Unmanned aerial vehicle image transmission signal detection and identification method based on improved YOLOv8

Abstract

The invention belongs to the technical field of wireless communication signal processing, and relates to an improved YOLOv8-based unmanned aerial vehicle image transmission signal detection and identification method, which comprises the following steps of: acquiring an unmanned aerial vehicle image transmission signal and interference data; generating a time-frequency graph through short-time Fourier transform, dimension reduction and gray level conversion, and marking the time-frequency graph; performing data enhancement by adopting frequency shift, noise addition and multi-signal fusion, and constructing an enhanced annotation set; a YOLOv8 model is improved, a convolution kernel of a CBS module is changed into a non-square to adapt to a signal form, a multi-branch variable receptive field unit is additionally arranged in a Bottleneck module to extract multi-scale features, and a detection model is obtained by using enhanced data training; a to-be-measured signal is preprocessed and input into the model, and the center frequency, the bandwidth and the unmanned aerial vehicle type of the signal are synchronously output. According to the invention, end-to-end signal feature automatic extraction and identification are realized, and the detection precision and robustness in a complex scene are improved.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication signal processing technology, specifically a method for detecting and recognizing UAV image transmission signals based on an improved YOLOv8. Background Technology

[0002] Currently, drone type identification is one of the key technologies in low-altitude security, countermeasure systems, and spectrum regulation. The mainstream drone type identification methods can be divided into radio frequency signal-based identification, radar-based identification, photoelectric / visual-based identification, and acoustic-based identification. Radio frequency signal-based identification is a passive detection method with a long operating range and has become the most mainstream approach.

[0003] With the development of drone technology, its signals are becoming increasingly complex and diverse, making accurate identification of drone signal types a challenge.

[0004] Currently, the more mature methods for identifying drone signal types include: The core process of traditional signal processing and feature engineering methods is as follows: 1) Signal preprocessing (down-conversion, filtering, etc.); 2) Feature extraction (time domain features, frequency domain features, modulation domain features, protocol features, etc.); 3) Classifier recognition (support vector machine, decision tree, random forest, K-nearest neighbor, etc.).

[0005] This approach, based on protocol analysis and behavioral fingerprinting, focuses more on the "behavioral patterns" of communication. The core idea is to intercept signals and analyze the communication protocol (frame structure, handshake process, retransmission mechanism, etc.) to extract unique identifiers or behavioral models, thereby identifying the type of UAV signal.

[0006] End-to-end recognition methods based on deep learning transform signals into time-frequency images (two-dimensional time-frequency images where intensity is represented by color) and then directly input them into deep learning networks for feature learning and classification. Typical deep learning methods use network models such as ResNet, VGG, and MobileNet.

[0007] However, methods based on traditional signal processing and feature engineering rely heavily on expert experience, requiring the manual design of complex features. Furthermore, when faced with new types of drones, previous features are no longer applicable and need to be redesigned. Methods based on protocol analysis and behavioral fingerprinting are only effective for drone types with known or reversible protocols. Once the protocol is encrypted or changed, the method becomes ineffective. Current mainstream deep learning-based end-to-end identification methods can only identify drone types and cannot output parameters such as center frequency and bandwidth. They also cannot adapt to scenarios with multiple drones. Summary of the Invention

[0008] To address the problems mentioned in the background section, this invention provides a method for detecting and recognizing UAV image transmission signals based on an improved YOLOv8, thereby solving the problem of low accuracy in detecting and recognizing UAV image transmission signals.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a method for detecting and recognizing UAV image transmission signals based on an improved YOLOv8, comprising the following steps: S1: Acquire raw data and interference data of training drone image transmission signals in multiple scenarios; S2: Process the raw data of the UAV image transmission signal for training according to the preset preprocessing logic to obtain grayscale image data for training, and mark the UAV image transmission signal area in the grayscale image data for training to generate initial annotation data; S3: Based on the interference data, the initial labeled data is subjected to scenario-based extension processing adapted to the signal transmission characteristics of the UAV to obtain enhanced labeled data; S4: Construct an improved YOLOv8 model. By adjusting the parameters and structure of the feature extraction module of the YOLOv8 model backbone network, adapt it to the morphological features of the UAV image transmission signal. Then, use the enhanced annotation data to train the improved YOLOv8 model to obtain the UAV image transmission signal detection and recognition model. S5: Acquire the radio frequency signal to be detected, process the radio frequency signal to be detected according to the preset preprocessing logic of S2 to obtain grayscale image data to be detected, input the grayscale image data to be detected into the UAV image transmission signal detection and recognition model, and output the center frequency, bandwidth and corresponding UAV type of the UAV image transmission signal.

[0010] Optionally, the raw data and interference data of the UAV image transmission signal used for training in step S1 include: Raw data of image transmission signals from training drones were collected from multiple dimensions, including drone attributes, signal parameters, transmission distance, and operating frequency band. Interference data is collected in the signal environment of the UAV, and the interference data is used to simulate electromagnetic interference scenarios in actual applications.

[0011] Optionally, the preset preprocessing logic in step S2 includes time-frequency conversion, dimensionality reduction, and grayscale conversion. The execution process of the preset preprocessing logic is as follows: The raw data of the training UAV image transmission signal is processed by time-frequency conversion to obtain time-frequency domain data; The time-frequency domain data is subjected to dimensionality reduction processing to reduce data redundancy; Convert the dimensionality-reduced time-frequency domain data into grayscale image data.

[0012] Optionally, the annotation process in step S2 is as follows: By using the annotation tool to select the area corresponding to the UAV image transmission signal in the training grayscale image data, the location range information of the area is marked to form the initial annotation data.

[0013] Optionally, step S3 includes: The original data of the training UAV image transmission signal corresponding to the initial labeled data is processed by frequency shift transformation to adapt to signal frequency fluctuation scenarios. The raw data of the training UAV image transmission signal corresponding to the initial labeled data is processed by noise addition to adapt to electromagnetic interference scenarios. A multi-signal fusion method is used to fuse the original data of the training UAV image transmission signal corresponding to the initial labeled data with the interference data to adapt to multi-target signal superposition scenarios. The signal after any of the above processing is processed according to the preset preprocessing logic of S2, and the region information corresponding to the processed signal is labeled to generate enhanced labeling data.

[0014] Optionally, adjusting the parameters and structure of the feature extraction module of the YOLOv8 model backbone network in step S4 includes: The convolution kernel shape parameters of the CBS module in the YOLOv8 model backbone network were adjusted, and a non-square convolution kernel was used to replace the original square convolution kernel to adapt to the elongated shape of the UAV image transmission signal. The CBS module consists of convolutional layers, batch normalization layers, and activation function layers.

[0015] Optionally, adjusting the feature extraction module structure of the YOLOv8 model backbone network in step S4 further includes: The structure of the Bottleneck module in the backbone network of the YOLOv8 model is optimized by adding a variable receptive field unit to the Bottleneck module. The variable receptive field unit includes multiple feature extraction branches with different receptive field sizes. By fusing the output features of the multiple branches, it can adapt to UAV image transmission signals with different parameters. The fused features are then superimposed on the input data of the Bottleneck module to complete feature enhancement.

[0016] Optionally, the training process in step S4 includes: The enhanced annotation data is divided into a training set, a validation set, and a test set; Set the model training parameters, select a loss function suitable for the object detection task, and iteratively train the improved YOLOv8 model; After each round of training, the model performance is evaluated using the validation set, and training stops when the model performance meets the preset stability conditions. The performance of the model after training was stopped was verified using the test set. Once the verification was successful, the UAV image transmission signal detection and recognition model was obtained.

[0017] Optionally, the time-frequency conversion process employs a short-time Fourier transform, using a window function to segment the signal and preserve the joint time-frequency characteristics of the signal.

[0018] Optionally, in the multi-signal fusion method, the fusion ratio of the original data of the training UAV image transmission signal and the interference data is within a preset range, and the value range of the preset range is 0.2 to 0.8.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention combines receptive field theory with the elongated shape of UAV image transmission signals to specifically improve the feature extraction module of the YOLOv8 model's backbone network. By optimizing the adaptability and accuracy of feature extraction, it enhances signal detection and recognition performance. The underlying principle is that after a UAV image transmission signal is converted into a two-dimensional time-frequency graph via short-time Fourier transform, it exhibits a significant elongated shape. The original YOLOv8 model's CBS module uses a 3×3 square convolution kernel, resulting in a receptive field aspect ratio of 1 during feature extraction. This leads to feature redundancy due to the inclusion of a large amount of irrelevant background information and insufficient adaptability to elongated signals with varying bandwidths. Furthermore, the single receptive field of the Bottleneck module cannot comprehensively capture the core features of signals with different parameters. Therefore, this invention replaces the square convolution kernel of the CBS module with a non-square convolution kernel, thus improving the receptive field shape and elongated shape. Signal matching reduces feature redundancy caused by background interference. A variable receptive field unit is added to the Bottleneck module, and features are extracted in parallel through three CBS sub-modules using different sized strip convolution kernels to cover the feature capture requirements of signals with different bandwidths. The three features are then concatenated in the channel dimension and fused through 1×1 convolution, and superimposed with the module input data to enhance effective features. Combined with scenario-specific extensions designed for UAV signals, such as frequency shift transformation, noise reduction, and multi-signal fusion, as well as preprocessing flows for time-frequency conversion, dimensionality reduction, and grayscale conversion, the improved model can automatically extract the core features of UAV image transmission signals in different scenarios without the need for manual design of complex features. It also achieves synchronous output of signal center frequency, bandwidth, and UAV type, maintaining high recognition accuracy even in scenarios with low signal-to-noise ratio, multiple interferences, and multiple UAVs overlapping. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall method steps of the present invention; Figure 2This is a structural diagram of the Bottleneck module in this invention; Figure 3 This is a structural diagram of the Bottleneck module based on the variable receptive field improvement in this invention; Detailed Implementation 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.

[0021] like Figures 1 to 3 As shown, this invention provides a method for detecting and recognizing UAV image transmission signals based on an improved YOLOv8, comprising the following steps: S1: Acquire raw data and interference data of training drone image transmission signals in multiple scenarios; S2: Process the raw data of the UAV image transmission signal for training according to the preset preprocessing logic to obtain grayscale image data for training, and mark the UAV image transmission signal area in the grayscale image data for training to generate initial annotation data; S3: Based on the interference data, the initial labeled data is subjected to scenario-based extension processing adapted to the signal transmission characteristics of the UAV to obtain enhanced labeled data; S4: Construct an improved YOLOv8 model. By adjusting the parameters and structure of the feature extraction module of the YOLOv8 model backbone network, adapt it to the morphological features of the UAV image transmission signal. Then, use the enhanced annotation data to train the improved YOLOv8 model to obtain the UAV image transmission signal detection and recognition model. S5: Acquire the radio frequency signal to be detected, process the radio frequency signal to be detected according to the preset preprocessing logic of S2 to obtain grayscale image data to be detected, input the grayscale image data to be detected into the UAV image transmission signal detection and recognition model, and output the center frequency, bandwidth and corresponding UAV type of the UAV image transmission signal.

[0022] Specifically, this method follows a complete workflow of "multi-scenario data acquisition - signal preprocessing and annotation - data scenario-based expansion - model improvement and training - actual signal detection and recognition": First, raw data of UAV image transmission signals for training is collected from four dimensions: UAV attributes, signal parameters, transmission conditions, and operating frequency band. Simultaneously, electromagnetic interference data from the UAV environment is collected to construct a basic dataset. Then, the one-dimensional raw signal is mapped to the time-frequency joint domain through time-frequency conversion. After dimensionality reduction to remove redundant information, it is converted into grayscale data. Annotation tools are used to accurately annotate the UAV image transmission signal areas in the grayscale image, generating initial annotation data. Subsequently, based on the collected interference data, the raw signal corresponding to the initial annotation data is scenario-based expanded using three methods: frequency shift transformation, noise addition processing, and multi-signal fusion. The signal undergoes a reprocessing and annotation process to obtain enhanced annotation data. Next, considering the elongated shape of the UAV image transmission signal, the backbone network feature extraction module of the YOLOv8 model is improved. The convolution kernel shape of the CBS module is adjusted to match the signal shape. A variable receptive field unit is added to the Bottleneck module, and effective features are enhanced through multi-branch feature fusion. The enhanced annotation data is used to divide the training, validation, and test sets. Training parameters and loss functions adapted to the target detection task are set to complete model training, resulting in a UAV image transmission signal detection and recognition model. Finally, the radio frequency signal to be detected is acquired and converted into grayscale data according to the same preprocessing logic as the training data. This grayscale data is then input into the trained detection and recognition model, simultaneously outputting the center frequency, bandwidth, and UAV type of the UAV image transmission signal.

[0023] The raw data and interference data of the UAV image transmission signal used for training in step S1 include: Raw data of image transmission signals from training drones were collected from multiple dimensions, including drone attributes, signal parameters, transmission distance, and operating frequency band. Specifically, the data acquisition dimensions and specific data acquisition targets are as follows: The data acquisition scheme is designed strictly according to four core dimensions: UAV attributes, signal parameters, transmission distance, and operating frequency band. The data acquisition targets for each dimension are selected from typical scenarios in actual applications. Drone attributes: Mainstream consumer and industrial drone models were selected, including Hubbleson, DJI Mini, DJI Mavic 3, DJI Mavic 4, Autel, ZOOM 909, and Skylink, covering drone types of different brands and performance levels, to ensure that the model can learn the characteristic differences of image transmission signals from different manufacturers.

[0024] Signal parameters: Focusing on the key parameter of signal bandwidth, three typical bandwidths are selected: 10M (narrowband), 20M (medium broadband), and 40M (broadband) to correspond to different transmission rate requirements of UAV image transmission signals and cover common signal bandwidth ranges.

[0025] Transmission distance: Five gradient distances are set: 50m (short distance, such as indoor testing scenarios), 300m (short to medium distance, such as small-site operations), 1000m (medium distance, such as regular outdoor operations), 3000m (medium to long distance, such as professional operations), and 6000m (long distance, such as extreme transmission scenarios) to simulate signal attenuation at different usage distances.

[0026] Operating frequency bands: The 2.4G and 5.8G ISM bands, which are the most commonly used ISM bands for drones, are selected. These two bands are the mainstream communication bands for the vast majority of consumer and industrial drones, ensuring that the collected data has broad applicability.

[0027] Data Acquisition Equipment and Parameter Settings: A professional radio frequency signal acquisition instrument is used for data acquisition. The equipment parameters are set as follows: the sampling rate is set to 200MHz to ensure that the frequency components of the UAV image transmission signal can be completely acquired; the sampling accuracy is 16bit to ensure the resolution of signal acquisition and reduce quantization error; the acquisition mode is set to continuous acquisition to avoid signal loss.

[0028] Data Acquisition Method and Data Storage: During data acquisition, the UAV is placed in different motion states such as hovering, constant speed flight, turning, and acceleration / deceleration. The acquisition time for each state is no less than 30 minutes to ensure coverage of signal changes during the dynamic operation of the UAV. The acquired raw data is stored in IQ data format, which can completely preserve the in-phase and quadrature components of the signal, thereby preserving the core features of the signal such as amplitude and phase, providing a complete signal foundation for subsequent preprocessing.

[0029] Interference data is collected in the signal environment of the UAV, and the interference data is used to simulate electromagnetic interference scenarios in actual applications.

[0030] Furthermore, regarding the selection of the data collection environment: two typical electromagnetic environments for drone activities were selected for interference data collection, namely an indoor laboratory environment (with a single type of interference, mainly equipment noise and mains power interference) and an open area in the suburbs (with a complex type of interference, including environmental electromagnetic noise, distant WiFi signals, Bluetooth signals, and mobile communication signals (4G / 5G)). This ensures that the interference data can cover electromagnetic scenarios of varying complexity.

[0031] Interference type and data collection duration: The types of interference to be collected include environmental electromagnetic noise (naturally occurring electromagnetic radiation), WiFi signals (2.4G / 5G band), Bluetooth signals (2.4G band), and mobile communication signals (4G / 5G band). These are the most common types of electromagnetic interference encountered in the actual application of drones. The data collection duration for each environment is no less than 60 minutes to ensure that the collected interference data has sufficient diversity and representativeness.

[0032] Data Acquisition Equipment and Data Storage: The same radio frequency signal acquisition instrument as that used for acquiring UAV image transmission signals is adopted to maintain consistent parameters such as sampling rate (200MHz) and sampling accuracy (16bit) to ensure that the format and acquisition standard of interference data and target signal data are consistent. The acquired interference data is also stored in IQ data format to facilitate subsequent fusion processing with UAV image transmission signals.

[0033] The preset preprocessing logic in step S2 includes time-frequency conversion, dimensionality reduction, and grayscale conversion. The execution process of the preset preprocessing logic is as follows: The raw data of the training UAV image transmission signal is processed by time-frequency conversion to obtain time-frequency domain data; The time-frequency domain data is subjected to dimensionality reduction processing to reduce data redundancy; Convert the dimensionality-reduced time-frequency domain data into grayscale image data.

[0034] Specifically, time-frequency conversion processing: The acquired raw signal is IQ data, represented as a complex sequence. ,in For in-phase components, These are orthogonal components. To obtain the joint time-frequency distribution of the signal, a Short-Time Fourier Transform (STFT) is used. The signal is processed in frames, with each frame having a length of... , for the The frame signal is windowed and then subjected to Fast Fourier Transform (FFT): ; in, For window functions, this embodiment uses a Hamming window, whose expression is: ; in The result Let be a two-dimensional matrix, representing the first... The time slice in the first The signal amplitude at each frequency point.

[0035] Going a step further, dimensionality reduction: To reduce the amount of data, the time-frequency matrix... Compression is performed along both time and frequency dimensions. The compression is set at each [time dimension]. The average of the summations of the points is calculated in the frequency dimension. Sum and average the points. The resulting dimension-reduced matrix. The following calculations were performed: ; This operation divides the original matrix into multiple parts of size . The data is divided into blocks, and the sum of the elements within the block is used to represent the block, thereby significantly reducing the data dimensionality.

[0036] Furthermore, grayscale conversion processing: The dimensionality-reduced data matrix Convert to grayscale. Grayscale images have pixel values ​​ranging from 0 to 255. The conversion process is achieved through linear normalization and rounding down: ; in, and They are matrices The minimum and maximum values ​​in This indicates a floor function (rounding down). The final result is... This refers to standard grayscale image data, where the brightness level corresponds to the strength of the signal energy.

[0037] The annotation process in step S2 is as follows: By using the annotation tool to select the area corresponding to the UAV image transmission signal in the training grayscale image data, the location range information of the area is marked to form the initial annotation data.

[0038] Specifically, the objects and tools used for annotation: The annotation target is the grayscale image output in step S2. In this image, the UAV image transmission signal is represented as a bright, elongated stripe region with a certain duration and bandwidth. This embodiment uses the open-source image annotation tool LabelMe for interactive manual annotation. This tool supports drawing polygons or rectangles on the image to identify target areas.

[0039] Annotation operation process: The annotator loads a grayscale image and observes its brightness distribution. They identify the energy concentration area formed by the UAV image transmission signal (i.e., a band significantly brighter than the background noise). Using the LabelMe tool's rectangle drawing function, they draw a rectangle that completely covers the signal energy band. The term "box selection" refers to this rectangle drawing operation.

[0040] Labeling information (location range information): For each drawn rectangle, its precise "positional range information" needs to be recorded. This information is usually automatically generated and stored in the annotation system, and mainly includes: Bounding box coordinates: With the top-left corner of the image as the origin (0,0), record the coordinates of the top-left vertex (x_min, y_min) and the bottom-right vertex (x_max, y_max) of the bounding box. The x-axis corresponds to the time dimension (i.e., the STFT and the time index k or m after dimensionality reduction), and the y-axis corresponds to the frequency dimension (i.e., the frequency index).

[0041] Category Label: Assign a text label to each bounding box, which corresponds to the drone model to which the signal belongs (e.g., "DJI_Mini4", "Autel_Evo_2", etc.). This label is used by the model to learn and identify drone types.

[0042] Annotated data generation and format: After selecting and labeling all signal regions in a single image, the LabelMe tool generates a JSON annotation file with the same name as the image file. This file structure includes the image path, image dimensions, and a list (shapes) containing information about all bounding boxes. Each element in the list corresponds to a bounding box and includes key fields such as label (category label) and points (the set of coordinates of the top-left and bottom-right points of the rectangle). All images and their corresponding JSON annotation files together constitute the "initial annotation data."

[0043] Labeling standards and quality: To ensure annotation quality, the bounding boxes should closely fit the visual boundaries of the signal energy region, avoiding the inclusion of excessive irrelevant background or omission of the effective signal portion. Annotation work on the same dataset must follow a consistent standard. All annotation results are recommended to undergo cross-validation or sampling review to ensure consistency and accuracy, providing reliable ground truth values ​​for subsequent model training.

[0044] Step S3 includes: The original data of the training UAV image transmission signal corresponding to the initial labeled data is processed by frequency shift transformation to adapt to signal frequency fluctuation scenarios. The raw data of the training UAV image transmission signal corresponding to the initial labeled data is processed by noise addition to adapt to electromagnetic interference scenarios. A multi-signal fusion method is used to fuse the original data of the training UAV image transmission signal corresponding to the initial labeled data with the interference data to adapt to multi-target signal superposition scenarios. The signal after any of the above processing is processed according to the preset preprocessing logic of S2, and the region information corresponding to the processed signal is labeled to generate enhanced labeling data.

[0045] Specifically, frequency shift transformation processing is used to simulate carrier frequency shifts that may occur in actual UAV image transmission signals or scenarios where they operate on different channels. This processing directly modifies the original IQ signal sequence corresponding to the initial annotation data. The operation is then carried out. Specifically, a controllable frequency offset is introduced in the digital domain. (Unit: Hz), spectrum shifting is achieved through complex exponential multiplication. The transformation formula is: ; in, The sampling rate of the signal (unit: Hz). This is the sampling point index. Frequency offset value. The signal is randomly selected within a certain range, for example, uniformly distributed within [-Δf, +Δf], where Δf is set based on the maximum possible frequency fluctuation. After this transformation, the signal's position on the time-frequency graph will be shifted along the frequency axis. For the new signal after the shift... The preprocessing steps in step S2 (time-frequency conversion, dimensionality reduction, and grayscale conversion) need to be repeated to generate a new grayscale image. Since the signal shape remains unchanged but its position has shifted, the position range information of its bounding box (especially the frequency axis coordinates) needs to be adjusted based on the frequency offset. The preprocessing parameters are calculated and updated accordingly to generate new labeled data.

[0046] Noise addition is used to enhance the model's recognition ability in environments with low signal-to-noise ratios or background electromagnetic noise. This processing also applies to the original IQ signals. Interference is simulated by adding random noise. Specifically, the signal is mixed with a sequence of random noise that follows a Gaussian distribution. Add: ; Among them, noise sequence The mean is 0, and the variance is 0. The target signal-to-noise ratio (SNR) is calculated based on the desired SNR. The target SNR value can be randomly selected within a preset range (e.g., from 0dB to 20dB) to cover noise scenarios of varying intensities. Adding noise increases the background noise level in the signal time-frequency plot, potentially reducing the clarity of the signal bars. This applies to the noisy signal... Repeat step S2 for preprocessing to obtain a new grayscale image. Since noise addition does not change the core location and shape of the signal, the position range information of its bounding box usually remains unchanged, but the type and parameters of this enhancement need to be recorded in the annotation data.

[0047] Multi-signal fusion processing is used to simulate real-world environments with multiple signal sources, especially scenarios where multiple drones operate simultaneously or their signals are mixed with background interference. This processing transforms a raw dataset containing drone image transmission signals... (The signal from the initial labeled data) and a set of pure interference data The interference data (collected in step S1) is linearly superimposed. The fusion formula is: ; in, The fusion ratio coefficient is limited to a preset range, for example... . Values ​​are randomly selected within this range to control the relative intensity ratio between the target signal and the interference signal. This process generates a synthetic signal. It includes both the target drone's signal components and complex environmental interference background. For After re-performing the preprocessing in step S2 to generate the grayscale image, the area where the target UAV signal is located in the new image (whose location may become less obvious due to interference) needs to be re-examined and labeled. During labeling, the original signal... The corresponding bounding box positions are used as a reference, and fine-tuned based on the actual performance of the synthesized image to ensure that the bounding boxes accurately cover the energy concentration areas of the fused signal. Similarly, this operation will generate new annotation data.

[0048] After any of the above-mentioned scenario-based extension processes are completed, new "signal-image-annotation" data pairs will be generated. All the new data generated through frequency shifting, noise addition, multi-signal fusion, etc., are merged with the original initial annotation data to form the "enhanced annotation data." This dataset has been significantly enhanced in terms of scale, diversity, and scenario complexity, laying a data foundation for subsequent training of detection and recognition models that can adapt to complex real-world environments.

[0049] Specifically, through the three types of scenario-based extension processing described above, the scale of the original initial annotation data is significantly expanded. In this embodiment, the initial annotation data contains 5,000 labeled grayscale images. After frequency shift enhancement, each original image is processed using three different random frequency offset values ​​to generate 15,000 new images; after noise enhancement, Gaussian noise is added to each original image at four different signal-to-noise ratio levels (0dB, 5dB, 10dB, 15dB) to generate 20,000 new images; after multi-signal fusion enhancement, each UAV signal data and interference data are fused in a random ratio to generate 10,000 new images. After removing some invalid samples, the final enhanced annotation data contains approximately 50,000 grayscale images and their corresponding annotation files, with the training set, validation set, and test set randomly divided in a 7:2:1 ratio. The enhanced dataset is large and diverse enough to support the improved YOLOv8 model in learning the essential features of UAV image transmission signals in different scenarios, effectively improving the model's generalization ability.

[0050] Step S4 involves adjusting the parameters and structure of the feature extraction module of the YOLOv8 model backbone network, including: The convolution kernel shape parameters of the CBS module in the YOLOv8 model backbone network were adjusted, and a non-square convolution kernel was used to replace the original square convolution kernel to adapt to the elongated shape of the UAV image transmission signal. The CBS module consists of convolutional layers, batch normalization layers, and activation function layers.

[0051] Specifically, when constructing the improved YOLOv8 model, a crucial adjustment was made to the convolution kernel shape of the CBS module, the fundamental feature extraction unit in the model's backbone network, to address the characteristic of UAV image transmission signals exhibiting a significantly elongated shape in the time-frequency graph. The original CBS module used a square convolution kernel for feature extraction, with a receptive field aspect ratio of 1. However, the aspect ratio of UAV image transmission signals (even the widest 40MHz bandwidth signal) in the grayscale image generated after preprocessing in step S2 is much smaller than 1 (e.g., possibly as low as 0.15625 or less). Using a square receptive field with an aspect ratio of 1 to capture such a narrow signal region results in the extracted features containing a large amount of irrelevant background information, causing feature redundancy and making the model overly sensitive to background interference.

[0052] To more accurately adapt to signal patterns, this invention replaces the default square convolutional kernel in the CBS module with a non-square convolutional kernel. Specifically, the kernel size used in the original CBS module is typically [size missing]. (This indicates that the convolution kernel height is 3 and the width is 3). After improvement, replace it with dimensions such as... A non-square convolution kernel with a height of 3 and a width of 2. The mathematical essence of this change is to adjust the shape of the receptive field during the convolution operation.

[0053] A CBS module typically consists of three sequentially connected sub-layers: a convolutional layer (Conv), followed by a batch normalization (BN) layer, and finally an activation function layer (typically using the SiLU activation function in this embodiment). Improvements focus on modifying the kernel size parameters of this convolutional layer.

[0054] This change in the shape of the convolutional kernel directly affects the size of the receptive field of the feature map. Assuming the network contains... A series of CBS modules (assuming step size and other parameters are set as usual). When using the original When convolution kernel, the first The receptive field size corresponding to a single pixel in the output feature map of each module is approximately calculated as follows: Its aspect ratio is 1. When replaced with After the convolution kernel, the first The receptive field size of the output feature points of each module becomes The improved receptive field maintains the same coverage in the frequency dimension (corresponding to the image height direction) as the original design, while its coverage in the time dimension (corresponding to the image width direction) is proportionally reduced as the convolution kernel width decreases. This makes the overall shape of the receptive field closer to a rectangle, with its width dimension decreasing relative to its height dimension, thus better matching the aspect ratio of the elongated shape of the UAV image transmission signal.

[0055] By adjusting the convolution kernel of the CBS module from a square to a non-square shape, the model can better fit the receptive field shape to the geometric characteristics of the target signal in the early stages of feature extraction. This helps to focus more on the main signal region during feature extraction, reducing invalid sampling of background noise on both sides of the signal, thereby reducing feature redundancy at the source, improving the signal-to-noise ratio of feature representation and adaptability to elongated signal shapes, and laying a better feature foundation for subsequent accurate detection and recognition.

[0056] The step S4, which involves adjusting the feature extraction module structure of the YOLOv8 model backbone network, also includes: The structure of the Bottleneck module in the backbone network of the YOLOv8 model is optimized by adding a variable receptive field unit to the Bottleneck module. The variable receptive field unit includes multiple feature extraction branches with different receptive field sizes. By fusing the output features of the multiple branches, it can adapt to UAV image transmission signals with different parameters. The fused features are then superimposed on the input data of the Bottleneck module to complete feature enhancement.

[0057] Specifically, based on the adjustment of the convolutional kernel shape of the CBS module, this invention further optimizes the structure of the deeper feature extraction unit in the YOLOv8 backbone network—the Bottleneck module—to solve the problem that a single fixed receptive field is insufficient to comprehensively capture the core features of UAV image transmission signals with different bandwidths and shapes. The original Bottleneck module typically consists of two cascaded CBS modules (using k=3×3, s=1×1 convolutional kernels), and the processed features are quickly concatenated (superimposed) with the module's input data. This structure provides a relatively fixed receptive field size and shape in the time-frequency plane.

[0058] To enhance the model's adaptability to diverse signal features, this embodiment adds a variable receptive field unit within the Bottleneck module. This unit employs a parallel branch structure, designed to simultaneously capture signal features at different scales. The specific improved structure is as follows: the feature data input to the Bottleneck module is simultaneously fed into three independent feature extraction branches for processing. Each branch contains two cascaded CBS modules, but the convolutional kernel sizes used in the CBS modules of these three branches differ, thus constructing different receptive fields.

[0059] The first path uses a CBS module with a kernel size of k=3×1 (height 3, width 1). This extremely narrow kernel has a certain receptive field in the frequency direction (height) but a very small receptive field in the time direction (width), making it particularly suitable for capturing signal components with obvious frequency characteristics and steep or narrow temporal variations. The second path uses a CBS module with a kernel size of k=5×1, providing a medium-width temporal receptive field, suitable for capturing the time-frequency profile of most typical UAV image transmission signals. The third path uses a CBS module with a kernel size of k=7×1, providing a larger temporal receptive field, which helps capture signal patterns with wide temporal broadening or some temporal domain extension. The stride of the convolution in all branches is set to s=1×1 to ensure that the spatial resolution is not reduced.

[0060] After processing the input features separately, the three branches output feature maps under three different receptive fields. Next, these feature maps from different branches are concatenated along the channel dimension. Assuming each branch's output feature map has C channels, the concatenated feature map has 3C channels. This operation aggregates multi-scale feature information.

[0061] Subsequently, to fuse this multi-scale information and control the number of channels, a 1×1 convolution operation is applied to the concatenated features. This 1×1 convolution has a kernel size of k=1×1 and s=1×1. Its main function is to perform cross-channel information interaction and feature fusion, and it can adjust the number of channels back to match the output of a single branch or meet the input requirements of the next layer. At this point, the variable receptive field unit has completed its multi-scale feature extraction and fusion process.

[0062] Finally, the features processed and fused by the variable receptive field unit are added element-wise (superimposed) to the original data input to this Bottleneck module (i.e., the data on the shortcut connection path). This residual connection structure helps alleviate the gradient vanishing problem in deep network training and ensures that the original basic feature information is preserved while introducing new features, thus completing an effective round of feature enhancement.

[0063] Through the above structural improvements, the improved Bottleneck module can adaptively extract and fuse features optimized for UAV image transmission signals with different parameters such as narrowband, medium-bandwidth, and wideband, enhancing the model's robustness to signal scale changes, thereby improving the overall performance of the detection and recognition system in complex and ever-changing scenarios.

[0064] The training process in step S4 includes: The enhanced annotation data is divided into a training set, a validation set, and a test set; Set the model training parameters, select a loss function suitable for the object detection task, and iteratively train the improved YOLOv8 model; After each round of training, the model performance is evaluated using the validation set, and training stops when the model performance meets the preset stability conditions. The performance of the model after training was stopped was verified using the test set. Once the verification was successful, the UAV image transmission signal detection and recognition model was obtained.

[0065] Specifically, model training is a systematic iterative optimization process aimed at enabling the improved YOLOv8 model to learn the ability to detect and identify UAV image transmission signals from time-frequency grayscale images. First, the augmented label data generated in step S3 is treated as a single dataset and randomly divided into three mutually exclusive subsets according to a preset ratio: a training set, a validation set, and a test set. A ratio of 7:2:1 or similar is typically used to ensure that the training set has sufficient data for model learning, the validation set is used for parameter tuning and monitoring the training process, and the test set is used to ultimately evaluate the model's generalization performance. During the partitioning, it is essential to ensure that each subset contains data from different UAV models, bandwidths, signal-to-noise ratios, and augmentation types to guarantee the comprehensiveness of the evaluation.

[0066] Next, configure the hyperparameters for model training. Key parameters include: optimizer (usually Adam or SGD), learning rate (setting the initial value and possible learning rate decay strategies, such as cosine annealing), batch size (determined based on GPU memory, e.g., 16 or 32), and the total number of training epochs. For object detection tasks, a suitable loss function needs to be selected. YOLOv8's loss function typically consists of several parts, including bounding box regression loss (such as CIoU loss), classification loss (such as binary cross-entropy), and possibly object confidence loss. The combination of these loss functions guides the model to learn how to accurately predict the location (the box corresponding to the center frequency and bandwidth) and category (drone type) of the target.

[0067] After training begins, in each epoch, the model uses the training set data for forward propagation to calculate predicted values, calculates the error between the predicted values ​​and the ground truth labels using a loss function, and then updates the model parameters (weights and biases) using backpropagation. After one epoch of training, the performance of the current model is evaluated using a validation set. Evaluation metrics include, but are not limited to, mean average precision (mAP), precision for each category, and recall. The training process continues for multiple epochs, monitoring performance metrics on the validation set.

[0068] The preset stopping conditions (i.e., the model performance meets the preset stability conditions) are usually set based on the performance of the validation set. A common strategy is to use early stopping: if the validation set loss no longer decreases within several consecutive epochs (e.g., 10 or 20 epochs), or the mAP metric no longer shows a significant improvement, then the model is considered to have been sufficiently trained or may have begun to overfit. At this point, training is stopped, and the checkpoint of the model with the best validation set performance is saved.

[0069] After training stops, the model cannot be used directly. Final performance validation must be performed using a test set that has not participated in any training or parameter tuning process. Run the trained model on the test set and calculate its mAP, precision, recall, and other metrics. If these metrics reach the preset application performance thresholds (e.g., mAP > 0.85), the model is considered validated. The resulting model is then a UAV image transmission signal detection and recognition model that can be used for practical signal detection and recognition. If the standards are not met, it is necessary to return to the inspection data, model structure, or training process, make adjustments, and retrain.

[0070] Specifically, the model performance evaluation uses the mean accuracy (mAP), a common indicator in the target detection field, as the core metric. This metric comprehensively considers the overall performance of precision and recall at different confidence thresholds. In this embodiment, the threshold for model validation is set to mAP > 0.85. This threshold is determined based on the high reliability requirements for UAV signal detection in actual low-altitude security scenarios: achieving this level of accuracy on the test set ensures that the model's detection rate and recognition accuracy for UAV image transmission signals meet the actual system deployment requirements in most complex electromagnetic environments. In addition to mAP, precision and recall for each category (i.e., different UAV models) are also examined simultaneously, requiring them to be no less than 0.80 to avoid blind spots or excessively high false positives in specific categories. These evaluation criteria collectively ensure that the final UAV image transmission signal detection and recognition model possesses high accuracy, strong robustness, and good category coverage.

[0071] The time-frequency conversion process employs a short-time Fourier transform, using a window function to segment the signal and preserve its joint time-frequency characteristics.

[0072] Specifically, the core of time-frequency conversion processing is to use short-time Fourier transform. This method uses a sliding window function to extract the signal and performs Fourier transform on the signal segment within each window, thereby obtaining the local spectrum of the signal frequency components as a function of time, effectively preserving the time-frequency joint characteristics of non-stationary signals.

[0073] In practice, for the collected IQ data sequences First, determine the parameters of the STFT: window function type, window length. and the number of overlapping points between adjacent windows. In this embodiment, a Hamming window is preferred as the window function, the expression of which has been given above. Window length The choice of overlap point requires a trade-off between time resolution and frequency resolution, and is usually determined based on the characteristics of the signal and the sampling rate, for example, set to an integer fraction of the number of sampling points. Typically smaller than the window length This is to ensure the continuity of time-frequency representation.

[0074] The processing procedure is as follows: The window function... The center of the signal is aligned with the starting point of the signal sequence, and a truncation length of [length missing] is extracted. signal segment Performing a dot-multiplication and windowing operation on this signal segment yields... Then perform the following steps on the results. Point Fast Fourier Transform (FFT) is used to obtain the spectrum estimate at that time. Then, slide the window function to the right. At each point, extract the next segment of the signal. Repeat the windowing and FFT operations to obtain This process continues until the entire signal sequence has been traversed. Finally, the spectra corresponding to all time slices are arranged in chronological order, forming a two-dimensional complex matrix. , where index Corresponding time (i.e., which window), index This corresponds to discrete frequency points. The amplitude value of this complex matrix is ​​calculated. This yields the time-frequency amplitude spectrum of the signal, which visually displays the distribution of signal energy in the time-frequency plane, providing a two-dimensional data foundation containing joint time-frequency information for subsequent processing.

[0075] In the multi-signal fusion method, the fusion ratio of the original data of the training UAV image transmission signal and the interference data is within a preset range, and the value range of the preset range is 0.2 to 0.8.

[0076] Specifically, in multi-signal fusion data enhancement methods, the fusion ratio coefficient This is a key control parameter that determines the relative intensity of the target UAV signal component and the background interference component in the synthesized signal. The value range is limited to a preset interval. The design incorporates clear engineering considerations.

[0077] In practice, for each data pair that needs to be fused (one drone signal) and a jamming signal (This does not use a fixed ratio, but rather a uniform distribution) A value is randomly sampled from the data and used as the basis for this fusion. For example, a random number generator might produce values ​​such as 0.35 or 0.62.

[0078] Set the lower limit to 0.2 (i.e.) This is to ensure that the energy contribution of the target UAV signal in the synthesized signal is no less than 20% of the total signal energy. This guarantees that although the target signal is interfered with, it will not be completely submerged in the background noise, thereby avoiding the generation of a large number of extremely difficult samples that the model cannot learn and that have no practical significance, and ensuring the effectiveness of the augmented data.

[0079] Set the upper limit to 0.8 (i.e.) This is to ensure that the energy contribution of the interference signal in the synthesized signal is not less than 20% of the total signal energy. This forces the model to learn to detect and identify interference in the presence of a certain intensity of background electromagnetic interference, avoiding the synthesized data being too "clean", thereby improving the robustness of the model in real complex electromagnetic environments.

[0080] By Randomization, taking values ​​within a reasonable range, can automatically generate a large number of training samples under different signal-to-interference ratio (STR) scenarios, greatly expanding the diversity of the dataset and enabling the trained model to adapt to various real-world situations, from strong target signals to relatively strong interference. After each sampling, the data is processed according to the formula... The fused signal data is calculated and generated for subsequent processing.

[0081] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0082] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for detecting and recognizing UAV image transmission signals based on an improved YOLOv8, characterized in that, Includes the following steps: S1: Acquire raw data and interference data of training drone image transmission signals in multiple scenarios; S2: Process the raw data of the UAV image transmission signal for training according to the preset preprocessing logic to obtain grayscale image data for training, and mark the UAV image transmission signal area in the grayscale image data for training to generate initial annotation data; S3: Based on the interference data, the initial labeled data is subjected to scenario-based extension processing adapted to the signal transmission characteristics of the UAV to obtain enhanced labeled data; S4: Construct an improved YOLOv8 model. By adjusting the parameters and structure of the feature extraction module of the YOLOv8 model backbone network, adapt it to the morphological features of the UAV image transmission signal. Then, use the enhanced annotation data to train the improved YOLOv8 model to obtain the UAV image transmission signal detection and recognition model. S5: Acquire the radio frequency signal to be detected, process the radio frequency signal to be detected according to the preset preprocessing logic of S2 to obtain grayscale image data to be detected, input the grayscale image data to be detected into the UAV image transmission signal detection and recognition model, and output the center frequency, bandwidth and corresponding UAV type of the UAV image transmission signal.

2. The method for detecting and recognizing UAV image transmission signals based on improved YOLOv8 according to claim 1, characterized in that, The raw data and interference data of the UAV image transmission signal used for training in step S1 include: Raw data of image transmission signals from training drones were collected from multiple dimensions, including drone attributes, signal parameters, transmission distance, and operating frequency band. Interference data is collected in the signal environment of the UAV, and the interference data is used to simulate electromagnetic interference scenarios in actual applications.

3. The method for detecting and recognizing UAV image transmission signals based on improved YOLOv8 according to claim 1, characterized in that, The preset preprocessing logic in step S2 includes time-frequency conversion, dimensionality reduction, and grayscale conversion. The execution process of the preset preprocessing logic is as follows: The raw data of the training UAV image transmission signal is processed by time-frequency conversion to obtain time-frequency domain data; The time-frequency domain data is subjected to dimensionality reduction processing to reduce data redundancy; Convert the dimensionality-reduced time-frequency domain data into grayscale image data.

4. The method for detecting and recognizing UAV image transmission signals based on improved YOLOv8 according to claim 3, characterized in that, The annotation process in step S2 is as follows: By using the annotation tool to select the area corresponding to the UAV image transmission signal in the training grayscale image data, the location range information of the area is marked to form the initial annotation data.

5. The method for detecting and recognizing UAV image transmission signals based on improved YOLOv8 according to claim 1, characterized in that, Step S3 includes: The original data of the training UAV image transmission signal corresponding to the initial labeled data is processed by frequency shift transformation to adapt to signal frequency fluctuation scenarios. The raw data of the training UAV image transmission signal corresponding to the initial labeled data is processed by noise addition to adapt to electromagnetic interference scenarios. A multi-signal fusion method is used to fuse the original data of the training UAV image transmission signal corresponding to the initial labeled data with the interference data to adapt to multi-target signal superposition scenarios. The signal after any of the above processing is processed according to the preset preprocessing logic of S2, and the region information corresponding to the processed signal is labeled to generate enhanced labeling data.

6. The method for detecting and recognizing UAV image transmission signals based on improved YOLOv8 according to claim 1, characterized in that, Step S4 involves adjusting the parameters and structure of the feature extraction module of the YOLOv8 model backbone network, including: The convolution kernel shape parameters of the CBS module in the YOLOv8 model backbone network were adjusted, and a non-square convolution kernel was used to replace the original square convolution kernel to adapt to the elongated shape of the UAV image transmission signal. The CBS module consists of convolutional layers, batch normalization layers, and activation function layers.

7. The method for detecting and recognizing UAV image transmission signals based on improved YOLOv8 according to claim 6, characterized in that, The step S4, which involves adjusting the feature extraction module structure of the YOLOv8 model backbone network, also includes: The structure of the Bottleneck module in the backbone network of the YOLOv8 model is optimized by adding a variable receptive field unit to the Bottleneck module. The variable receptive field unit includes multiple feature extraction branches with different receptive field sizes. By fusing the output features of the multiple branches, it can adapt to UAV image transmission signals with different parameters. The fused features are then superimposed on the input data of the Bottleneck module to complete feature enhancement.

8. The method for detecting and recognizing UAV image transmission signals based on improved YOLOv8 according to claim 1, characterized in that, The training process in step S4 includes: The enhanced annotation data is divided into a training set, a validation set, and a test set; Set the model training parameters, select a loss function suitable for the object detection task, and iteratively train the improved YOLOv8 model; After each round of training, the model performance is evaluated using the validation set, and training stops when the model performance meets the preset stability conditions. The performance of the model after training was stopped was verified using the test set. Once the verification was successful, the UAV image transmission signal detection and recognition model was obtained.

9. The method for detecting and recognizing UAV image transmission signals based on improved YOLOv8 according to claim 3, characterized in that, The time-frequency conversion process employs a short-time Fourier transform, using a window function to segment the signal and preserve its joint time-frequency characteristics.

10. The method for detecting and recognizing UAV image transmission signals based on improved YOLOv8 according to claim 5, characterized in that, In the multi-signal fusion method, the fusion ratio of the original data of the training UAV image transmission signal and the interference data is within a preset range, and the value range of the preset range is 0.2 to 0.8.

Images