Adaptive optimization method of training parameters for online learning of UAV target recognition

Abstract

The present invention discloses a method for adaptively optimizing training parameters for online learning of drone target recognition. The method comprises the following steps: using images actually captured by a drone as subsequent online learning samples to continuously train the drone target recognition model; during the continuous training process, dividing the online learning samples according to a set batch size, and iteratively training the drone target recognition model using the divided batch samples; calculating the convergence rate of the drone target recognition model based on the computation time and loss value obtained after each iteration; comparing the calculated convergence rate of the current iteration cycle with the historical optimal convergence rate, and updating the optimal convergence rate or batch size based on the comparison result. The present invention effectively improves robustness and reduces training latency.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle (UAV) target recognition, and in particular to a training parameter adaptive optimization method for online learning of UAV target recognition. Background Art

[0002] Drones, due to their flexibility, portability, and 3D mobility, play a vital role in both civilian and military applications. Leveraging their rapid mobility to track and identify targets is a key application of drone intelligence, playing a crucial role in disaster relief, power line inspection, and other tasks. To achieve these applications, the use of artificial intelligence (AI), particularly deep learning, is inevitable.

[0003] Deep learning is a research direction in machine learning that aims to enable machines to acquire human-like analytical and learning capabilities, enabling them to recognize data such as text, images, and sound. Unlike shallow learning, deep learning models have more layers in their structure and enhance feature learning capabilities. Applying deep learning to drones effectively leverages their maneuverability and flexibility, providing a large number of learning samples for deep learning. Furthermore, these samples, thanks to the drone's three-dimensional maneuverability, can be obtained from various target heights and angles, resulting in more accurate and versatile trained models.

[0004] Although deep learning has proven its ability to make drones intelligent, traditional deep learning models still have the following shortcomings:

[0005] 1. Difficulty adapting to changing environments. Since the actual application environment of drones is often different from the environment of training samples, the detection effect of drone models trained well offline is not ideal in dynamic environments.

[0006] 2. The training process is time-consuming. Online learning can effectively improve model accuracy, but it requires a significant amount of computational time. Furthermore, as the number of samples increases, the training time per iteration also increases. This is a fatal flaw for drones that prioritize timeliness, and the flight time of small drones is insufficient to support training requirements.

[0007] In summary, the root cause of these two issues lies in the inability of drones to obtain high-quality samples in real time during target detection and recognition, and to adjust the sample size and target detection model online based on their specific circumstances. Therefore, adaptive optimization methods for training parameters for online learning of drone target recognition are needed to support real-time model optimization.

[0008] Patent number CN112732960B discloses an image classification method based on online federated learning. The method sets batch size selection criteria based on the characteristics of machine learning for image classification. By adaptively and individually adjusting the batch size of each working node, the method effectively improves the utilization rate of image data during training, while effectively mitigating fluctuations in batch size and further improving the stability of convergence during training. Furthermore, by introducing a learning rate proportional to the batch size, the negative impact of a decrease in batch size on training is mitigated, further improving the convergence speed of training. This invention only uses historical experience from locally stored datasets to adjust the batch size, which cannot meet the needs of drones performing target detection in dynamic environments. The batch adjustment method it employs is also not suitable for online learning samples acquired in dynamic environments. Summary of the Invention

[0009] To address the existing technical issues of drones' inability to balance accuracy and timeliness in target detection and recognition, this paper proposes an adaptive optimization method for training parameters for online learning of drone target recognition. By dynamically adjusting the batch size based on model performance and the number of online samples during online learning, the method meets the needs of drones for target detection in dynamic environments, effectively improving robustness and reducing training latency. The entire training process maintains a high convergence rate and strong robustness, enabling online learning to be completed in the shortest possible time. Furthermore, the present invention incorporates a mechanism for identifying and discarding outdated samples, further reducing computing resources for the device and enabling the algorithm to be better deployed on drone embedded platforms.

[0010] In order to achieve the above technical objectives, the technical solution adopted by the present invention is:

[0011] A training parameter adaptive optimization method for online learning of unmanned aerial vehicle target recognition, the training parameter adaptive optimization method comprising the following steps:

[0012] A drone target recognition model is built based on a neural network. After preliminary training of the drone target recognition model, images actually taken by the drone are used as subsequent online learning samples to continuously train the drone target recognition model.

[0013] During continuous training, the online learning samples are divided according to the set batch size, and the UAV target recognition model is iteratively trained using the divided batch samples. The convergence speed of the UAV target recognition model is calculated based on the calculation time and loss value obtained after each iteration. The calculated convergence speed of the current iteration cycle is compared with the historical optimal convergence speed, and the optimal convergence speed or batch size is updated based on the comparison result.

[0014] Among them, if the calculated convergence speed is greater than the historical optimal convergence speed, maintain the same batch size in the next iteration cycle, and update the historical optimal convergence speed to the convergence speed of the current iteration cycle; if the calculated convergence speed is less than the historical optimal convergence speed, increase the batch size to speed up the convergence speed of the next iteration cycle or reduce the batch size to reduce the calculation time of the next iteration cycle.

[0015] Furthermore, a UAV target recognition model was constructed based on the Yolov5 network.

[0016] Furthermore, several Visdrone2021 dataset images were used as initial samples to conduct preliminary training of the drone target recognition model.

[0017] Furthermore, based on the calculation time and loss value obtained after each iteration, the process of calculating the convergence speed of the UAV target recognition model includes the following steps:

[0018] Before each round of model iteration, the training data set D = (x1, y1), (x2, y2), ... (x n ,y n ), the number of online samples N in n∈N, where x n is the eigenvalue of the nth sample, y n is the label of the nth sample; according to the batch size S of the tth iteration t Divide the online samples into batches and calculate the convergence rate θ of the tth iteration based on the training results t :

[0019]

[0020]

[0021] Among them, L t Indicates the loss value of the model at the tth iteration, B t Indicates the number of batches for the tth iteration, T t represents the computation time of the tth iteration, (x i，j ,y i，j ) represents the i-th sample of the j-th batch, w t (x i，j ) are model parameters, f(y i，j |w t (x i，j )) is its loss function.

[0022] Furthermore, if the calculated convergence rate is lower than the historical optimal convergence rate, the following formula is used to increase the batch size to speed up the convergence rate of the next iteration cycle or to reduce the batch size to reduce the calculation time of the next iteration cycle:

[0023] S t+1 =η t *S t

[0024]

[0025]

[0026] Among them, σ(θ t-1 ,θ t-2 …θ t-u ) represents the θ of the first u iterations t The variance of E(θ t-1 ,θ t-2 …θ t-u ) represents the θ of the first u iterations t The mean of the training dataset is CoV, S t+1 Indicates the number of batches for the t+1th iteration, S t Indicates the batch size of the t-th iteration.

[0027] Furthermore, the calculation formula of the historical optimal convergence speed is:

[0028] θ * =θ′*ε;

[0029] Where θ′ is the minimum convergence rate in the historical iteration cycle actually calculated; ε is the delay factor used to prevent the convergence rate θ of any round in the training process t Less than θ * , the value range is [0.05, 1.00].

[0030] Furthermore, the training parameter adaptive optimization method further includes the following steps:

[0031] The sample convergence rate φ of the jth batch i-th online learning sample is calculated using the following formula: i，j :

[0032]

[0033] where θ j is the convergence rate of the jth batch, τ i，j is the number of training iterations for the i-th sample in the j-th batch, i∈1,2…S t , j∈1,2…B t ; It is the iteration factor, which is used to represent the increase in the number of iterations and has a value range of [0.05, 1.00];

[0034] The sample convergence rate φ is taken from the training set i，jSamples with a convergence speed lower than the preset threshold are eliminated.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] First, the adaptive optimization method for training parameters for online learning of drone target recognition of the present invention achieves faster convergence of the online learning training of the target recognition model by dynamically adjusting the batch size, thereby reducing the total training time without adding additional overhead.

[0037] Second, the adaptive optimization method for training parameters for online learning of drone target recognition of the present invention has excellent robustness and effectiveness and can be applied in a variety of scenarios;

[0038] Third, the adaptive optimization method of training parameters for online learning of drone target recognition of the present invention saves device computing resources by discarding outdated samples, so that the algorithm can be better deployed on the drone embedded platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 The impact of different batch sizes on single iteration time and accuracy during model training in an embodiment of the present invention;

[0040] Figure 2 is the accuracy of different adaptive batch size methods during the model training process of an embodiment of the present invention;

[0041] Figure 3 is the computation time of different adaptive batch size methods during the model training process of an embodiment of the present invention;

[0042] Figure 4 This is a flow chart of a method for adaptively optimizing training parameters for online learning of drone target recognition according to an embodiment of the present invention. DETAILED DESCRIPTION

[0043] The embodiments of the present invention are described in further detail below with reference to the accompanying drawings.

[0044] See also Figure 4 This embodiment discloses a method for adaptively optimizing training parameters for online learning of drone target recognition. The method comprises the following steps:

[0045] A drone target recognition model is built based on a neural network. After preliminary training of the drone target recognition model, images actually taken by the drone are used as subsequent online learning samples to continuously train the drone target recognition model.

[0046] During continuous training, the online learning samples are divided according to the set batch size, and the UAV target recognition model is iteratively trained using the divided batch samples. The convergence speed of the UAV target recognition model is calculated based on the calculation time and loss value obtained after each iteration. The calculated convergence speed of the current iteration cycle is compared with the historical optimal convergence speed, and the optimal convergence speed or batch size is updated based on the comparison result.

[0047] Among them, if the calculated convergence speed is greater than the historical optimal convergence speed, maintain the same batch size in the next iteration cycle, and update the historical optimal convergence speed to the convergence speed of the current iteration cycle; if the calculated convergence speed is less than the historical optimal convergence speed, increase the batch size to speed up the convergence speed of the next iteration cycle or reduce the batch size to reduce the calculation time of the next iteration cycle.

[0048] Figure 1 The Visdrone2021 drone dataset was used to test model performance at different batch sizes. The comparison metrics used were loss, convergence speed, and average time per iteration. Batch sizes of 2, 4, 8, and 16 were selected. Observing the images, it is clear that as the batch size decreases, the loss curve fluctuates more widely, the model gradient shifts more pronounced, and the model converges faster. In contrast, as the batch size increases, the time to complete a single iteration decreases due to the increased number of samples computed per iteration, while the model converges more slowly. Furthermore, given the limited computing resources of drone embedded devices, overly large batch sizes can actually increase training time per iteration. For example, increasing the batch size from 2 to 16 increases overall training time. This is because the CPU is required to load and preprocess online samples, while the GPU is required to calculate the loss and gradient descent. When the batch size is too large, the CPU and GPU struggle to provide sufficient computing resources, increasing waiting time and ultimately decreasing performance. Since the batch size is closely related to the convergence speed of the training model, selecting an appropriate batch value is a direct means to improve the efficiency of the model.

[0049] The core idea of ​​the batch adaptive optimization algorithm based on online learning proposed in this paper is to calculate the convergence rate θ and keep it above the expected value. Before each round of iteration, the machine learning system will reconfirm the number of online samples N in the training data set and adjust the number of samples based on the current batch size S. t Divide the samples into batches, and calculate the convergence rate θ based on the training results after each iteration t .

[0050]

[0051]

[0052] Among them L t Indicates the loss value of the model in the t-th iteration process, B t Indicates the value of batch, T t Indicates the computation time of the t-th iteration process, (x i，j ,y i，j ) represents the i-th sample of the j-th batch, w t (x i，j ) are model parameters, f(y i，j |w t (x i，j )) is its loss function. In the present invention, let L t and T t The ratio of represents the convergence rate θ t By comparing the current convergence rate θ t and the historical optimal convergence rate θ * , the present invention can judge the current S t The performance of the model is suitable and optimized. In addition, as the number of model training increases, the performance of the model is also continuously optimized. Compared with the loss value L in the early stage of training, L t gradually decreases, and accordingly θ t Therefore, the present invention sets a delay factor ε to prevent the convergence speed θ of any round in the training process. t Less than θ * ε is also the criterion for adaptive adjustment of batch size. The specific judgment process is as follows:

[0053] 1) When θ t >θ * When the current S t The model training effect is better under , there is no need to adjust the batch size, and maintain the same batch in the next iteration, that is, S t+1 =S t ;

[0054] 2) When θ t <θ * When the current S t The computing resources cannot be fully utilized, and the batch size needs to be adjusted before the next iteration to improve the model performance.

[0055] Generally speaking, in order to adapt to the parallel accelerator in the UAV embedded device, the batch size is usually set to 2 m , m∈M, so the batch of the next iteration is adjusted to S t twice or half, that is:

[0056] S t+1 =η t *S t

[0057]

[0058]

[0059] Where σ(θ t-1 ,θ t-2 …θ t-u ) represents the θ of the first u iterations t The variance of E(θ t-1 ,θ t-2 …θ t-u ) represents its mean, and the ratio of the two represents the coefficient of variation, recorded as CoV. The coefficient of variation is suitable for comparing the degree of dispersion between data with different dimensions or the same dimensions but with large differences in the concentration of the data. The specific comparison results are as follows:

[0060] When CoV is relatively small, it means that the data distribution is concentrated and the model is still in the early training stage. At this time, it is necessary to consider reducing S t To improve the convergence speed; when CoV is relatively large, it means that the data tends to be dispersed and the model tends to converge. At this time, the model has a high accuracy and fluctuates around the local optimal solution. It can be seen that there is little room for improvement in this model. Reducing S t It does not improve the convergence speed, but instead makes the model move away from the local optimal solution. Therefore, the present invention will adaptively increase S t , reducing the computation time of each iteration. In addition, considering the limited computation freedom of the UAV embedded device, the present invention also sets the maximum value of the batch S max , to prevent insufficient resources from slowing down the training process.

[0061] Due to the real-time updating nature of the online learning process, there is a time lag in the addition of online samples. By the time new samples are added, some outdated samples have already undergone multiple rounds of iteration, and their contribution to model performance is almost zero. This paper experimentally verifies the importance of screening and eliminating outdated samples in the training set for improving model performance. Without adjusting the batch size, the loss values ​​for different batch values ​​were tested. The experimental results show that the samples added earliest to the dataset basically converged after 15 rounds of iterations, and subsequent iterations did not promote model improvement.

[0062] Based on the above experimental results, in order to save a lot of computing resources and single-iteration training time, the present invention designs a sample convergence rate φ to screen and eliminate outdated samples. At the same time, the elimination of outdated samples will affect the training convergence rate θ. For batch j (j∈1, 2…B t), the online sample i(i∈1,2…S t )The sample convergence rate φ in the current iteration is as follows:

[0063]

[0064] where θ j is the convergence rate of the jth batch, τ i，j is the number of training iterations for the i-th sample in the j-th batch. Similar to the delay factor ε mentioned in the previous section, the present invention defines represents the increase in the number of iterations, with a value range of [0.05, 1.00]. As the model continues to train, the sample convergence rate φ of the online samples will decrease accordingly. During the training process, when the model accuracy reaches the expected value, the present invention automatically identifies and eliminates some samples with the lowest sample convergence rate φ.

[0065] The actual experimental operation process of the embodiment is introduced below.

[0066] This embodiment uses the Jetson agx (32G GPU) developed by NVIDIA as the embedded device on the reconnaissance drone. The yolov5 based on the PyTorch framework is selected as the training model. Its single-step detection algorithm can quickly identify the target while ensuring a high accuracy rate. A small number of Visdrone2021 drone dataset images are selected as the initial samples for preliminary training of the model, and the images actually taken by the drone are used as subsequent online learning samples. At the same time, the maximum batch size is set to S max =16, the initial batch size is S0=2, the first delay factor ε=0.95, and the decision factor ψ of CoV=0.15.

[0067] In order to simulate the scenario where drones continuously add online samples for model training, the experiment is set up to add new online samples in fixed iteration rounds, and the average average precision (mAP) and calculation time of the model are recorded after each iteration.

[0068] A) Performance of different adaptive batch size methods

[0069] In this part of the experiment, this example uses different adaptive batch size methods to evaluate model performance. The specific settings are as follows:

[0070] OLBF (Online Learning based Adaptive Batch Size Fitting): This embodiment proposes an adaptive optimization method for training parameters for online learning of drone target recognition.

[0071] DBF (Dynamic batch size fitting): After every n iterations (n=5), the batch size is adjusted using the loss value;

[0072] FixUp (Designated Batch Size on Empirical Rules): Adjust the batch size at fixed intervals based on experience.

[0073] Fixed: Use a fixed batch size.

[0074] Figure 2 The graph shows the curves of mAP changing with computation time under the four different batch adaptation methods mentioned above, and training is stopped when mAP is greater than 0.99. Obviously, the DBF method has a higher convergence speed in the early stages of training, but with the addition of online samples, it is difficult to maintain high performance. This is because DBF does not consider the impact of online samples, and a large time difference is generated after multiple rounds of comprehensive evaluation of historical loss values. In addition, the DBF method does not take into account the phenomenon that the convergence speed will decrease as the model is continuously trained, while the present invention takes this into account. FixUp adjusts the batch size based on historical experience, and its performance is similar to that of DBF, and even outperforms DBF in the later stages of training. However, FixUp requires a lot of preliminary experiments and can only make a rough adjustment, which is difficult to achieve in actual scenarios. However, the OLBF method provided in this embodiment maintains high performance and robustness throughout the training process and can complete the online learning task in the shortest time.

[0075] Figure 3 The total training time for different adaptive batch adjustment methods is recorded. Intuitively, this embodiment significantly shortens training time, by 31.6%, 34.4%, and 38.8% compared to the other three methods, respectively. This embodiment reduces the number of iterations by improving the convergence speed, thereby shortening the total training time. Furthermore, embedded devices require sample input in a fixed batch size, so the number of training data sets is not 2. m , m∈M, will cause computational redundancy and reduce model performance. The real-time batch size adaptation designed in this embodiment can effectively avoid such computational redundancy and maintain and improve efficiency.

[0076] B) Performance of the stale sample elimination mechanism

[0077] This example records the training time required for each 100 iterations of the adaptive batch scaling method with and without an elimination mechanism. OLBF identifies and eliminates samples at t-40, 60, and 80. It can be clearly seen that, in the same iteration, OLBF reduces training time by 9%, 12%, and 18%, respectively, compared to the method without an elimination mechanism. As the model training process progresses, the time difference between the two methods increases.

[0078] Furthermore, to further validate the accuracy of this invention, this example tested the mAP of two experimental groups at different IoU thresholds. It can be seen that the elimination mechanism included in this example reduced the total training time required for 100 iterations by 10%, and the mAP decreased by approximately 0.1%, a negligible decrease. This indicates that eliminating outdated samples does not affect the accuracy of the model, but it significantly reduces training time and improves training efficiency.

[0079] C) The elimination sample mechanism is outdated

[0080] To test the robustness of the adaptive batch size adjustment method in various scenarios, we used the present invention in a non-online learning scenario. The curves show that even without online samples added to the training dataset, this embodiment still improves the model's mAP and convergence speed. This is because a fixed batch size can limit the exploration of the policy space, while adaptive batch adjustment facilitates the exploration of various possible solutions.

[0081] In addition, different YOLOv5 models are also used in actual operation to test the strong robustness of the present invention. Since the differences between the neural network structures of different YOLOv5 models are applicable to different scenarios, the requirements for the input image size are different and the training speed is different. As shown in the data in the table, the present embodiment shows better performance on the YOLOv5n, YOLOv5s and YOLOv5m models, but poor performance on the YOLOv51. This is due to the differences between embedded devices. Limited computing resources make it difficult for embedded devices to train large models such as YOLOv51. However, even if computing resources are limited, the present embodiment still improves its performance in model training, which can confirm the strong robustness of the method described in the present embodiment.

[0082] In summary, the method described in this embodiment is suitable for drone target detection in dynamic environments. It can dynamically adjust the batch size according to the current performance of the model and the number of online samples. The convergence speed on the mainstream target detection model YOLV is greatly improved, and the total training time is reduced without adding any additional overhead. It has high accuracy and strong robustness, enabling drones to perform better on embedded platforms.

[0083] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiment of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and literal translation scripting language JavaScript, etc.

[0084] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0085] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0086] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions for executing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0087] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present application.

[0088] Obviously, those skilled in the art may make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is intended to include these modifications and variations.

Claims

1. A training parameter adaptive optimization method for online learning of UAV target recognition, characterized by: The training parameter adaptive optimization method comprises the following steps: A drone target recognition model is built based on a neural network. After preliminary training of the drone target recognition model, images actually taken by the drone are used as subsequent online learning samples to continuously train the drone target recognition model. During continuous training, the online learning samples are divided according to the set batch size, and the UAV target recognition model is iteratively trained using the divided batch samples. The convergence speed of the UAV target recognition model is calculated based on the calculation time and loss value obtained after each iteration. The calculated convergence speed of the current iteration cycle is compared with the historical optimal convergence speed, and the optimal convergence speed or batch size is updated based on the comparison result. Among them, if the calculated convergence speed is greater than the historical optimal convergence speed, maintain the same batch size in the next iteration cycle, and update the historical optimal convergence speed to the convergence speed of the current iteration cycle; if the calculated convergence speed is less than the historical optimal convergence speed, increase the batch size to speed up the convergence speed of the next iteration cycle or reduce the batch size to reduce the calculation time of the next iteration cycle; According to the calculation time and loss value obtained after each iteration, the process of calculating the convergence speed of the UAV target recognition model includes the following steps: Before each round of model iteration, the training data set D=(x1,y1),(x2,y2),…(x n ,y n ), the number of online samples N in n∈Ν, where x n is the eigenvalue of the nth sample, y n is the label of the nth sample; according to the batch size S of the tth iteration t Divide the online samples into batches and calculate the convergence rate θ of the tth iteration based on the training results t : Among them, L t Indicates the loss value of the model at the tth iteration, B t Indicates the number of batches for the tth iteration, T t represents the computation time of the tth iteration, (x i,j ,y i,j ) represents the i-th sample of the j-th batch, w t (x i,j ) are model parameters, f(y i,j |w t (x i,j )) is its loss function; If the calculated convergence rate is lower than the historical optimal convergence rate, use the following formula to increase the batch size to speed up the convergence rate of the next iteration cycle or reduce the batch size to reduce the calculation time of the next iteration cycle: S t+1 =the t *S t Among them, σ(θ t-1 ,θ t-2 …θ t-u ) represents the θ of the first u iterations t The variance of E(θ t-1 ,θ t-2 …θ t-u ) represents the θ of the first u iterations t The mean of the training dataset is CoV, S t+1 Indicates the number of batches for the t+1th iteration, S t Indicates the batch size of the t-th iteration.

2. The method for adaptive optimization of training parameters for online learning of UAV target recognition according to claim 1 is characterized in that: The UAV target recognition model is constructed based on the Yolov5 network.

3. The method for adaptive optimization of training parameters for online learning of UAV target recognition according to claim 1 is characterized in that: Several Visdrone2021 dataset images are used as initial samples to conduct preliminary training of the drone target recognition model.

4. The method for adaptive optimization of training parameters for online learning of UAV target recognition according to claim 1 is characterized in that: The calculation formula of the historical optimal convergence speed is: i * =θ′*ε; Where θ′ is the minimum convergence rate in the historical iteration cycle actually calculated; ε is a delay factor that prevents the convergence rate θ of any round during training. t Less than θ * , the value range is [0.05,1.00].

5. The method for adaptive optimization of training parameters for online learning of UAV target recognition according to claim 1 is characterized in that: The training parameter adaptive optimization method further comprises the following steps: The sample convergence rate φ of the jth batch i-th online learning sample is calculated using the following formula: i,j : where θ j is the convergence rate of the jth batch, τ i,j is the number of training iterations for the i-th sample in the j-th batch, i∈1,2…S t ,j∈1,2…B t ; It is the iteration factor, which represents the increase in the number of iterations and has a value range of [0.05, 1.00]; The sample convergence rate φ is taken from the training set i,j Samples with a convergence speed lower than the preset threshold are eliminated.

Images