spike inputs.

[0028] When the membrane potential is close to the threshold, the neuron is in a state of readiness to fire. The neuron holds important information that can be interpreted at this stage. At this time, the leak of the membrane potential should be reduced, thereby increasing the firing rate of the neuron within a short time step. Therefore, the network can respond faster, which is beneficial for the UAV to complete the object detection task under the condition of lower power consumption.

[0029] After the neuron fires, there are two cases that can reset the membrane potential. Hard reset directly sets the membrane potential to the resting potential. In contrast, the soft reset equation is as follows, subtracting the threshold from the membrane potential, which can preserve the memory of the network: Considering the need to further preserve the information of each time step in the network, the present application proposes an exponential function reset (exp-reset) scheme as follows: By taking advantage of the unique properties of the exponential function, the efficiency of information preservation in the neural network can be improved. When a neuron receives a significant stimulus that causes it to fire, the membrane potential will recover through a smooth and exponential recovery process. This method can maximize the preservation of information from the previous state, thereby achieving better memory preservation and accurate signal processing in subsequent steps. This setup greatly improves the network trigger rate, thus requiring a shorter inference time.

[0030] In summary, the memory-enhanced spiking neuron of the present application can be mathematically modeled as follows: Step 2.3. Construct a low-parameter object detection network using the spiking neuron of the present application. Integrate the ME-LiF into the feature extraction module, convert each convolution into a fully spiking convolution layer to reduce computational complexity. In this network architecture, in order to effectively reduce the computational burden of the edge device, all high-resolution images captured by the UAV are uniformly adjusted to 256x256 pixels. This preprocessing step significantly reduces the computational complexity and memory usage of the subsequent network stages, while retaining enough detail to ensure the ability to distinguish small targets. The object detection network finally outputs a set of predictions in the form of a five-tuple , where and denote the center coordinates of the predicted bounding box, and denote its width and height, respectively, is the confidence score. This compact output format is particularly suitable for UAV object detection tasks, achieving a good balance between detection accuracy and inference efficiency.

[0031] Further, the target detection method used in step S2 faces poor visibility environments such as rain and snow, and the visual information collected by the UAV will contain noise, affecting the final detection accuracy. Traditional image processing methods, such as convolutional neural network (CNN) based algorithms, try to obtain a larger receptive field by deploying larger convolution kernels. However, these methods are difficult to effectively handle noise during feature extraction. In high-noise videos captured by UAVs, simple convolution operations can lose or mix some representative features, affecting the accuracy of object detection. The reason for this problem is that traditional convolution-based methods are actually not sensitive to image details and noise, resulting in a decrease in detection accuracy. To this end, the present application selects to use Haar wavelet transform for image processing because of its simple structure and high processing efficiency, making it an ideal choice for processing large-scale UAV video datasets. Moreover, Haar wavelets can be combined with convolution operations, and Haar wavelets can effectively extract low-frequency and high-frequency information from each UAV image, thereby enhancing visual detail preservation and noise suppression. This integration is computationally efficient for UAV image feature extraction and target detection tasks, and preserves key image features in high-noise environments, improving the accuracy of final target detection. The present application uses four sets of Haar wavelet transform filters for convolution, with a convolution step size of 2, and the convolution kernel is as follows: where is a low-pass filter, , and form a set of high-pass filters. For each input channel, the output of the convolution layer is as follows: The output consists of four channels, each with a resolution of half the input image in each spatial dimension. denotes the low-frequency component of , while and correspond to its horizontal, vertical, and diagonal high-frequency components. Since the Haar wavelet filters form an orthonormal basis, the wavelet inverse transform can be implemented by transposing the convolution layer , i.e. The present application integrates Haar wavelet into the traditional down-sampling layer, thereby enhancing the representation ability of the model while expanding its receptive field. The present application designs a module named MT-downsampling, which consists of two parallel paths: one is the max-pooling path, as follows: wherein is the feature map of the previous layer, represents Batch Normalization is the batch normalization layer, represents Convolution is the convolution layer, represents Max Pooling is the max-pooling. The other is the wavelet transform path, as follows: wherein denotes Haar wavelet transform, is the inverse Haar wavelet transform. Then the feature maps obtained by the two paths are added element by element, as follows: Further, for steps S2 and S3, in order to support distributed training and protect privacy, the present application uses the FedAvg algorithm to implement a federated learning framework, wherein there are edge-end UAVs, each of which holds a local data set. The federated learning objective function is as follows: wherein, denotes the local target detection model, wherein is the model parameter, is the input and prediction result of the local target detection model.

[0032] In federated learning, the present application divides the data to UAVs, wherein denotes the data index set of UAV , denotes the number of data samples owned by UAV . The total number of data samples of all UAVs is .

[0033] Further, the present application converts the global optimization problem into a distributed form. Since the data is distributed on different clients, the global loss function is rewritten as the weighted average of the local loss function of each UAV. Specifically, the present application allocates weights according to the proportion of the data amount of each UAV, wherein the weight of UAV is reflecting the relative importance of the client's data in the total dataset. The joint objective can be redefined as: Further, the framework training process based on federated learning is as follows: the server first initializes the global model parameters and broadcasts them to all drones. Each drone performs local training on its private dataset to optimize the local objective function after receiving the global parameters. After completing the local training, the server collects the parameter updates of all drones and performs weighted aggregation using the FedAvg algorithm. Then the server broadcasts the updated global parameters to all drones and starts the next round of training.

[0034] Further, the training process is iterated until the objective function converges or meets the predefined convergence criteria, and the original data is kept locally on each client throughout the process, thus achieving privacy-preserving distributed collaborative training. To further enhance system security and privacy protection, a differential privacy mechanism is adopted during parameter transmission. Specifically, before sending the local parameter updates to the server, the drone adds calibrated noise to the local parameter updates: where is the model parameter of the kth drone after adding noise, is the original model parameter of the kth drone; is a multidimensional Gaussian random vector, subject to a normal distribution with mean 0 and variance . Where is the variance of the noise, used to control the size of the noise. is an identity matrix with the same dimension as . The noise variance is determined according to the differential privacy parameter , as follows: where is the sensitivity bound of the parameter update. Through this mechanism, even if an attacker obtains the parameter update information, it is impossible to accurately infer the private data of a specific client. is the privacy strength parameter in differential privacy, is the failure probability parameter in differential privacy. After receiving the noisy parameters, the server performs aggregation operations: ​ wherein, is the i-th trained global model parameter, is the result of aggregating all the uploaded from all the UAVs. is the i-th sample size of the i-th UAV, is the total sample size and is weighted aggregated according to the sample size of all the UAVs.

[0035] This differential privacy protection mechanism ensures that the difference in the distribution of algorithm output for any adjacent data set is strictly limited, thereby providing a quantifiable privacy protection level for each participant while maintaining the effectiveness of model training.

[0036] Further, the present application collects a multi-scene UAV target detection data set MultiDrone-Car. It is a newly constructed, multi-scene, UAV-based crowd-sensing vehicle detection data set developed specifically for high-speed inference in UAV applications. The data set contains 256x256 pixel images sampled from high-resolution aerial images taken by UAVs, covering 8 different scenes with diverse aerial view angles, including overhead, oblique, and side views. In addition, considering that UAV images are greatly affected by factors such as lighting conditions, different intensity of simulated noise is introduced in the data set. The data set contains 10,425 annotated images, a total of 47,765 vehicles, presenting a real distribution: small targets dominate, followed by medium targets.

[0037] To construct the MultiDrone-Car data set, the present application first collects a large number of high-resolution aerial photos from UAV monitoring tasks, covering various urban and rural environments. Initially, a team of three trained volunteers manually pre-annotates a small batch of images to establish ground truth labels. Based on these initial annotation data, a target detection network is trained to assist the annotation process of the remaining images. After the automatic annotation phase, volunteers manually verify all generated labels and find that about 28.75% of the automatic annotations have errors that need to be corrected. To ensure data set quality, 30% of the images are randomly selected for cross-validation, with an average Intersection over Union (IoU) of 0.923 between different annotators, confirming the reliability of the annotations.

[0038] In this experiment, the present application compares and analyzes the performance of multiple architectures (ANN and SNN models) under different power consumption constraints. Specifically, the present application evaluates the recall rate, mAP@50, and mAP@50:95 of models in three power ranges. The experimental results are shown in Table 1 below.

[0039] Table 1 Low-power architecture (less than 10 mJ)

[0040] Under the condition of low power consumption constraint, the method of the present application is compared with various traditional ANN models and existing SNN models. The experimental results show that the target detection model proposed in the present application achieves the optimal performance in the key indicators such as recall rate, mAP@50 and mAP@50:95. Especially in the mAP@50:95 indicator, the model of the present application is significantly better than the lightweight ANN models such as Tiny-YOLO and YOLOv8n, and also shows obvious advantages compared with the SpikeYOLO and EMS-YOLO pulse neural network methods, which shows that the present application can still maintain high small target detection precision and stability under very low power consumption condition.

[0041] Table 2 Medium power consumption architecture (10-30mJ)

[0042] As shown in Table 2, in the medium power consumption range, the overall detection performance of various models is improved, but the method of the present application still maintains a leading advantage. The experimental results show that the target detection model of the present application is superior to the comparison model in the recall rate and mAP indicators, which verifies that the memory-enhanced pulse neuron and wavelet down-sampling structure proposed have good robustness and generalization ability under different power consumption constraints. Compared with the traditional ANN model, the present application realizes higher detection accuracy under the premise of not significantly increasing power consumption, which reflects the effective balance between performance and energy consumption.

[0043] Table 3 High power consumption architecture (more than 30mJ)

[0044] As shown in Table 3, under the condition of high power consumption, although part of the large-scale ANN model has advantages in parameter quantity and computing power, the method of the present application can still maintain the same or even better performance in key detection indicators. The experimental results show that the target detection model of the present application is close to or reaches the optimal level in the recall rate and mAP@50 indicators, and is significantly better than part of the high-complexity model in the mAP@50:95 indicator, which shows that even in the case of relaxed power consumption limit, the present application still has strong detection ability and stability, which verifies the effectiveness of its network structure design.

[0045] In summary, by comparing the three sets of experimental results under different power consumption constraints, it can be seen that the low-power multi-UAV small target detection method based on pulse neural network proposed in the application has excellent comprehensive performance in low, medium and high power consumption scenarios. This method significantly reduces energy consumption while effectively improving small target detection accuracy and robustness, especially suitable for unmanned aerial vehicle application scenarios with limited computing power and energy consumption, fully verifying the advancement and practical value of the technical solution of the application.

[0046] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the application, but not to limit them; although the application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the application.

Claims

1. A low-power multi-UAV small target detection method based on spiking neural networks, characterized in that, Includes the following steps: S1. Distribute the data collection task to edge drones, which use crowdsourced perception technology to collect data. S2. The central server distributes the initial weights of the global object detection model to the edge drones, and the edge drones train their local object detection models using the collected data. S3. The edge drone sends the gradient of the trained local object detection model back to the central server, and the central server uses an aggregation algorithm to update the object detection model; S4. The updated target detection model on the central server is used for target detection to obtain target detection results.

2. The low-power multi-UAV small target detection method based on spiking neural network according to claim 1, characterized in that, The data collection using crowdsourced sensing technology in step S1 is as follows: First, by designing a unified data acquisition platform, each edge drone uploads image or video data in the natural environment; second, an automated data verification and labeling mechanism is set up; then, the acquired image data is bounded to targets and categorized; finally, a task incentive mechanism is used to enhance participation and achieve continuous data acquisition and diversification.

3. The low-power multi-UAV small target detection method based on spiking neural networks according to claim 1, characterized in that, The global object detection model in step S2 is implemented as follows: Step S2.

1. Employ a spiking neural network to represent features using sparse binary representation; Step S2.

2. Propose a memory-enhanced dynamic leakage constant model Me-LiF improved spiking neural network to obtain the corresponding charging function and recovery function; Step S2.

3. Construct an object detection network using the improved spiking neural network, integrate ME-LiF into the feature extraction module, and convert each convolution into a fully spiked convolutional layer. The object detection network is finally constructed using a quintuple. Output a set of predictions in the form of , where and Indicates the center coordinates of the predicted bounding box. and These represent its width and height, respectively. It is the confidence score.

4. The low-power multi-UAV small target detection method based on spiking neural network according to claim 3, characterized in that, The specific implementation process of step S2.1 is as follows: Using a spiking neural network, feature representation is performed through sparse binary representation. The spiking neuron is modeled as follows: in Indicates time The total current flowing through the neuron, of which This represents the current flowing through the resistor. This represents the current flowing through the capacitor; ,in Indicates capacitance. It was a tiny moment. yes The voltage change over time, as described above, can be solved as follows: in, For a moment membrane potential, The initial membrane potential, Let the membrane resistance be denoted as ; Let be the leakage time constant, and control the membrane potential leakage intensity. Then the above formula can be expressed as: By Set as To reduce computational complexity, the equations are discretized as follows: in To simulate the time step, the above equation is solved iteratively, where the th... The steps are as follows: 。 5. The low-power multi-UAV small target detection method based on spiking neural network according to claim 4, characterized in that, The specific implementation process of step S2.2 is as follows: The Me-LiF model for dynamic leakage constants in memory enhancement is proposed, and the spiking neuron is represented as follows: definition This is an adjustable parameter of the capacitor; when When the capacitance increases, the leakage of the membrane potential is reduced, suppressing pulse discharge; when When the capacitor adjusts, it reduces membrane potential leakage and promotes pulsed discharge; when At that time, it behaves as a standard neuron model; By discretizing the time step, the dynamic behavior of neurons is simulated by progressively updating the membrane potential of neurons. This yields the mathematical formula describing the dynamics of neurons: in, It represents Membrane potential at time step, time constant The membrane controls the potential rate of decay. Represents neurons The weight, It is the first The binary input values ​​of a presynaptic neuron; Indicates membrane potential leakage behavior, It is the weighted sum of all incoming spike inputs; After a neuron fires, there are two scenarios for resetting the membrane potential: a hard reset, which directly sets the membrane potential to the resting potential, and a soft reset, which subtracts a threshold from the membrane potential to preserve the network's memory. A scheme for resetting the exponential function is proposed as follows: In summary, the mathematical model of the memory-enhancing spiking neuron is as follows: 。 6. The low-power multi-UAV small target detection method based on spiking neural network according to claim 5, characterized in that, The specific process by which the edge-end drone trains a local target detection model using the collected data is as follows: Image processing is performed using Haar wavelet transform, employing four sets of Haar wavelet transform filters for convolution, with the following convolution kernels: in It is a low-pass filter. , and A set of high-pass filters is formed; for each input channel, the output of the convolutional layer is as follows: The output consists of four channels, each with a resolution equal to that of the input image in each spatial dimension. Half of; express The low-frequency components, and These correspond to its horizontal, vertical, and diagonal high-frequency components, respectively; the inverse wavelet transform is transposed through a convolutional layer. Implementation, that is: Integrating Haar wavelets into the downsampling layer consists of two parallel paths: one is a max pooling path, as shown below: in It is the feature map of the previous layer. Representative of the standardization layer, Represents a convolutional layer. Represents max pooling; Another is the wavelet transform path, as follows: in Represents the Haar wavelet transform. It is the inverse Haar wavelet transform; then the feature maps obtained from the two paths are added element by element to obtain the feature map. .

7. The low-power multi-UAV small target detection method based on spiking neural network according to claim 6, characterized in that, The central server uses an aggregation algorithm to update the target detection model as follows: A federated learning framework is implemented using the FedAvg algorithm, which includes... There are 16 edge drones, each holding a local dataset; the federated learning objective function is as follows: in, Describes the local object detection model, where For model parameters, This serves as the input and prediction result for the local object detection model. In federal learning, Indicates drone The data index set, Indicates drone The number of data samples available, the total number of data samples from all drones is ; The global optimization problem is transformed into a distributed form. The global loss function is rewritten as a weighted average of the local loss functions of each drone, with weights allocated according to the proportion of data volume for each drone. The weight is This reflects the relative importance of the client's data within the overall dataset; the joint objective is redefined as: In each round of communication, the drone Add calibration noise to the local parameter update before sending it to the server: in These are the model parameters for the k-th UAV after adding noise. It is the first The original model parameters of the drone; It is a multidimensional Gaussian random vector with a mean of 0 and a variance of . The normal distribution; where It is the variance of the noise; It is an identity matrix, with dimensions equal to... Same; noise variance Based on differential privacy parameters OK; After receiving the noise parameters, the server performs the aggregation operation: in, It is the first The global model parameters after each training round are aggregated from all drone uploads. The result.

Images