A method and system for recognizing noise of a UAV based on a lightweight YOLO network

Abstract

The application discloses a kind of unmanned aerial vehicle noise identification method and system based on light YOLO network, it is related to unmanned aerial vehicle acoustic signal processing and target identification technical field, including: obtaining unmanned aerial vehicle noise sample and city background noise sample, based on unmanned aerial vehicle noise sample and city background noise sample processing sample, obtain unmanned aerial vehicle noise data set under complex noise environment, based on the preset mel frequency cepstrum coefficient MFCC feature extraction method one-dimensional time domain audio signal to two-dimensional feature map conversion under complex noise environment unmanned aerial vehicle noise data set, obtain MFCC feature extraction graph data set, MFCC feature extraction graph data set is input into the YOLOv8n-cls light network model established in advance and is trained, obtain the YOLOv8n-cls light network model after training, input into the YOLOv8n-cls light network model after training to be identified unmanned aerial vehicle noise data, output obtains unmanned aerial vehicle noise identification result, to be able to improve identification performance.

Description

Technical Field

[0001] This invention relates to the field of acoustic signal processing and target recognition technology for unmanned aerial vehicles (UAVs), specifically a method and system for UAV noise recognition based on a lightweight YOLO network. Background Technology

[0002] With the rapid development of drone technology, drones have been widely used in logistics delivery, power line inspection, agricultural plant protection, aerial photography and mapping, and emergency rescue. However, the widespread use of drones has also raised low-altitude airspace safety issues, such as illegal flights, air traffic intrusions, and unauthorized reconnaissance of important areas, seriously threatening public safety and the protection of key targets. Therefore, research on drone target detection and identification is of significant practical importance. Existing drone detection technologies mainly include radar detection, radio detection, photoelectric detection, and acoustic detection. Among these, acoustic detection has good application potential in close-range low-altitude monitoring scenarios due to its advantages such as low cost, ease of deployment, adaptability to low-light or obstructed environments, and less dependence on the electromagnetic environment. Drones generate regular rotor and motor noise during flight; these acoustic characteristics can serve as important evidence for drone target identification.

[0003] Traditional UAV noise recognition methods typically combine manual feature extraction with shallow classifiers. For example, acoustic features such as Mel-frequency cepstral coefficients (MFCC) and short-time Fourier transform (STFT) spectra are extracted and then input into support vector machines, K-nearest neighbors, multilayer perceptrons, or hidden Markov models for classification. These methods are simple in structure and fast in training, showing some recognition performance in laboratory environments. However, when applied to complex noise environments, existing UAV noise recognition methods struggle to adapt to the complex acoustic differences caused by variations in UAV flight status, distance, attitude, and background environment. This leads to a decline in their ability to learn deep discriminative information in complex noise, resulting in a significant drop in recognition performance. Summary of the Invention

[0004] To address the shortcomings mentioned in the background section, the present invention aims to provide a method and system for drone noise identification based on a lightweight YOLO network.

[0005] Firstly, the objective of this invention can be achieved through the following technical solution: a method for drone noise identification based on a lightweight YOLO network, the method comprising the following steps: We acquire drone noise samples and urban background noise samples, and perform sample processing based on the drone noise samples and urban background noise samples to obtain a drone noise dataset under complex noise environment. Based on the MFCC feature extraction method with preset Mel frequency cepstral coefficients, a one-dimensional time-domain audio signal of a UAV noise dataset under complex noise environment is converted into a two-dimensional feature map to obtain an MFCC feature extraction map dataset. The MFCC feature extraction map dataset is input into a pre-built YOLOv8n-cls lightweight network model for training, resulting in a trained YOLOv8n-cls lightweight network model. The noise data of the drone to be identified is then obtained and input into the trained YOLOv8n-cls lightweight network model, and the drone noise identification result is output.

[0006] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the Mel frequency cepstral coefficient (MFCC) feature extraction method includes pre-emphasis, framing and windowing, fast Fourier transform, Mel filter bank filtering, logarithmic operation and discrete cosine transform.

[0007] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the pre-emphasis process, comprising: Raw signals within a UAV noise dataset under complex noise environments Through a first-order high-pass filter: in These are preset coefficients; The process of framing and windowing includes: The pre-emphasized signal is multiplied by a window function that smoothly transitions to zero at both ends. The expression for the window function is as follows: in Given a frame length, the windowed signal is... .

[0008] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the process of the Fast Fourier Transform, as follows: Perform an FFT on the windowed signal s[n] of each frame to transform it from the time domain to the frequency domain, and obtain the spectral distribution S[k] of that frame: Then calculate the power spectrum: ; The filtering process of the Mel filter bank is as follows: The power spectrum is passed through a set of Mel-scale triangular filters, with the center frequencies of each filter being equally spaced on the Mel scale, exhibiting a distribution characteristic of dense low frequencies and sparse high frequencies on the Hz scale. No. The frequency response of a triangular filter is defined as: in For the first The center frequency of each filter For frequency indexing.

[0009] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the process of the logarithmic operation, as follows: Take the natural logarithm of the output energy of each Mel filter: The process of discrete cosine transform is as follows: Log-log energy sequence By performing a Discrete Cosine Transform (DCT) and utilizing its energy concentration characteristics, the information output by the filter bank is compressed into low-dimensional coefficients. Simultaneously, the linear correlation between features in each dimension is removed, ultimately yielding the MFCC coefficients, as follows: in, Indicates the first Vermeer frequency cepstral coefficients; This represents the total number of filters within the Mel filter bank. The filter number; For the first The value obtained by performing a logarithmic operation on the output energy of a Mel filter; The dimension of the final output MFCC coefficients; This refers to the MFCC coefficient sequence number.

[0010] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the YOLOv8n-cls lightweight network model performs multi-scale feature learning through the backbone network, combines the classification neck network and the classification head network to output a multi-dimensional probability vector, and takes the category corresponding to the maximum value after normalization by the Softmax function as the UAV noise recognition result.

[0011] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the backbone network is composed of a first convolutional module, a second convolutional module, four C2f modules and one SPPF module cascaded sequentially; the classification neck network is an adaptive average pooling layer that reduces the spatial size of the feature map output by the backbone network; the classification head network sequentially includes a flattening layer, a first dropout layer, a first fully connected layer, a second dropout layer and a second fully connected layer, wherein the first fully connected layer and the second fully connected layer map the global feature vector to a high-dimensional space through linear transformation to enhance class separability; the first dropout layer and the second dropout layer randomly discard some neurons during training to prevent overfitting, and finally output a multi-dimensional probability vector.

[0012] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the expression of the YOLOv8n-cls lightweight network model is as follows: Let the input be YOLOv8n-cls as a whole can be regarded as a composite function. Mapping the input to a class probability vector ,in, Number of categories; The mathematical meanings of each part are as follows: backbone network It is composed of convolution, batch normalization, SiLU activation, C2f module and SPPF module in sequence: Where X is the input feature map, each... Represents the triples of convolution with batch normalization and SiLU; each It consists of multiple Bottleneck residual blocks stacked together, with cross-layer stitching and fusion; SPPF is the concatenation and convolution after serial max pooling, and the final output is a feature map. ; Classification of neck For adaptive average pooling layers, the spatial dimension is compressed to 1×1: Obtain the global vector ; Category Header global feature vectors The forward propagation process, which maps the probability distributions of each category, is as follows: in, For flattening operation This is a randomly deactivated layer with a dropout probability of 0.2; This is the first fully connected layer; For the second fully connected layer Let be the category score vector output by the fully connected layer, and let its be the . The component indicates that the input belongs to the first... The unnormalized score of the class will After processing with the Softmax normalization function, the probability distributions of each class are obtained: in, For input belonging to the first The probability of a class The sum of the scores for all categories.

[0013] Secondly, in order to achieve the above objectives, this invention discloses a drone noise recognition system based on a lightweight YOLO network, comprising: The data acquisition module is used to acquire drone noise samples and urban background noise samples, and to perform sample processing based on the drone noise samples and urban background noise samples to obtain drone noise dataset under complex noise environment. The data processing module is used to convert one-dimensional time-domain audio signals into two-dimensional feature maps of UAV noise datasets in complex noise environments based on the preset Mel frequency cepstral coefficients (MFCC) feature extraction method, and obtain MFCC feature extraction map datasets. The recognition output module is used to input the MFCC feature extraction map dataset into the pre-established YOLOv8n-cls lightweight network model for training, obtain the trained YOLOv8n-cls lightweight network model, acquire the drone noise data to be identified, input the drone noise data to be identified into the trained YOLOv8n-cls lightweight network model, and output the drone noise recognition result.

[0014] The beneficial effects of this invention are: This invention uses the YOLOv8n-cls lightweight network as the classification model, which can reduce model complexity and improve inference efficiency while ensuring high recognition accuracy. Through deep feature extraction, it can effectively improve the model's accuracy in dealing with complex background noise and improve recognition performance. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of MFCC feature extraction for nine types of sounds, excluding drone and drilling noise, according to the present invention. Figure 3 This is a schematic diagram illustrating the extraction of drone and drilling sound and MFCC features according to the present invention; Figure 4 This is a schematic diagram of the YOLOv8n-cls network structure of the present invention; Figure 5 This is a schematic diagram of the SqueezeNet network structure of the present invention; Figure 6 This is a schematic diagram of the MobileNetV2 network structure of the present invention; Figure 7 This is a schematic diagram of the ResNet50 network structure of the present invention; Figure 8 This is a schematic diagram of the confusion matrix of the YOLOv8n-cls network on the test set according to an embodiment of the present invention; Figure 9 This is a schematic diagram of the confusion matrix of the SqueezeNet network on the test set according to an embodiment of the present invention; Figure 10 This is a schematic diagram of the confusion matrix of the MobileNetV2 network on the test set according to an embodiment of the present invention; Figure 11 This is a schematic diagram of the confusion matrix of the ResNet50 network on the test set according to an embodiment of the present invention; Figure 12 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0016] 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, and 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.

[0017] Example 1: like Figure 1 As shown, a method for drone noise recognition based on a lightweight YOLO network includes the following steps: S101: Obtain drone noise samples and urban background noise samples, perform sample processing based on drone noise samples and urban background noise samples, and obtain drone noise dataset under complex noise environment. Specifically, the process of acquiring drone noise samples and urban background noise samples includes: (1) Drone noise samples from the Drone Audioset Detest dataset were selected as the target category data, i.e., "positive samples". This dataset includes two quadcopter models, large (DJIF450, wheelbase 45 cm) and small (DJIF330, wheelbase 33 cm), as well as low and high thrust states, multiple microphone positions above and below (distances of 25 cm and 50 cm), and a wide signal-to-noise ratio range of -57.2 dB to -2.5 dB. The drone noise samples include audio data under different models, different thrust states, different acquisition positions, and different signal-to-noise ratio conditions.

[0018] (2) In the selection of non-target category data, the present invention introduces urban environmental sound samples as "non-drone" negative samples, specifically from the publicly available UrbanSound8k dataset, which is one of the benchmark datasets widely used in the field of urban sound recognition. It contains 8,732 labeled audio segments of real on-site recordings, each with a duration of no more than 4 seconds, and comprehensively covers 10 categories of common urban noise, including: noise from mechanical equipment such as air conditioner outdoor units, engine idling, hand drills, and drilling; traffic warning noise such as car horns and alarms; noise from human activities such as children playing and street music; and biological sounds such as dog barking and sudden transient noise such as gunshots. The UrbanSound8k dataset fully meets the requirements of this invention for noise coverage in complex environments for three main reasons: First, all samples are from real-world recordings rather than laboratory synthesis, preserving the complex acoustic characteristics of multi-source aliasing, reverberation, and sudden interference in real urban environments. Second, the 10 categories exhibit significant differences in frequency domain distribution, temporal envelope, and energy characteristics, forming a complete characterization of the diversity of urban sound fields in terms of sound source properties. Third, the scale of 8732 labeled samples ensures sufficient acoustic variation within each category, allowing the model to be exposed to various manifestations of the same type of interference under different environmental conditions and recording devices during the training phase. By introducing this dataset as a source of negative samples, this invention ensures sufficient coverage of the training data in the interference space, enabling the model to be widely exposed to various potential interference sources during the training phase, thus possessing good adaptability and anti-interference performance in complex and ever-changing urban acoustic environments during actual deployment.

[0019] When the UrbanSound8k dataset was officially released, all audio files were distributed across 10 different subdirectories according to a preset number of folds. Samples of the same category were randomly scattered and assigned to different folds without being organized by category. Therefore, the original downloaded data presented a mixed category structure in terms of file organization. To construct training data with a clear structure and distinct categories, this invention performs systematic preprocessing on the downloaded raw data: parsing the metadata file released with the dataset to obtain the category label corresponding to each audio file; based on the label information, each audio file is extracted from its original fold directory and re-archived into a category subdirectory according to the category name for archiving.

[0020] (3) Perform stratified sampling and class balance processing on the drone noise samples and environmental sound samples to construct a balanced classification dataset. In the Drone Audioset Detest dataset, 1000 drone noise samples were selected as positive samples by stratified random sampling to achieve a balance between the number of data items in each category and the UrbanSound8k dataset.

[0021] (4) The UrbanSound8k dataset, which has been classified and archived, and 1,000 data points randomly extracted from the Drone AudiosetDetest dataset are fused together to form the UAV noise data under complex noise environment of this invention. This dataset is divided into training set, validation set and test set in a ratio of 98:1:1.

[0022] S102: Based on the preset Mel frequency cepstral coefficient MFCC feature extraction method, the one-dimensional time-domain audio signal of the UAV noise dataset under complex noise environment is converted into a two-dimensional feature map to obtain the MFCC feature extraction map dataset. Mel frequency cepstral coefficient (MFCC) feature extraction methods include pre-emphasis, framing and windowing, fast Fourier transform, Mel filter bank filtering, logarithmic operation, and discrete cosine transform.

[0023] The specific processing procedure is as follows: Step 1: Pre-weighting When a sound signal propagates through the air, the high-frequency components attenuate much faster than the low-frequency components, resulting in a high-frequency fading in the received signal spectrum. To compensate for this loss and eliminate the smoothing effect of the sound-generating system on the spectrum, it is necessary to process the original signal... Through a first-order high-pass filter: Where the coefficient Typically, a value of 0.97 is used. This processing keeps the signal spectrum flat across the high and low frequency bands, facilitating subsequent analysis.

[0024] Step 2: Frame Segmentation and Windowing Sound signals exhibit time-varying characteristics, but within a very short time window, typically considered to be 20-40 seconds. The continuous sound signal can be approximated as a stationary signal. Therefore, it is necessary to divide the continuous sound signal into several short frames, each with a duration of 20-40 seconds. To smooth the transition between frames, a 50% overlap is set between adjacent frames. However, directly framing the signal is equivalent to applying a rectangular window at the frame boundary, and this abrupt change introduces spurious high-frequency components into the spectrum—i.e., spectral leakage. To solve this problem, each frame signal needs to be multiplied by a window function that smoothly transitions to zero at both ends. The Hamming window is widely used due to its good sidelobe suppression characteristics, and its window function is expressed as follows: in Given a frame length, the windowed signal is... .

[0025] Step 3: Fast Fourier Transform Perform an FFT on the windowed signal s[n] of each frame to transform it from the time domain to the frequency domain and obtain the spectral distribution of that frame: Then calculate the power spectrum Step 4: Mel filter bank filtering The power spectrum is passed through a set of Mel-scale triangular filters, with the center frequencies of each filter equally spaced on the Mel scale, exhibiting a distribution characteristic of dense low frequencies and sparse high frequencies on the Hz scale.

[0026] No. The frequency response of a triangular filter is defined as: in For the first The center frequency of each filter For frequency indexing.

[0027] Step 5: Logarithmic Operations The human ear's perception of sound pressure is also non-linear—its logarithmic response means that the ear's sensitivity to changes in sound pressure decreases as the sound pressure level increases. To simulate this characteristic, the natural logarithm of the output energy of each Mel filter obtained in step four is taken: Step Six: Discrete Cosine Transform The logarithmic energy sequence output in step 5 The sequence still exhibits high correlation between its various dimensions, which reduces the training efficiency of subsequent classification models. Therefore, a Discrete Cosine Transform (DCT) is required. Utilizing the energy concentration properties of DCT, the information output by the filter bank is compressed into low-dimensional coefficients, while simultaneously removing the linear correlation between features in each dimension, ultimately yielding the MFCC coefficients.

[0028] In this embodiment, the invention is implemented using Python, and MFCC feature extraction and visualization transformation are performed on all original audio samples in the dataset based on the librosa audio processing library. When loading audio, the sampling rate parameter sr=None is set to preserve the original sampling rate of each audio file to maintain the frequency band integrity of the original signal. During the extraction process, the number of MFCC coefficients is set to 13, the frame shift length is set to 512 sampling points, the FFT window size uses the default value of 2048, and the number of Mel filter banks is 128. After completing the above parameter configuration, the librosa.feature.mfcc function is called to extract the MFCC coefficients of each frame. The coefficients are then converted from amplitude to dB scale, and the corresponding two-dimensional spectrogram is generated using the librosa.display.specshow function. The spectrogram is saved as a 224×224 pixel PNG image with an output resolution of 100 dpi. The main parameters for generating the MFCC image are shown in Table 1. After the above processing, each audio sample in the dataset is converted into a corresponding MFCC feature image and archived in the same directory structure as the original audio, finally generating a complete feature map dataset for model training, validation and testing in step three.

[0029] Table 1 Main parameters for generating MFCC diagrams MFCC features were extracted from 11 types of sound samples, including ten typical environmental noises such as drone noise, air conditioner noise, car horn noise, children playing noise, dog barking noise, drilling noise, engine idling noise, gunshot noise, jackhammer noise, police siren noise, and street music, as well as drone noise. The results are as follows: Figure 2 The image shown is an MFCC feature extraction map excluding drone and drilling noise. Figure 3 The image shows the MFCC feature extraction results for mechanical noises such as drilling noise. These two sounds have certain similarities in acoustic characteristics.

[0030] Figure 3The results show that there are significant differences between drilling noise and UAV noise in terms of temporal continuity, periodicity, and frequency energy distribution in the MFCC spectrum. This demonstrates that the MFCC-based visualization method can intuitively transform the physical characteristics of the two sound sources into pattern differences on the spectrum, thus providing clear feature basis for subsequent classification and recognition tasks.

[0031] S103: Input the MFCC feature extraction map dataset into the pre-established YOLOv8n-cls lightweight network model for training to obtain the trained YOLOv8n-cls lightweight network model. Obtain the drone noise data to be identified, input the drone noise data to be identified into the trained YOLOv8n-cls lightweight network model, and output the drone noise identification result.

[0032] The YOLOv8n-cls lightweight network model learns multi-scale features through a backbone network, and combines a classification neck network and a classification head network to output a multi-dimensional probability vector. After normalization by the Softmax function, the category corresponding to the maximum value is taken as the drone noise recognition result.

[0033] The backbone network consists of a first convolutional module, a second convolutional module, four C2f modules, and one SPPF module cascaded together. Specifically, the first convolutional module includes a convolutional layer with a kernel size of 3×3, a stride of 2, a padding of 1, and a number of kernels of 64, a batch normalization layer, and a SiLU activation layer; the second convolutional module includes a convolutional layer with a kernel size of 3×3, a stride of 2, a padding of 1, and a number of kernels of 128, a batch normalization layer, and a SiLU activation layer; the four C2f modules are respectively set in layers 2, 4, 6, and 8, and each C2f module is formed by stacking Bottleneck modules with a depth factor of 0.33, with output channels of 128, 256, 512, and 1024 respectively; the SPPF module includes three parallel max pooling layers with pooling kernel sizes of 5×5, 9×9, and 13×13 respectively, with an output channel count of 1024.

[0034] The classification neck network is an adaptive average pooling layer that reduces the spatial size of the feature map output by the backbone network. The classification head network consists of a flattening layer, a first dropout layer, a first fully connected layer, a second dropout layer, and a second fully connected layer. The first and second fully connected layers map the global feature vector to a high-dimensional space through linear transformation to enhance class separability. The first and second dropout layers randomly drop some neurons during training to prevent overfitting, and finally output a multi-dimensional probability vector.

[0035] The first fully connected layer has an input dimension of 1024 and an output dimension of 1280. The second fully connected layer has an input dimension of 1280 and an output dimension equal to the number of classes. The dropout probability of the first and second Dropout layers is 0.2, and the number of classes is 1000 by default. The stride of all convolutional layers is set layer by layer, the padding mode is "same", and the activation function, except for the head fully connected layer, is the SiLU function.

[0036] The specific parameters of the network structure are shown in Table 2.

[0037] Table 2. Network Structure of YOLOv8n-cls In specific processing, the YOLOv8n backbone network extracts features from the input image through stepwise downsampling: the input is a 3×224×224 MFCC feature map. First, it passes through a convolutional layer with a stride of 2 to halve the spatial size to 112×112 and expand the number of channels to 64. Then, it passes through a second convolutional layer with a stride of 2 to reduce the size to 56×56 and expand the number of channels to 128. Subsequently, C2f modules and convolutional layers with a stride of 2 are stacked alternately to complete multi-stage feature learning at three scales: 28×28, 14×14, and 7×7. Within each C2f module, rich local frequency domain features are extracted through multi-branch Bottleneck residual stacking. Finally, the SPPF module aggregates multi-scale receptive field information through serial max pooling and outputs a 7×7 deep feature map with 1024 channels.

[0038] The neck and head networks further map deep features to the recognition results: the neck uses an adaptive average pooling layer to directly compress the 7×7 feature map into a 1×1×1024 global feature vector, preserving the global response in the channel dimension and significantly reducing the number of parameters. This vector is then input into the classification head, first flattened into a 1024-dimensional vector, and then passed sequentially through the first Dropout layer (dropout probability 0.2), a fully connected layer (1024→1280), a second Dropout layer (dropout probability 0.2), and a second fully connected layer (1280→number of categories), finally outputting the probability distribution of the drone and various environmental noises. The fully connected layer maps the global feature vector to a high-dimensional space through linear transformation to enhance class separability. The Dropout layer randomly drops some neurons during training to prevent overfitting. The final output multidimensional probability vector is normalized by the Softmax function, and the category corresponding to the maximum value is taken as the drone noise recognition result.

[0039] The expression for the YOLOv8n-cls lightweight network model is as follows: Let the input be (3 channels, size 224×224). YOLOv8n-cls as a whole can be regarded as a composite function. Mapping the input to a class probability vector ( (Number of categories) The mathematical meanings of each part are as follows: (1) Backbone network It is composed of several basic operations (convolution, batch normalization, SiLU activation, C2f module, SPPF module) in sequence: Where X is the input feature map, each... This represents a triple of "convolution + batch normalization + SiLU"; each It consists of multiple Bottleneck residual blocks stacked together, with cross-layer stitching and fusion; SPPF is the stitching and convolution after serial max pooling. The final output feature map. (2) Classification of neck The classification neck layer is an adaptive average pooling layer that compresses the spatial dimension to 1×1: Obtain the global vector .

[0040] (3-category head) The classification head will display the global feature vector. The forward propagation process, which maps the probability distributions of each category, is as follows: in, To flatten the operation, The tensor is converted into a 1024-dimensional vector; This is a randomly deactivated layer with a dropout probability of 0.2, used to prevent overfitting during training; The first fully connected layer maps the 1024-dimensional input to 1280 dimensions; The second fully connected layer maps the 1280-dimensional input to the number of categories. ; Let be the category score vector output by the fully connected layer, and let its be the . The component indicates that the input belongs to the first... The unnormalized score of the class. Ultimately, it will be... After processing with the Softmax normalization function, the probability distributions of each class are obtained: in, For input belonging to the first The probability of a class The sum of the exponents of the scores for all categories is used to transform the score vector into a probability distribution with a sum of 1. The category corresponding to the maximum probability is taken as the drone noise recognition result.

[0041] Specifically, the training process of the YOLOv8n-cls lightweight network model includes: During the training phase, the data loader reads preprocessed MFCC feature maps in batches from the `train` subdirectory under the root directory of the specified dataset. Each training sample is a pseudo-color image with a uniform size of 224×224 pixels. This image is obtained by extracting Mel-frequency cepstral coefficients from the original UAV or environmental noise audio signal through framing, windowing, and Fourier transform, then stacking them temporally to form a two-dimensional spectrogram. Finally, it is mapped to a three-channel color image through energy normalization and pseudo-color encoding. Its pixel matrix intuitively represents the energy distribution pattern of the noise signal in the time-frequency domain. The dataset covers UAV acoustic signals and ten typical urban environmental noise categories, including but not limited to traffic noise, crowd noise, and animal calls. The true category of each sample is directly encoded by the name of its subfolder, without the need for external label files. The model's forward output is a K-dimensional probability vector, where K corresponds to the total number of all categories. This vector is normalized by the Softmax activation function to obtain the predicted probability of each semantic category.

[0042] This invention uses YOLOv8 as the backbone network and maps the input feature map to the category probability space through multi-scale convolutional kernels. Forward prediction and the generated true labels based on folder names are compared using cross-entropy loss to calculate the error, and backpropagation iteratively updates the network parameters. This forces the convolutional kernels to adaptively capture the local texture and global envelope differences of different acoustic categories under fixed scales and multiple receptive fields, thereby establishing a robust mapping from time-frequency features to UAV and urban environmental noise categories. To adapt to different data scales and accuracy requirements, two optimal hyperparameter combinations are provided during training: (1) Balanced and robust type - AdamW optimizer with weight decay, initial learning rate set to 0.001, weight decay coefficient 5×10⁻ 4 With a dropout rate of 0.3, a batch size of 16, and a total of 30 training rounds with an early stopping patience value of 10, this configuration balances convergence speed and generalization performance on medium-sized datasets and can effectively suppress the risk of overfitting the model to specific noisy samples in the training set.

[0043] (2) High-precision, large-iteration type – adopts an SGD optimizer with Nesterov momentum, an initial learning rate of 0.01, and a cosine annealing scheduling strategy, with a weight decay coefficient of 1×10⁻ 4 With a dropout rate of 0.1%, batch size of 32, 60 training rounds, and an early stop patience value of 20, this configuration is suitable for tasks with a large number of categories or highly similar acoustic features between categories. It can fully explore the subtle differences between drone sound and various environmental noises, thereby further improving recognition accuracy. Users can flexibly choose the configuration based on the data scale, category complexity, and computing resources of the actual task to achieve the optimal balance between training efficiency and recognition accuracy.

[0044] Specifically, the present invention will be further illustrated below through embodiments: Comparative experiments were conducted between YOLOv8n-cls and SqueezeNet, MobileNetV2, and ResNet50 models. Precision, recall, specificity, and confusion matrix were used as evaluation metrics to assess the recognition performance and anti-interference capabilities of different models, and YOLOv8n-cls was determined to be the preferred model for UAV noise recognition.

[0045] In recognition tasks, accuracy alone often fails to fully reflect a model's performance, especially when there is an imbalance in the distribution of positive and negative samples or when focusing on a specific category. This paper focuses on drone recognition, a specific category within a multi-class classification task. To evaluate the model's performance in recognizing this category, a "one-vs-rest" evaluation strategy is adopted, defining the target category (drone) as positive and all other categories as negative. Under this setting, three widely used evaluation metrics in classification are selected: precision, recall, and specificity.

[0046] SqueezeNet, a lightweight convolutional neural network proposed by Forrest Iandola et al. in 2016, is innovatively based on its Fire module. This module first compresses the number of input channels using 1×1 convolutional layers to reduce computation, then uses a combination of 1×1 and 3×3 convolutional kernels for feature expansion and extraction. Through this "compression-then-expansion" design, SqueezeNet maintains classification accuracy similar to AlexNet while reducing the number of parameters to one-fiftieth of the latter. MobileNetV2, a lightweight network architecture proposed by Google in 2018, is innovatively based on its inverted residual block and linear bottleneck layer. Unlike traditional residual structures, MobileNetV2's inverted residual block first expands the low-dimensional input to a high-dimensional space using 1×1 convolutions, then uses depthwise separable convolutions for lightweight feature filtering, and finally compresses the features back to the low-dimensional bottleneck layer using 1×1 convolutions. This "expansion-then-compression" design significantly reduces computation while effectively preserving feature representation capabilities. ResNet50, a deep residual network proposed by HeKaiming et al. in 2015, contains 49 convolutional layers and 1 fully connected layer, with a total of approximately 25.5 million parameters. Its core innovation lies in the introduction of a residual connection mechanism, which effectively solves the gradient vanishing problem in deep network training by directly adding input features to output features (identity mapping). ResNet50 employs a bottleneck block design for each residual block—first, 1×1 convolutions are used to reduce dimensionality to decrease computation, then 3×3 convolutions are used to extract features, and finally 1×1 convolutions are used to restore dimensionality. Meanwhile, yolov8n-cls.pt, as the Nano version of the YOLOv8 series specifically designed for image classification, balances extremely high inference speed with moderate recognition accuracy, making it suitable for deployment on edge computing devices. The network's design structure is as follows... Figure 4 As shown in the figure. The structure diagrams of the three networks, SqueezeNet, MobileNetV2, and ResNet50, are respectively shown in the figure. Figure 5 , Figure 6 , Figure 7 As shown.

[0047] Specifically, the backbone of the YOLOv8n network utilizes the concept of cross-stage local networks and C2f modules. Unlike traditional image classification models, this model is derived from the YOLOv8 object detection framework, removing the detection head and multi-scale feature fusion components, retaining only the backbone network and classification head. The backbone consists of multiple Conv modules and C2f modules stacked alternately. A Conv module is a standard unit containing convolutional layers, SiLU activation functions, and batch normalization. See [link to specific structure] for details. Figure 5 The pinkish-nude portion. The core approach of the C2f module is to process the input features... Figure 1 It is divided into two parts: one part is directly transmitted, while the other part is processed step by step through multiple Bottleneck sub-layers before being pieced together, such as... Figure 5 As shown in the light green module, this design effectively reduces computational redundancy while ensuring gradient diversity. The network ultimately outputs the classification result through a Classify module, which consists of global average pooling and fully connected layers; the structure is shown below. Figure 5 The blue module section; the entire network has 29 layers, and the input image size is 224×224×3.

[0048] The pre-trained backbone network is responsible for extracting general visual feature representations, while the classification head performs customized reconstruction for 11 classification tasks. The entire model adopts an end-to-end fine-tuning approach, enabling the pre-trained knowledge to fully adapt to the target task and achieve better generalization performance on a limited training set of 8650 data points. After continuous parameter adjustments, the optimal hyperparameter configuration for model performance was finally selected. The main hyperparameters of each model are shown in Table 3 below: Table 3. Main hyperparameters of the model The three PyTorch models used the exact same training configuration. This ensures that performance differences between the models stem solely from their architectures and lays the foundation for a fair comparison with YOLOv8 later. A stable and moderate initial learning rate of 0.0001 was used, coupled with a StepLR decay strategy—the learning rate decreases to 0.1 times its original value every 5 training epochs—achieving a balance between fast initial convergence and fine-tuning in later stages. This unified configuration simplifies the control of experimental variables and makes the comparative results more convincing.

[0049] YOLOv8 employs a different training strategy than PyTorch models, primarily following the framework's official recommendations for classification tasks. The training epochs are set to 30, more than the previous 20, and include cosine annealing for learning rate scheduling and a 3-epoch warm-up mechanism. This ensures a steady increase in the learning rate initially, followed by a smooth decrease later, avoiding the loss value oscillations that can occur with step-wise decay. The batch size is set to 16 to balance memory usage and convergence stability. The optimizer chosen is AdamW, with weight decay set to 0.0005, providing decoupling regularization compared to standard Adam, effectively suppressing overfitting. Furthermore, a Dropout value of 0.3 and an early stopping mechanism with a patience value of 10 epochs form a double safety net. The former randomly discards some neurons during training, while the latter automatically terminates training if the validation set metrics do not improve over a prolonged period, preventing the model from overlearning on limited data. Overall, the configurations in Table 2 allow the model to achieve good performance with minimal human intervention. This was performed in a Python 3.12 environment; the specific library specifications used are shown in Table 4.

[0050] Table 4. Specifications of Python libraries used for training models Based on the previously defined evaluation index system, the performance of four network models—SqueezeNet, MobileNetV2, ResNet50, and YOLOv8n-cls—in the UAV recognition task was quantitatively analyzed. The index results obtained from a test set of 100 UAV noise samples are shown in Table 5 below. Table 5 Comparison of UAV noise recognition performance across different networks Before introducing specific indicators, let's first explain the four basic cases of binary classification results. Based on the correspondence between the model's predicted results and the true categories, the sample statistics can be divided into the following four categories: 1) True Positive (TP): The number of samples that are actually positive and the model predicts to be positive; 2) False Positive (FP): The number of samples that are actually negative but the model predicts as positive; 3) True Negative (TN): The number of samples that are actually negative and that the model predicts to be negative; 4) False Negative (FN): The number of samples that are actually positive but the model predicts as negative.

[0051] The four statistical measures mentioned above form the basis for calculating the evaluation indicators below. It should be noted that this experiment verifies the accuracy of drone noise recognition in complex environments, thus it is a multi-classification task. Under the multi-classification "one class versus the remaining classes" setting, the "positive class" specifically refers to the drone category of interest, while the "negative class" is the union of all other environmental noises besides drones. Based on the above statistical measures, precision, recall, and specificity can be calculated to evaluate the model's performance in identifying drone categories. The three selected indicators will be explained in detail below.

[0052] 1) Precision Precision, also known as accuracy, measures the proportion of samples that a model predicts to be positive, but which are actually positive. Its calculation formula is as follows: This represents the probability that the model correctly identifies the drone, and its value ranges from [0,1]. The higher this metric, the more reliable the positive class prediction and the fewer false positives the model generates in identifying drones.

[0053] 2) Recall Recall, also known as the detection rate, measures the proportion of all true positive samples correctly detected by the model. In other words, it represents the probability that drone noise is correctly identified as a drone by the model. Its calculation formula is as follows: The recall rate ranges from [0,1]. A higher recall rate indicates a stronger ability of the model to capture positive samples and fewer false negatives. For drone identification, a high recall rate means that very few real drones are missed. In this paper, we can treat this rate as the hit rate of drone noise, which is crucial for ensuring safety.

[0054] 3) Specificity Specificity measures the proportion of all true negative samples that are correctly classified as negative by the model. Its calculation formula is as follows: This represents the probability that the model will classify an input as "not a drone" when the input is not drone noise. The specificity value ranges from [0,1]. The higher this metric, the stronger the model's ability to identify negative samples, and the lower the probability of generating false alarms when drones do not actually exist.

[0055] In a multi-class classification setting, high specificity means that when the input sample is actually of another category, the model can correctly not classify it as a drone. Subtracting the specificity probability from 1 gives the scenario where the model classifies it as a "drone", representing a false alarm.

[0056] Therefore, the recall and specificity metrics needed further definition to better reflect the performance of drone recognition. The two metrics are the drone noise hit rate and false alarm rate, respectively, and their specific expressions are as follows: Based on the experimental data in Table 5, it can be observed that there are certain performance gradient differences in the horizontal comparison of the three classic convolutional neural network architectures: SqueezeNet, MobileNetV2, and ResNet50. Although SqueezeNet achieves a high degree of lightweighting, its hit rate is 0.860 and its false alarm rate is 0.041, indicating that it may lose some effective features under deep compression, resulting in significant missed detections and false alarms. In contrast, MobileNetV2, by introducing an inverse residual structure, improves the hit rate to 0.930 and reduces the false alarm rate to 0.024 while maintaining lightweighting, demonstrating a better balance between efficiency and performance. ResNet50, with its 50-layer residual connection structure, further explores the potential of deep networks, achieving a hit rate of 0.940 and a relatively low false alarm rate of 0.018, demonstrating superior performance among the classic CNN models in this UAV recognition task. However, YOLOv8n-cls outperformed all three CNN models across all three core metrics, with its hit rate soaring to 0.980 and its false alarm rate plummeting to 0.007. Specifically, YOLOv8n-cls reduced the false negative rate from 6% of ResNet50 to 2%, while lowering the false alarm rate by approximately 61%. This "high hit rate, low false alarm rate" characteristic is precisely what security monitoring scenarios urgently need. Considering the limited number of drone noise samples in the test set, a comparative analysis of the models' confusion matrices is also necessary to comprehensively compare the performance of the models in drone noise recognition.

[0057] and Figure 9 Compared to SqueezeNet, Figure 8The YOLOv8n-cls model demonstrates significant performance advantages. In 100 test samples, SqueezeNet correctly identified 86 drone sound signals, but many were misclassified as similar mechanical noises such as engine idling or drilling noise. Furthermore, its accuracy in other categories was low, with air conditioner sound recognition accuracy at only 40% and children's play sound recognition accuracy at only 71%. Children's play sounds were significantly confused with street music, dog barking, and other noises. In contrast, the YOLOv8n-cls model correctly identified 98 groups, with only a few samples misclassified as drilling noise. Its accuracy across all categories was significantly higher, with a smaller overall off-diagonal misclassification value. It particularly alleviated the confusion between children's play sounds and street music, exhibiting stronger noise robustness and drone target recognition stability.

[0058] and Figure 10 Compared to MobileNetV2, Figure 8 The YOLOv8n-cls model demonstrates a significant performance advantage. While MobileNetV2 achieves a 93% drone recognition accuracy, outperforming SqueezeNet, it still suffers from one misclassification as street music and five misclassifications as drilling sounds, indicating insufficient ability to distinguish industrial noise. YOLOv8n-cls, on the other hand, only had one minor misclassification as drilling sounds and no other cross-category confusion. Overall, MobileNetV2 exhibits more misclassifications for complex background noises such as children playing and street music, while YOLOv8n-cls achieves higher accuracy in these categories. In conclusion, YOLOv8n-cls significantly outperforms MobileNetV2 in drone noise recognition in low signal-to-noise ratio scenarios, making it better suited for complex real-world environments.

[0059] like Figure 11 As shown, while ResNet50 outperforms the first two models, it is still inferior to... Figure 8 The YOLOv8n-cls correctly identified 94 drone samples, while YOLOv8n-cls correctly identified 98. Regarding anti-interference capabilities, taking the easily confused category of car horns as an example, ResNet50 misclassified a total of 13 samples as either drones or sirens, while YOLOv8n-cls only had 3 minor cross-category misclassifications and did not misclassify them as drones. In terms of overall robustness, ResNet50's confusion matrix shows multiple light-colored distributions in the off-diagonal region, indicating that misclassification still spreads between different categories; YOLOv8n-cls's diagonal distribution is more concentrated and deeper, and its off-diagonal values ​​are generally extremely low, consistent with its extremely low false alarm rate of 0.007, demonstrating that YOLOv8n-cls is not only more accurate in identifying positive samples but also more decisive in eliminating negative samples.

[0060] The experimental results above demonstrate that YOLOV8n-cls, through its more advanced network structure, can more effectively extract the unique rotor acoustic features of UAVs from background noise, significantly reducing false alarms caused by environmental noise interference. Therefore, it is a more suitable model choice for real-world UAV acoustic signal recognition scenarios.

[0061] Example 2: To achieve the above objective, such as Figure 12 As shown, based on Embodiment 1, this invention discloses a drone noise recognition system based on a lightweight YOLO network, comprising: Data acquisition module 11 is used to acquire drone noise samples and urban background noise samples, and to perform sample processing based on drone noise samples and urban background noise samples to obtain drone noise dataset under complex noise environment. Data processing module 12 is used to convert one-dimensional time-domain audio signals into two-dimensional feature maps of UAV noise datasets under complex noise environments based on the preset Mel frequency cepstral coefficient MFCC feature extraction method, so as to obtain MFCC feature extraction map datasets. The recognition output module 13 is used to input the MFCC feature extraction map dataset into the pre-established YOLOv8n-cls lightweight network model for training, obtain the trained YOLOv8n-cls lightweight network model, acquire the drone noise data to be identified, input the drone noise data to be identified into the trained YOLOv8n-cls lightweight network model, and output the drone noise recognition result.

[0062] Based on the same inventive concept, this invention also provides a computer device, comprising: one or more processors, and a memory for storing one or more computer programs; the programs include program instructions, and the processor executes the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, used to implement one or more instructions, specifically for loading and executing one or more instructions stored in a computer storage medium to implement the above-described method.

[0063] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, performs the above-described method. This storage medium can be any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0064] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0065] The foregoing has shown and described the basic principles, main features, and advantages of this disclosure. Those skilled in the art should understand that this disclosure is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of this disclosure. Various changes and modifications can be made to this disclosure without departing from its spirit and scope, and all such changes and modifications fall within the scope of this disclosure as claimed.

Claims

1. A method for drone noise recognition based on a lightweight YOLO network, characterized in that, The method includes the following steps: We acquire drone noise samples and urban background noise samples, and perform sample processing based on the drone noise samples and urban background noise samples to obtain a drone noise dataset under complex noise environment. Based on the MFCC feature extraction method with preset Mel frequency cepstral coefficients, a one-dimensional time-domain audio signal of a UAV noise dataset under complex noise environment is converted into a two-dimensional feature map to obtain an MFCC feature extraction map dataset. The MFCC feature extraction map dataset is input into a pre-built YOLOv8n-cls lightweight network model for training, resulting in a trained YOLOv8n-cls lightweight network model. The noise data of the drone to be identified is then obtained and input into the trained YOLOv8n-cls lightweight network model, and the drone noise identification result is output.

2. The method for drone noise recognition based on a lightweight YOLO network according to claim 1, characterized in that, The Mel frequency cepstral coefficient (MFCC) feature extraction method includes pre-emphasis, framing and windowing, fast Fourier transform, Mel filter bank filtering, logarithmic operation, and discrete cosine transform.

3. The method for drone noise recognition based on a lightweight YOLO network according to claim 2, characterized in that, The pre-emphasis process includes: Raw signals within a UAV noise dataset under complex noise environments Through a first-order high-pass filter: in These are preset coefficients; The process of framing and windowing includes: The pre-emphasized signal is multiplied by a window function that smoothly transitions to zero at both ends. The expression for the window function is as follows: in Given a frame length, the windowed signal is: .

4. The method for drone noise recognition based on a lightweight YOLO network according to claim 3, characterized in that, The process of the Fast Fourier Transform is as follows: Perform an FFT on the windowed signal s[n] of each frame to transform it from the time domain to the frequency domain, and obtain the spectral distribution S[k] of that frame: Then calculate the power spectrum: ； The filtering process of the Mel filter bank is as follows: The power spectrum is passed through a set of Mel-scale triangular filters, with the center frequencies of each filter being equally spaced on the Mel scale, exhibiting a distribution characteristic of dense low frequencies and sparse high frequencies on the Hz scale. No. The frequency response of a triangular filter is defined as: in For the first The center frequency of each filter For frequency indexing.

5. The method for drone noise recognition based on a lightweight YOLO network according to claim 4, characterized in that, The process of the logarithmic operation is as follows: Take the natural logarithm of the output energy of each Mel filter: The process of discrete cosine transform is as follows: Log-log energy sequence By performing a Discrete Cosine Transform (DCT) and utilizing its energy concentration characteristics, the information output by the filter bank is compressed into low-dimensional coefficients. Simultaneously, the linear correlation between features in each dimension is removed, ultimately yielding the MFCC coefficients, as follows: in, Indicates the first Vermeer frequency cepstral coefficients; This represents the total number of filters within the Mel filter bank. The filter number; For the first The value obtained by performing a logarithmic operation on the output energy of a Mel filter; The dimension of the final output MFCC coefficients; This refers to the MFCC coefficient sequence number.

6. The method for drone noise recognition based on a lightweight YOLO network according to claim 1, characterized in that, The YOLOv8n-cls lightweight network model performs multi-scale feature learning through the backbone network, and combines the classification neck network and classification head network to output a multi-dimensional probability vector. After normalization by the Softmax function, the category corresponding to the maximum value is taken as the drone noise recognition result.

7. The method for drone noise recognition based on a lightweight YOLO network according to claim 6, characterized in that, The backbone network consists of a first convolutional module, a second convolutional module, four C2f modules, and one SPPF module cascaded sequentially. The classification neck network is an adaptive average pooling layer that reduces the spatial dimension of the feature map output by the backbone network. The classification head network includes a flattening layer, a first dropout layer, a first fully connected layer, a second dropout layer, and a second fully connected layer. The first and second fully connected layers map the global feature vector to a high-dimensional space through linear transformation to enhance class separability. The first and second dropout layers randomly drop some neurons during training to prevent overfitting, and finally output a multi-dimensional probability vector.

8. The method for drone noise recognition based on a lightweight YOLO network according to claim 7, characterized in that, The expression for the YOLOv8n-cls lightweight network model is as follows: Let the input be YOLOv8n-cls as a whole can be regarded as a composite function. Mapping the input to a class probability vector ,in, Number of categories; The mathematical meanings of each part are as follows: backbone network It is composed of convolution, batch normalization, SiLU activation, C2f module and SPPF module in sequence: Where X is the input feature map, each... Represents the triples of convolution with batch normalization and SiLU; each It consists of multiple Bottleneck residual blocks stacked together, with cross-layer stitching and fusion; SPPF is the concatenation and convolution after serial max pooling, and the final output is a feature map. ; Classification of neck For adaptive average pooling layers, the spatial dimension is compressed to 1×1: Obtain the global vector ; Category Header global feature vectors The forward propagation process, which maps the probability distributions of each category, is as follows: in, For flattening operation; This is a randomly deactivated layer with a dropout probability of 0.2; This is the first fully connected layer; This is the second fully connected layer; Let be the category score vector output by the fully connected layer, and let its be the . The component indicates that the input belongs to the first... The unnormalized score of the class will After processing with the Softmax normalization function, the probability distributions of each class are obtained: in, For input belonging to the first The probability of a class The sum of the scores for all categories.

9. A drone noise recognition system based on a lightweight YOLO network, employing the drone noise recognition method based on a lightweight YOLO network as described in any one of claims 1 to 8, characterized in that, include: The data acquisition module is used to acquire drone noise samples and urban background noise samples, and to perform sample processing based on the drone noise samples and urban background noise samples to obtain drone noise dataset under complex noise environment. The data processing module is used to convert one-dimensional time-domain audio signals into two-dimensional feature maps of UAV noise datasets in complex noise environments based on the preset Mel frequency cepstral coefficients (MFCC) feature extraction method, and obtain MFCC feature extraction map datasets. The recognition output module is used to input the MFCC feature extraction map dataset into the pre-established YOLOv8n-cls lightweight network model for training, obtain the trained YOLOv8n-cls lightweight network model, acquire the drone noise data to be identified, input the drone noise data to be identified into the trained YOLOv8n-cls lightweight network model, and output the drone noise recognition result.

10. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, The memory stores a computer program that can run on the processor. When the processor loads and executes the computer program, it employs a drone noise recognition method based on a lightweight YOLO network as described in any one of claims 1 to 8.

Images