Unmanned aerial vehicle signal identification method, electronic equipment and storage medium

Abstract

The present application relates to the technical field of unmanned aerial vehicle signal identification, and discloses an unmanned aerial vehicle signal identification method, an electronic device and a computer readable storage medium, the method comprising: obtaining a target signal image obtained by detecting a target unmanned aerial vehicle, the target signal image comprising frequency distribution information of a target signal; the duty ratio of the target signal is greater than a preset duty ratio threshold; amplifying the resolution of the target signal image according to a preset amplification factor to obtain an amplified signal image; identifying a plurality of edge contours of a target signal in the amplified signal image to obtain a plurality of first target areas on the amplified signal image; and reducing the resolutions of the plurality of first target areas according to a preset magnification factor to obtain a plurality of second target areas so as to identify target information of the target unmanned aerial vehicle according to the plurality of second target areas. In this way, the accuracy of unmanned aerial vehicle signal identification is improved.

Description

A method for identifying drone signals, electronic devices, and storage media. Technical Field

[0001] This application relates to the field of unmanned aerial vehicle (UAV) signal recognition technology, specifically to a method for recognizing UAV signals, an electronic device, and a computer-readable storage medium. Background Technology

[0002] With the booming development of the low-altitude economy, drones are increasingly used in logistics, inspection and other fields, and their signal recognition technology has become crucial.

[0003] A typical drone signal identification process usually includes: radio frequency acquisition of the target drone's signal, time-frequency transformation of the acquired signal to obtain a signal image, detection of the signal region in the signal image, feature extraction of the signal region, and classification or localization of the target drone based on the extracted features.

[0004] However, drone signals typically appear periodically in signal images, and existing technologies struggle to distinguish the boundaries between multiple periodic signal regions, making it impossible to accurately identify drone signals.

[0005] Therefore, improving the accuracy of drone signal recognition has become a pressing technical problem. Summary of the Invention

[0006] In view of the above problems, embodiments of this application provide a method for identifying drone signals, an electronic device, and a computer-readable storage medium to solve the problem that the accuracy of drone signal identification in the prior art needs to be improved.

[0007] According to one aspect of the embodiments of this application, a method for identifying unmanned aerial vehicle (UAV) signals is provided. The method includes: acquiring a target signal image obtained by detecting a target UAV, wherein the target signal image includes frequency distribution information of the target signal and the duty cycle of the target signal is greater than a preset duty cycle threshold; magnifying the resolution of the target signal image according to a preset magnification factor to obtain a magnified signal image; identifying multiple edge contours of the target signal in the magnified signal image to obtain multiple first target regions on the magnified signal image; and reducing the resolution of the multiple first target regions according to a preset magnification factor to obtain multiple second target regions, so as to identify target information of the target UAV based on the multiple second target regions.

[0008] Preferably, before magnifying the resolution of the target signal image according to a preset magnification factor to obtain an amplified signal image, the method further includes: identifying interference information in the target signal image to obtain the interference region where the interference information is located in the target signal image; generating a binary mask image according to the image size of the target signal image and the interference region in the target signal image, wherein the image size of the binary mask image is consistent with the image size of the target signal image, the binary mask image includes a mask region, and the mask region corresponds to the interference region in the target signal image; repairing the pixel values ​​of multiple pixels in the target signal image that correspond to the mask region in the binary mask image according to the pixel values ​​of multiple pixels in the target signal image to obtain a target repaired image; and magnifying the resolution of the target signal image according to a preset magnification factor to obtain an amplified signal image, including: magnifying the resolution of the target repaired image according to a preset magnification factor to obtain an amplified signal image.

[0009] Preferably, identifying interference information in the target signal image to obtain the interference region where the interference information is located includes: identifying single-tone interference information and constant-frequency interference information in the target signal image; determining the region where the single-tone interference information or constant-frequency interference information is located in the target signal image as the interference region; wherein, the method for identifying single-tone interference information includes: in time-frequency analysis, if an energy concentration region with a frequency bandwidth lower than a preset bandwidth threshold and a signal duration exceeding a preset stability threshold is identified from the target signal image, then it is determined that the region has single-tone interference, and single-tone interference information is obtained; or, morphological feature detection is performed on the target signal image in the form of a time-frequency graph, and if a straight line is detected in the target signal image, then it is determined that the region where the straight line is located has single-tone interference, and single-tone interference information is obtained; wherein, the method for identifying constant-frequency interference information includes: performing power spectral density statistics on the target signal image within a preset time window to obtain a power spectrum; determining a background noise energy reference value based on the power spectrum; traversing the frequency points in the power spectrum, if there is a frequency point whose average power value within the preset time window is continuously higher than the background noise energy reference value, then it is determined that the frequency point has constant-frequency interference, and constant-frequency interference information is obtained.

[0010] Preferably, the target signal image includes multiple interference regions; generating a binary mask image based on the image size of the target signal image and the interference regions in the target signal image includes: determining interference regions with a spatial distance less than a preset distance threshold as neighboring regions; merging the neighboring regions; performing a slight dilation process on the interference regions obtained after merging according to a preset dilation parameter to obtain an expanded interference region, wherein the expanded interference region is used to cover the edge of the interference region; and generating a binary mask image based on the image size of the target signal image and the expanded interference region in the target signal image.

[0011] Preferably, before identifying multiple edge contours of the target signal in the magnified signal image to obtain multiple first target regions on the magnified signal image, the method further includes: performing grayscale processing on the magnified signal image to obtain a grayscale magnified signal image; performing noise reduction processing on the grayscale magnified signal image; determining a first structuring element parameter based on the morphological features corresponding to the single-tone interference information; performing an opening operation on the noise-reduced magnified signal image based on the first structuring element parameter; determining a second structuring element parameter based on a preset magnification factor; and performing a closing operation on the opening magnified signal image based on the second structuring element parameter.

[0012] Preferably, before identifying multiple edge contours of the target signal in the magnified signal image and obtaining multiple first target regions on the magnified signal image, the method further includes: determining a grayscale histogram of the magnified signal image after the closing operation, wherein the horizontal axis of the grayscale histogram is the grayscale value of a pixel, and the vertical axis of the grayscale histogram is the number of pixels; determining multiple grayscale coordinate points based on the vertices of each vertical cylinder in the grayscale histogram, wherein each grayscale coordinate point corresponds one-to-one with a vertex of a vertical cylinder; determining the maximum value among the multiple grayscale coordinate points as the peak coordinate point; and determining the peak coordinate point with the largest number of pixels among the multiple peak coordinate points as the background peak point. The signal peak point is determined by identifying the point with the largest gray value among multiple peak coordinate points; the line connecting the background peak point and the signal peak point is defined as the energy trend line; the gray coordinate points located between the gray values ​​of the background peak point and the signal peak point are defined as valid coordinate points; the valid coordinate points farthest from the energy trend line are defined as the target coordinate point; the gray value of the target coordinate point is defined as the gray threshold; and the amplified signal image is converted from grayscale form to binary form based on the gray threshold. The binary form of the amplified signal image includes background points with pixel values ​​of the first type and signal points with pixel values ​​of the second type.

[0013] Preferably, identifying multiple edge contours of the target signal in the magnified signal image to obtain multiple first target regions on the magnified signal image includes: performing a closing operation on the binary form of the magnified signal image according to the second structuring element parameters; extracting the signal contours from the binary form of the magnified signal image after the closing operation to obtain the contour information of the target signal, wherein each contour information corresponds to a detection box; selecting detection boxes from the multiple detection boxes whose aspect ratio is within a preset aspect ratio range and whose area is within a preset area range; and determining the region defined by the selected detection boxes in the magnified signal image as the first target region.

[0014] Preferably, determining the region bounded by the selected detection boxes in the magnified signal image as the first target region includes: sorting the selected detection boxes in descending order according to their area to obtain the detection box arrangement order; initializing a retention set, wherein the retention set is initially empty; performing repeated detection on each detection box in the order of the detection boxes to obtain an updated retention set; determining the region bounded by the detection boxes in the updated retention set in the magnified signal image as the first target region; wherein the process of repeatedly detecting the current detection box includes: determining the intersection area between the current detection box and each retained detection box in the retention set; determining whether there is a retained detection box that satisfies condition one, wherein condition one includes: the intersection area between the retained detection box and the current detection box is equal to the area of ​​the current detection box, and the area of ​​the current detection box is equal to the area of ​​the current detection box. The area of ​​the current detection frame is less than or equal to the area of ​​the retained detection frame. If a retained detection frame satisfies condition one, the current detection frame is determined to be contained within a retained detection frame, and the repeated detection of the current detection frame is completed, and the next detection frame is processed. If no retained detection frame satisfies condition one, the intersection-union ratio (IUR) of the current detection frame with each retained detection frame in the retained set is determined, and it is determined whether a retained detection frame satisfies condition two. Condition two includes: the IUR corresponding to the retained detection frame is greater than a preset IUR threshold. If a retained detection frame satisfies condition two, the overlap between the current detection frame and the retained detection frame exceeds a preset overlap threshold, and the repeated detection of the current detection frame is completed, and the next detection frame is processed. If no retained detection frame satisfies condition two, the current detection frame is added to the retained set, and the next detection frame is processed.

[0015] According to another aspect of the embodiments of this application, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the method as described in any of the preceding claims.

[0016] According to another aspect of the embodiments of this application, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the method as described in any of the above.

[0017] This application embodiment amplifies the resolution of the target signal image by a preset magnification factor to obtain an amplified signal image, and identifies multiple edge contours of the target signal in the amplified signal image to obtain multiple first target regions on the amplified signal image. For UAV signals with high duty cycles, the boundaries of multiple periodic signal regions in the target signal image are blurred to the point of almost sticking together. By amplifying the resolution, the clarity between the difficult-to-distinguish boundaries can be improved, avoiding the identification of multiple periodic signal regions as a single signal region, thereby improving the accuracy of signal recognition.

[0018] The above description is merely an overview of the technical solutions of the embodiments of this application. In order to better understand the technical means of the embodiments of this application and to implement them in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the embodiments of this application more obvious and understandable, specific implementation methods of this application are described below. Attached Figure Description

[0019] The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Throughout the drawings, the same reference numerals denote the same components. In the drawings: Figure 1 shows a flowchart of a method for identifying drone signals according to an embodiment of this application; Figure 2 shows a schematic diagram of a first target region extracted from a target signal image according to an embodiment of this application; Figure 3 shows a schematic diagram of a first target region extracted from a signal image magnified at 2x magnification according to an embodiment of this application; Figure 4 shows a schematic diagram of a first target region extracted from a signal image magnified at 3x magnification according to an embodiment of this application; Figure 5 shows a schematic diagram of a binary mask image according to an embodiment of this application; Figure 6 shows a schematic diagram of a grayscale histogram according to an embodiment of this application; Figure 7 shows a structural schematic diagram of a drone signal identification device according to an embodiment of this application; Figure 8 shows a structural schematic diagram of an electronic device according to an embodiment of this application. Detailed Implementation

[0020] Exemplary embodiments of the present application will now be described in more detail with reference to the accompanying drawings. Although exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application may be implemented in various forms and should not be limited to the embodiments set forth herein.

[0021] Figure 1 shows a flowchart of the drone signal identification method provided in this application embodiment, which is executed by a drone signal identification device. This device can be a device with data reception and processing computing capabilities, such as a server deployed at a ground station, a dedicated signal processing module integrated into an anti-drone system, a handheld spectrum analyzer, or even an edge computing device with sufficient computing power. As shown in Figure 1, the method includes the following steps S110~S140: S110, acquiring a target signal image obtained by detecting the target drone, wherein the target signal image includes frequency distribution information of the target signal, and the duty cycle of the target signal is greater than a preset duty cycle threshold.

[0022] Unmanned aerial vehicle (UAV) signal identification equipment detects target UAVs through its hardware front-end, such as antennas and receivers. The target UAV is an unknown UAV whose identity needs to be identified; its type is not limited and can include consumer-grade multi-rotor UAVs, industrial-grade fixed-wing UAVs, or military UAVs.

[0023] The detection process for a target drone involves an antenna capturing the electromagnetic signals emitted or reflected by the drone. The receiver then amplifies, filters, down-converts, and performs analog-to-digital conversion on these signals, converting them into digital baseband or intermediate frequency (IF) signal data. Subsequently, the identification equipment processes these digital signals to generate a target signal image. It is important to emphasize that this target signal image is not an optical photograph, but rather a visual representation of the signal in a specific domain; its core content is the frequency distribution information of the target signal.

[0024] Typically, the target signal image is a time-frequency analysis image, such as a spectrum obtained from a short-time Fourier transform, a scaled image obtained from a wavelet transform, or a more advanced cyclic spectrum, or a time-frequency image. These images can intuitively show the two-dimensional or three-dimensional distribution of signal energy as time, frequency, or cyclic frequency change.

[0025] After generating the target signal image, the recognition device calculates the duty cycle of the target signal. The duty cycle is defined as the ratio of the pulse duration (i.e., the time the signal is "high" or in an active state) to the total period within a signal cycle. For example, if a signal cycle is 10 milliseconds and the pulse lasts for 9 milliseconds, its duty cycle is 90%. A high duty cycle signal means that the signal is in an active state for most of the time, resulting in it appearing as a large, continuous energy region on the time-frequency image, making the periodic pulse boundaries very blurred.

[0026] Specifically, refer to Figure 2. P1 in Figure 2 is the original image of the target region where the target signal needs to be identified. P2 in Figure 2 is the identification result of directly identifying the target region at the original resolution. In P2 of Figure 2, the identified target region is defined by a green box. That is, multiple periodic energy regions will be identified as one energy region. The target region is actually represented as multiple small yellow rectangles in P2 of Figure 2.

[0027] However, UAV signals generally exhibit a high duty cycle, primarily due to the practical requirements and technical implementation of their communication links. Modern UAV systems, especially those requiring continuous high-definition image transmission or real-time telemetry, typically employ continuous wave, high-repetition-frequency burst frames, or wideband spread spectrum modulation in their physical layer protocols. This design aims to maintain a stable, high-throughput data stream, resulting in signals that persist continuously in the time domain or are widely distributed in the frequency domain. Consequently, in the two-dimensional image constructed by time-frequency analysis, this manifests as large areas with continuous or dense energy distribution. Furthermore, time-frequency analysis methods such as the short-time Fourier transform essentially trade off between time resolution and frequency resolution. Their windowing operations lead to spectral leakage and energy blurring. For two high-duty-cycle signals that are adjacent in the time-frequency domain, their main lobes and side lobes overlap and merge after transformation, forming a connected energy block in the time-frequency image. This block lacks clear valleys or grooves that descend to the noise floor. Additionally, the limited signal-to-noise ratio and dynamic range in actual receiving environments cause the energy level attenuation at signal edges to transition gradually rather than drop sharply, further obscuring any potential gaps.

[0028] This adhesion, caused by the combined effects of signal physical characteristics, the mathematical nature of analysis methods, and observation noise, renders traditional image segmentation algorithms based on local gradients or thresholds ineffective. These algorithms cannot accurately define the precise boundaries of each independent signal entity, thus directly affecting the accuracy of subsequent parameter measurements, signal sorting, and identification.

[0029] To filter out such high duty cycle signals, this method sets a preset duty cycle threshold, which is an empirical value or a scalar set according to the specific application scenario, such as 80%, 85%, or 90%. When the calculated duty cycle of the target signal is greater than the preset duty cycle threshold, the system determines that the signal is a high duty cycle signal and needs to initiate subsequent steps S120-S140 for processing.

[0030] Specifically, the complete operation of step S110 is illustrated below: An anti-drone system deployed in a certain area uses an omnidirectional antenna to continuously monitor airspace electromagnetic signals. At a certain moment, the device detects an unknown signal. After preliminary analysis, it is found to be a frequency-hopping signal with a center frequency of 2.4 GHz. The device then performs a short-time Fourier transform on the signal, generating a 1024x1024 pixel spectrum as the target signal image. Simultaneously, the device analyzes the time-domain waveform of the signal, calculating its pulse repetition period to be 10 milliseconds and its pulse width to be 9.5 milliseconds, resulting in a duty cycle of 95%. Since 95% is greater than the system's preset duty cycle threshold of 90%, the signal is identified as a high duty cycle target signal, and subsequent procedures are triggered.

[0031] S120: The resolution of the target signal image is magnified according to the preset magnification factor to obtain an magnified signal image.

[0032] S120 aims to enhance the clarity of blurred boundaries within the target signal by increasing image resolution, laying the foundation for accurate segmentation of subsequent signal region boundaries.

[0033] The preset magnification factor is a pre-defined scalar or vector used to control the degree of resolution enhancement, such as 2x, 4x, or 8x. The choice of this factor requires a trade-off between processing effect and computational cost. A higher factor may result in better edge enhancement, but also a greater computational burden. This factor can also be adaptive, for example, dynamically adjusted based on the signal's duty cycle or the degree of edge blurring.

[0034] As shown in Figure 3, P1 in Figure 3 is the magnified signal image obtained after the resolution of the target signal image is magnified by 2 times, and P2 in Figure 3 is the recognition result of the target region at twice the original resolution. In P2 in Figure 3, the identified target region is represented by a green box. That is, some energy regions in multiple periodic energy regions are identified as one energy region. Similarly, the target region is actually represented by multiple small yellow rectangles in P2 in Figure 3. Further, refer to Figure 4. As shown in Figure 4, P1 in Figure 4 is the magnified signal image obtained after the resolution of the target signal image is magnified by 3 times, and P2 in Figure 4 is the recognition result of the target region at three times the original resolution. In P2 in Figure 4, the identified target region is represented by a green box. At this time, each energy region in multiple periodic energy regions is identified as a separate energy region. Thus, by magnifying the resolution to three times the original resolution, the boundaries of the signal region can be accurately distinguished.

[0035] It should be noted that, generally speaking, the higher the resolution, the more accurate the distinction of the boundaries of the signal region.

[0036] In the examples corresponding to Figures 2 to 4 above, a resolution of 3x is sufficient to accurately distinguish the boundaries of the signal area. In actual use, the preset magnification factor can be adjusted according to the actual needs.

[0037] Resolution refers to the number of pixels contained in an image per unit area, usually expressed in pixels for width and height, such as 1024x1024. Upscaling a target signal image does not simply involve physically enlarging the image; rather, it uses algorithms to generate a magnified signal image with more pixels and richer details.

[0038] The basic implementation of resolution upscaling can employ traditional interpolation algorithms, such as nearest neighbor interpolation, bilinear interpolation, or bicubic interpolation. To achieve better results, embodiments of this application preferably employ learning-based super-resolution algorithms, particularly deep learning-based models. For example, advanced models such as SRCNN (Super-Resolution Convolutional Neural Network), EDSR (Enhanced Deep Super-Resolution Network), and ESRGAN (Enhanced Super-Resolution Generative Adversarial Networks) can be used. These models, trained on massive pairs of high- and low-resolution images, learn complex mapping relationships for recovering high-frequency details (such as edges and textures) from low-resolution images. When applied to target signal images, these models can intelligently reconstruct signal boundaries that were originally blurred due to insufficient resolution, making the transition regions between different signal pulses steeper and clearer.

[0039] S130, identify multiple edge contours of the target signal in the magnified signal image to obtain multiple first target regions on the magnified signal image.

[0040] Among them, the edge contour refers to a set of pixels on the boundary of the target object in the image, where the gray value or color value of these points has changed drastically; the first target region is a closed image region surrounded by these identified edge contours, and each first target region corresponds to an original, independent UAV signal pulse.

[0041] Edge detection algorithms aim to find points of abrupt changes in grayscale in an image. Classic edge detection operators include the Sobel operator, the Prewitt operator, the Roberts operator, and the superior Canny operator.

[0042] After identifying the edge contours, they usually need to be transformed into closed regions through subsequent processing. This can be achieved using contour finding algorithms (such as the findContours function in OpenCV). These algorithms can construct a hierarchical contour structure based on the detected edge pixels. Then, multiple primary target regions can be obtained by contour filling or by directly defining regions based on the contour vertex coordinates.

[0043] S140: The resolution of multiple first target regions is reduced according to a preset magnification factor to obtain multiple second target regions, so as to identify the target information of the target UAV based on the multiple second target regions.

[0044] S140 is executed to map the recognition results in the magnified image coordinate system back to the original image coordinate system, and the final signal feature extraction and UAV recognition are performed based on the precisely mapped region.

[0045] The second target region is the result of the first target region after coordinate transformation. It represents the precise position and range occupied by each independent signal pulse on the scale of the original, unmagnified target signal image.

[0046] The target information is the intelligence that is ultimately desired about the target drone. It can be the drone's specific model, manufacturer, communication protocol type (such as OCU_Link, Lightbridge), or even the drone's flight status information (such as whether it is in recording or return-to-home mode).

[0047] Since the zoom in step S120 is isotropic (i.e., the length and width are zoomed in by the same factor), the reverse zooming operation also maintains isotropic properties. Specifically, you can obtain the corresponding coordinates in the original image coordinate system by dividing all the vertex coordinates (or bounding box coordinates) of each first target region by the preset zoom factor used in step S120. The region defined by these coordinates is the second target region.

[0048] For example, the boundary of a first target region in a magnified image is from (800, 1000) to (1200, 1400). If the magnification is 4, then the boundary of the corresponding second target region is from (200, 250) to (300, 350).

[0049] After obtaining multiple second target regions, the recognition device performs in-depth analysis on each region. The analysis includes, but is not limited to: calculating the duration of each region (corresponding to the signal pulse width), measuring the frequency bandwidth of each region (corresponding to the signal bandwidth), analyzing the modulation method of the signal within the region (such as FSK, PSK), and calculating the time interval between multiple regions (corresponding to the pulse repetition interval PRI). These extracted features constitute a multi-dimensional feature vector. Finally, the recognition device compares and matches this feature vector with a pre-established database containing various known UAV signal features. The matching algorithm can use Euclidean distance, cosine similarity, or more complex classifiers such as support vector machines (SVM), deep neural networks, etc. When the best match is found, the corresponding target information can be output.

[0050] This application embodiment amplifies the resolution of the target signal image by a preset magnification factor to obtain an amplified signal image, and identifies multiple edge contours of the target signal in the amplified signal image to obtain multiple first target regions on the amplified signal image. For UAV signals with high duty cycles, the boundaries of multiple periodic signal regions in the target signal image are blurred to the point of almost sticking together. By amplifying the resolution, the clarity between the difficult-to-distinguish boundaries can be improved, avoiding the identification of multiple periodic signal regions as a single signal region, thereby improving the accuracy of signal recognition.

[0051] Furthermore, the above method is described in full. The steps of the method generally include S210~S270: S210, Image acquisition: Acquire the target signal image obtained by detecting the target UAV; S220, Interference removal: Identify interference information in the target signal image, remove the interference information in the target signal image, and obtain the target restoration image; S230, Resolution magnification: Magnify the resolution of the target restoration image according to a preset magnification factor to obtain an magnified signal image; S240, Grayscale processing: Perform grayscale processing on the magnified signal image to obtain a grayscale magnified signal image, and perform noise reduction processing on the grayscale magnified signal image; S250, Threshold segmentation: Convert the magnified signal image from grayscale form to binary form according to a grayscale threshold; S260, Contour recognition: Identify multiple edge contours of the target signal in the magnified signal image to obtain multiple first target regions on the magnified signal image; S270, Reduction mapping: Reduce the resolution of the multiple first target regions according to a preset magnification factor to obtain multiple second target regions, so as to identify the target information of the target UAV based on the multiple second target regions.

[0052] However, when addressing the issue of signal boundary adhesion, there may be other factors affecting the accuracy of UAV signal recognition.

[0053] Next, we will explain the remaining influencing factors and the improvements made to address them.

[0054] This application, building upon the solution to the problem of high duty cycle signal boundary adhesion, further identifies and addresses another technical issue affecting the accuracy of UAV signal recognition: signal edge disruption and connected component errors caused by penetrating interference. Specifically, strong single-tone interference or equipment noise manifests as thin straight lines or bright bands on the time-frequency graph. This interference can penetrate the target UAV signal, causing subsequent edge contour recognition algorithms (such as the Canny operator) to misjudge the interference lines as signal boundaries, thus resulting in a break in the target signal contour. Alternatively, in morphological processing, interference may incorrectly connect unconnected signal regions into a large connected component, severely affecting the accuracy of feature extraction.

[0055] As shown in Figure 5, P1 is the target signal image, and P2 is the image generated by interference detection of the target signal image. The blue area in P2 is the background area, the yellow area is the signal area, and the yellow area enclosed by the green line is the interference area. The interference area that runs through the signal area, which is represented by the overlapping of two green lines in P2, is single-tone interference. Above the area corresponding to single-tone interference, there are many interference areas that do not appear periodically, which are specifically constant frequency interference.

[0056] Traditional global filters (such as notch filters) often damage useful signal components with frequencies close to the interference frequency when removing such interference, especially for frequency-hopping signals, which may lead to the loss of signal at some frequencies. Therefore, this application proposes a local selective interference removal scheme, which accurately identifies the interference region and generates a binary mask, and then uses image inpainting technology to repair only the masked region, thereby completely eliminating the interference while preserving the original features of the target signal to the maximum extent.

[0057] Therefore, further, S220 includes sub-steps S221~S223: S221, identify interference information in the target signal image and obtain the interference area where the interference information in the target signal image is located.

[0058] In S220, interference information refers to the collective term for useless signal components in the target signal image that are not generated by the target UAV and have a negative impact on identification. Its specific manifestations can be single-tone interference, constant frequency interference, or other sudden noise. The interference area is the specific location and range occupied by the interference information in the two-dimensional pixel space of the target signal image. The target restoration image is the image obtained after interference removal processing. Its characteristic is that the interference information is effectively removed, while the energy distribution and morphological characteristics of the target UAV signal are well preserved.

[0059] This step is further refined into S211a and S211b, which identify two typical types of interference respectively. Furthermore, S211 includes sub-steps S211a~S211b: S211a, identifies single-tone interference information and constant-frequency interference information in the target signal image.

[0060] S211b, the region where the single-tone interference information or constant-frequency interference information is located in the target signal image is determined as the interference region.

[0061] The method for identifying single-tone interference information includes: in time-frequency analysis, if an energy concentration region with a frequency bandwidth lower than a preset bandwidth threshold and a signal duration exceeding a preset stability threshold is identified from the target signal image, then the region is determined to have single-tone interference, and single-tone interference information is obtained; or, morphological feature detection is performed on the target signal image in the form of a time-frequency diagram, and if a straight line is detected in the target signal image, then the region where the straight line is located is determined to have single-tone interference, and single-tone interference information is obtained.

[0062] The method for identifying constant frequency interference information includes: performing power spectral density statistics on the target signal image within a preset time window to obtain the power spectrum; determining the background noise energy reference value based on the power spectrum; traversing the frequency points in the power spectrum, and if there is a frequency point whose average power value within the preset time window is continuously higher than the background noise energy reference value, then it is determined that there is constant frequency interference at that frequency point, and constant frequency interference information is obtained.

[0063] Single-tone interference appears as a thin line with highly concentrated energy and a relatively constant frequency position on the time-frequency graph, corresponding to a stable sine wave in the time domain; constant-frequency interference appears as a rectangular or striped area with a fixed frequency range and continuous energy, usually generated by carrier leakage from continuous operation of the equipment or non-frequency hopping communication signals.

[0064] For the identification of single-tone interference information, this application provides two complementary methods. The first is statistical analysis based on time-frequency features: In the time-frequency analysis image, the algorithm traverses all regions with concentrated energy and calculates the frequency bandwidth and signal duration of each region. If the frequency bandwidth of a certain region is lower than a preset bandwidth threshold (e.g., less than 1.5 times the signal resolution) and its signal duration exceeds a preset stability threshold (e.g., greater than 80% of the signal analysis duration), then it is determined that single-tone interference exists in that region, thus obtaining single-tone interference information. The second is line detection based on morphological features: Since single-tone interference appears as a straight line on the time-frequency image, line detection algorithms such as Hough Transform can be used to directly search for straight lines that satisfy specific length and angle constraints in the target signal image. If such a straight line is detected, it is determined that the region where the straight line is located contains single-tone interference.

[0065] For the identification of constant-frequency interference information, stability analysis of the power spectrum is used. Specifically, the following steps are taken: First, power spectral density statistics are performed on the target signal image within a preset time window (e.g., the entire image duration) to obtain a time-averaged power spectrum. Then, based on the distribution of this power spectrum, a background noise energy benchmark value is determined, for example, by calculating the median or lower quantile of the power spectrum. Finally, the algorithm iterates through each frequency point in the power spectrum. If it finds that the average power value of a certain frequency point within the preset time window is consistently significantly higher than the background noise energy benchmark value (e.g., more than 10 dB higher), then it is determined that constant-frequency interference exists at that frequency point, thus obtaining the constant-frequency interference information.

[0066] S222, Generate a binary mask image based on the image size of the target signal image and the interference region in the target signal image, wherein the image size of the binary mask image is consistent with the image size of the target signal image, the binary mask image includes a mask region, and the mask region corresponds to the interference region in the target signal image.

[0067] A binary mask image is a black-and-white image with the same width and height as the target signal image. Specifically, refer to P3 in Figure 5. Its pixel values ​​are typically only 0 and 1. Pixels with a value of 1 constitute the mask area, as shown by the white area in P3 in Figure 5. It precisely corresponds to the interference area in the target signal image that needs to be repaired. Pixels with a value of 0 correspond to the non-interference area that needs to be retained, as shown by the black area in P3 in Figure 5.

[0068] However, directly using the initially identified interference area may have shortcomings. For example, multiple neighboring interference points may not be connected, or the edges of the interference area may not be completely covered. Therefore, this step is further refined into S222a-S222d to optimize the mask quality.

[0069] The target signal image includes multiple interference regions. Further, S222 also includes sub-steps S222a~S222d: S222a, the interference regions with a spatial distance less than a preset distance threshold among the multiple interference regions are determined as neighboring regions.

[0070] Spatially adjacent or close independent disturbance patches are merged to avoid leaving untreated gaps between them during repair.

[0071] S222b, merges neighboring areas.

[0072] Neighboring areas are merged, meaning these neighboring areas are treated as a single, unified interference area.

[0073] S222c, the interference region obtained after merging is slightly expanded according to the preset expansion parameters to obtain the expanded interference region, wherein the expanded interference region is used to cover the edge of the interference region.

[0074] Dilation is a fundamental operation in mathematical morphology. It involves scanning an image using a structuring element (such as a 3x3 square kernel). If any pixel in the area covered by the structuring element belongs to the interference region, the pixel corresponding to the center of the structuring element is also designated as the interference region.

[0075] The purpose of micro-dilation is to slightly expand the mask area to cover any undetected energy diffusion at the edges of the interference area, ensuring thorough repair. For example, using a 3x3 dilation kernel to dilate the merged area, extending its edges outward by one pixel, yields the dilated interference area.

[0076] S222d generates a binary mask image based on the image size of the target signal image and the dilation interference region in the target signal image.

[0077] A zero-based image of the same size as the original image is created, and all pixel values ​​within the dilated interference region are set to 1, generating a binary mask image containing white blocks (mask regions).

[0078] S223, Based on the pixel values ​​of multiple pixels in the target signal image, repair the pixel values ​​of multiple pixels in the target signal image that correspond to the mask area in the binary mask image, and obtain the target repaired image.

[0079] There are several algorithms available for implementing this process, and the embodiments in this application can be selected according to the needs of the application scenario. An efficient and commonly used method is to use the `cv2.inpaint` function provided by the OpenCV library, which supports two algorithms: the Navier-Stokes method and the Telea method. The Telea method is recommended for narrowband interference removal from UAV time-frequency maps due to its speed and effectiveness. The Telea method smoothly fills in information by solving a partial differential equation, starting from the boundary of the mask region and gradually working inwards. It prioritizes pixels on the boundary, using known pixels in their neighborhood to quickly propagate and estimate the values ​​of unknown pixels, thereby efficiently removing interference while preserving the texture and structure of the surrounding signal. For more complex background interference, block-based inpainting algorithms, such as PatchMatch, can be used. This algorithm finds the small image patch that is most similar to the surrounding environment of the block to be repaired in the undamaged area of ​​the entire image, and then fills the area to be repaired with the most similar small image patch. This method has a better effect on preserving complex textures, but the computational cost is relatively large. With sufficient training data, even image inpainting models based on generative adversarial networks (GANs) can be used to generate more visually realistic repair content that is highly integrated with the surrounding environment by learning from a large amount of image data.

[0080] For example, the process of step S223 is as follows: The binary mask image generated in S222 and the original target signal image are simultaneously input into the cv2.inpaint function (using the Telea algorithm). The function iterates through each pixel in the mask region. For each pixel to be repaired, it analyzes the grayscale value and gradient of the non-masked pixels within a certain radius around it, and then calculates the most likely grayscale value of the pixel to be repaired based on this information, and fills it with this value. After processing, the interfering pixels in the original image corresponding to the mask region are replaced with new pixel values ​​calculated based on the surrounding signal characteristics. Abrupt interference lines or bands in the image are eliminated, and the signal energy distribution becomes smooth and natural.

[0081] By utilizing the undisturbed pixel information around the mask area (i.e., pixels in the target signal image corresponding to the non-masked area in the binary mask image) in steps S221-S223, pixel values ​​within the mask area are inferred and reconstructed to remove interference. By repairing only locally, the negative impact of interference on subsequent processing can be eliminated while fully preserving the time-frequency characteristics of the target UAV signal, such as frequency hopping patterns and bandwidth changes. This provides a good data source for subsequent resolution magnification and edge contour recognition, thereby improving the accuracy of UAV signal recognition.

[0082] S240 includes sub-steps S241~S242: S241, performing grayscale processing on the amplified signal image to obtain a grayscale amplified signal image; S242, performing noise reduction processing on the grayscale amplified signal image.

[0083] If the input is a color pseudo-color image (commonly a heatmap), first convert it to grayscale (brightness weighting can be used), and then perform Gaussian filtering, with specific parameters such as GaussianBlur(kernel=(5,5), σ=0). This is used to smooth noise on the magnified image and avoid Canny from producing a large number of false edges. The kernel size can be increased when the SNR is low.

[0084] After image restoration (S220~S223), resolution upscaling (S230), and preliminary noise reduction (S240), the quality of the target signal image has been greatly improved. However, two types of minor imperfections may still exist that affect subsequent processing: First, the image restoration process may not completely cover all interference, leaving some small, linear interference structures; second, after upscaling and noise reduction, the target signal itself may have some tiny holes or gaps inside. These imperfections may lead to the misidentification of residual interference as independent targets or the incorrect segmentation of a continuous signal region into multiple fragments during subsequent thresholding (S250) and contour recognition (S260).

[0085] Therefore, this application introduces morphological processing steps S243~S246 after S240 and before S250 to accurately remove residual linear interference through an opening operation and effectively enhance the internal connectivity of the target signal through a closing operation, thereby providing a cleaner image for subsequent steps.

[0086] Specifically, after executing S240 and before executing S250, the method further includes S243~S246: S243, determining the first structural element parameter based on the morphological features corresponding to the monotone interference information.

[0087] Single-tone interference information specifically refers to the small linear interference that remains in the image after the interference removal steps S220-S223. These may be interference edges that the repair algorithm failed to fully cover, or weak device noise.

[0088] Morphological features refer to the geometric shape characteristics of these residual interferences in the image, namely elongation and linearity, which is the main basis for distinguishing them from blocky target signals.

[0089] Since residual monotone interference typically appears as thin horizontal or vertical lines on the time-frequency graph, the first structuring element parameter should be set to effectively match and remove such lines. For example, a structuring element with a length (k-value) greater than the width of the interference line but much smaller than the vertical dimension of the target signal can be set. To handle both horizontal and vertical interference, two orthogonal structuring elements are often used: one vertical (e.g., 1x7 pixels) and one horizontal (e.g., 7x1 pixels).

[0090] S244, perform opening operation on the amplified signal image after noise reduction according to the first structural element parameters.

[0091] Opening is a fundamental operation in mathematical morphology, typically performed by first erosion and then dilation. Erosion uses a structuring element as a probe, scanning every pixel in the image. If any pixel within the area covered by the structuring element belongs to the background (low energy), the pixel corresponding to the center of the structuring element is set to the background. This process eliminates bright features smaller than the structuring element. Dilation, on the other hand, works by setting the center point to the target if at least one pixel within the area covered by the structuring element belongs to the target (high energy). Combinations of opening operations effectively remove isolated bright spots and thin lines smaller than the structuring element, while having minimal impact on the basic shape and size of targets larger than the structuring element.

[0092] S245, determine the parameters of the second structural element according to the preset magnification factor.

[0093] The preset magnification factor is the magnification factor used in step S230, which directly determines the size of the target signal in the magnified signal image and the relative size of internal holes and gaps. For example, a hole that is only 1 pixel in the original image will become a 4x4 pixel hole after 4x magnification.

[0094] The second structuring parameter defines the shape and size of the structuring element used in the closing operation. This second structuring parameter is designed for connection and padding to match the size of the structuring element to the magnification of the image.

[0095] As the magnification increases, the feature sizes in the image become larger. Therefore, the structuring element used for filling should also be increased accordingly to maintain consistent processing results. Typically, the second structuring element parameter is set to a small, compact shape, such as a rectangle or circle, whose size can be proportional to the preset magnification. For example, a 3x3 rectangular structuring element can be set, or a rectangle with a size of (1 + magnification / 2).

[0096] For example, if the preset magnification factor used in step S230 is 4x, the system determines the second structuring element parameter as follows: a rectangle in shape and a size of 3x3 pixels. This size selection takes into account the size of typical holes and gaps that may appear inside the target signal in a 4x magnified image. The 3x3 structuring element is sufficient to effectively bridge them without excessively blurring the signal boundaries, ensuring that the closing operation achieves the desired filling effect regardless of how many times the image is magnified.

[0097] S246, perform closing operation processing on the amplified signal image after opening operation processing according to the second structuring element parameters.

[0098] Closing is also a fundamental operation in mathematical morphology, typically involving dilation followed by erosion. The dilation operation expands the target object outwards, filling its internal cavities and connecting adjacent, smaller, independent parts. The subsequent erosion operation shrinks the target object back to approximately its original size, but the filled cavities and connected gaps remain closed.

[0099] In the magnified signal image after opening processing, several small black holes caused by noise or signal instability, as well as some tiny gaps, exist inside the main body of the UAV signal. During the dilation step, the signal boundary expands inward and outward, these holes are filled with signal energy, and the gaps are bridged. The subsequent erosion step shrinks the outwardly expanding boundary back, but the internal filling effect is preserved. Ultimately, the target signal region in the image becomes more complete and solid, without any internal voids, and the overall connectivity is significantly enhanced. By filling the holes and bridging the gaps, it ensures that the target signal can be completely extracted in the subsequent thresholding segmentation, and a single, continuous, and smooth edge contour can be formed in contour recognition, thereby avoiding contour breakage or missegmentation caused by incomplete internal signal.

[0100] S250, Threshold segmentation: Convert the amplified signal image from grayscale form to binary form based on the grayscale threshold.

[0101] Traditional thresholding methods face significant challenges in contour recognition. Fixed thresholding is clearly inadequate to adapt to the vast differences in pixel brightness distribution caused by different sampling devices, different detection backgrounds, and different drone models. Even the classic Otsu algorithm often fails when processing drone signal images with uneven background noise and target signal energy distribution and severely skewed histograms, failing to find an effective boundary point that accurately distinguishes between background noise and drone signals.

[0102] Based on this, the threshold selection problem can be transformed into a geometric problem of finding the point furthest from the trend line between the background and signal bimodal peaks. This allows for the determination of the optimal segmentation threshold based on the unique grayscale distribution characteristics of each image, ensuring accurate separation of the background and signal. S250 includes sub-steps S251–S259: S251, determining the grayscale histogram of the amplified signal image after the closing operation, where the horizontal axis of the grayscale histogram represents the grayscale value of the pixel, and the vertical axis represents the number of pixels.

[0103] A grayscale histogram describes the number of pixels in an image with each grayscale level. The horizontal axis of the histogram represents the grayscale value of the pixel, typically ranging from 0 (pure black) to 255 (pure white); the vertical axis represents the number of pixels in the image with a specific grayscale value. For example, a vertical column with coordinates (50, 1200) indicates that there are 1200 pixels in the image with a grayscale value of 50. Please refer to Figure 6, which shows a schematic diagram of a grayscale histogram provided in an embodiment of this application. The horizontal axis of the grayscale histogram is labeled x in Figure 6, and the vertical axis is labeled H[x] in Figure 6.

[0104] 252. Based on the vertices of each vertical cylinder in the grayscale histogram, multiple grayscale coordinate points are determined, where each grayscale coordinate point corresponds one-to-one with a vertex of the vertical cylinder.

[0105] Vertical cylinders are the basic building blocks of a grayscale histogram, with each cylinder representing a grayscale level. A vertex is the highest point of each cylinder; its x-coordinate represents the grayscale value corresponding to the center point within the grayscale range of that vertical cylinder, and its y-coordinate represents the number of pixels corresponding to that grayscale value.

[0106] As shown in Figure 6, points (p, H[p]) and (r, H[r]) are both vertices.

[0107] S253 determines the maximum value point among multiple grayscale coordinate points as the peak coordinate point.

[0108] In a grayscale histogram, a grayscale point is a local maximum if its number of pixels (vertical coordinate value) is greater than the number of pixels corresponding to its left and right adjacent grayscale levels.

[0109] Peak coordinates are the set of coordinates of the maximum points. Peak coordinates usually correspond to the main components of the image (such as background, signal, etc.).

[0110] S254 determines the background peak point as the point with the largest number of pixels among multiple peak coordinate points, and determines the signal peak point as the point with the largest gray value among multiple peak coordinate points.

[0111] Background peaks represent the dominant background noise peaks in an image. They are characterized by their wide coverage area, thus corresponding to the largest number of pixels. As shown in Figure 6, the point (p, H[p]) is the background peak.

[0112] The signal peak point represents the peak value of the target UAV signal. Its characteristic is that the energy is usually stronger than the background, so the corresponding gray value is the highest. The point (r, H[r]) in Figure 6 is the signal peak point.

[0113] S255 defines the line connecting the background peak point and the signal peak point as the energy trend line.

[0114] Geometrically, the energy trend line is a straight line segment connecting the background peak point and the signal peak point. It represents a linear transition trend from background energy to signal energy and is used to measure the transition between the background and the signal. The optimal segmentation threshold is located at the position furthest from this baseline.

[0115] S256, determine the gray coordinate points between the gray values ​​of the background peak point and the signal peak point among multiple gray coordinate points as valid coordinate points.

[0116] Valid coordinate points refer to all grayscale coordinate points located between the background peak and the signal peak. This interval is the transition zone between the background and the signal, and contains the true threshold information. As shown in Figure 6, the grayscale coordinate points corresponding to the grayscale values ​​between p and r are the valid coordinate points.

[0117] S257 determines the target coordinate point as the one that is furthest from the energy trend line among the valid coordinate points.

[0118] The target coordinate point is the point geometrically furthest from the energy trend line among multiple valid coordinate points. This point can characterize the difference between noise energy and signal energy; that is, the difference between signal energy and noise energy is greatest at the target coordinate point. The target coordinate point is the optimal dividing point for distinguishing between background and signal. Even if the troughs of the histogram are very flat or there are multiple small fluctuations, this method can stably find the global optimum.

[0119] In Figure 6, the dashed line segment perpendicular to the energy trend line is used to represent the distance between the effective coordinate point at the gray value of 124.0 and the energy trend line.

[0120] S258, determine the gray value of the target coordinate point as the gray threshold.

[0121] The red dashed line at the grayscale value of 124.0 in Figure 6 represents the grayscale threshold.

[0122] S259, convert the magnified signal image from grayscale form to binary form according to the grayscale threshold, wherein the magnified signal image in binary form includes background points with pixel values ​​of the first type and signal points with pixel values ​​of the second type.

[0123] A binary image is an image in which pixel values ​​have only two possibilities, typically 0 and 255.

[0124] The first type of value usually represents the background and is assigned a value of 0 (black). The second type of value usually represents the target signal and is assigned a value of 255 (white).

[0125] Background points and signal points refer to pixels in the original image whose grayscale values ​​are below and above the threshold, respectively.

[0126] By using an adaptive threshold, the target signal in the image is extracted completely and clearly, while background noise is suppressed.

[0127] S260, contour recognition: identify multiple edge contours of the target signal in the magnified signal image to obtain multiple first target regions on the magnified signal image; the specific execution of S260 can be found in S130, and will not be repeated here.

[0128] After image binarization, the crucial contour recognition stage begins, requiring the extraction of independent regions representing the target UAV signal from the binary image and the removal of all invalid regions caused by noise or processing defects. Specifically, S260 further includes sub-steps S261~S264: S261, performing a closing operation on the amplified binary signal image based on the second structuring element parameters.

[0129] Execute S261 to fix potential edge breakage issues in binary images.

[0130] After thresholding, the edges of the target signal may have slight breaks or discontinuities. These broken edges can be connected by performing morphological closing operations, i.e., dilation followed by erosion.

[0131] For example, a 3x3 rectangular structural element can be used to fill small holes and gaps inside the target area and connect adjacent broken parts, thereby making the target signal area a more complete and connected whole, thus avoiding the misidentification of a complete signal pulse as multiple fragments due to edge breaks.

[0132] S262, perform signal contour extraction on the amplified signal image in binary form after the closing operation to obtain the contour information of the target signal, wherein each contour information corresponds to a detection box.

[0133] Execute S262 to locate and delineate the boundaries of all connected regions in the image.

[0134] Specifically, standard contour finding algorithms can be applied, such as the cv2.findContours function in the OpenCV library. This algorithm scans the entire binary image, identifies all connected regions with non-zero pixel values ​​(i.e., white), and returns the set of coordinates of the boundary pixels of each region, i.e., contour information.

[0135] Typically, a minimum bounding rectangle, or detection box, is used to approximate each contour.

[0136] S263, Filter out the detection frames from multiple detection frames whose aspect ratio is within the preset aspect ratio range and whose area is within the preset area range.

[0137] S263 is executed to filter the initially extracted detection boxes based on prior knowledge, removing obvious noise.

[0138] Drone signal pulses typically have a specific geometry on the time-frequency graph, with their aspect ratio and area within a reasonable range. By setting a preset aspect ratio range (e.g., between 0.5 and 2.0) and a preset area range, detection boxes such as thin interference lines or tiny noise points can be filtered.

[0139] Specifically, the area range needs to be adaptively adjusted according to the preset magnification factor in step S230, for example, by multiplying the original minimum area threshold by the square of the magnification factor.

[0140] S264, the region defined by the selected detection box in the magnified signal image is determined as the first target region.

[0141] In time-frequency analysis, the signal center usually has the strongest energy (brightest) and gradually decays at the edges. It is a normal physical phenomenon that a small bright area of ​​signal (high-energy core) is surrounded by a large bright area of ​​signal (complete signal range).

[0142] If a small area is completely surrounded by a large area, they are highly likely to be components of the same signal source rather than two independent signals. Therefore, it is necessary to retain the outer frame to more accurately represent the actual coverage of the signal. In other words, it is necessary to further deduplicate the overlapping detection frames that may still exist after geometric screening to ensure that each signal pulse is represented by only one detection frame.

[0143] Specifically, S264 includes sub-steps S264a~S264d: S264a, sorts the selected detection boxes in descending order according to their area size to obtain the detection box arrangement order.

[0144] S264b, initialize a hold set, wherein the hold set is initially an empty set.

[0145] S264c: According to the order of the detection boxes, each detection box is repeatedly detected to obtain the updated retain set.

[0146] S264d defines the region bounded by the detection box in the updated retain set of the amplified signal image as the first target region.

[0147] The process of repeatedly detecting the current detection box in step S264c includes S311~S316: S311, determining the intersection area between the current detection box and each retained detection box in the retention set; S312, determining whether there is a retained detection box that satisfies condition one, wherein condition one includes: the intersection area between the retained detection box and the current detection box is equal to the area of ​​the current detection box, and the area of ​​the current detection box is less than or equal to the area of ​​the retained detection box; if there is a retained detection box that satisfies condition one, then jump to S313 and execute S313; if there is no retained detection box that satisfies condition one, then jump to S314 and execute S314~S316.

[0148] S313, determine that the current detection box is contained by the retained detection box, complete the duplicate detection of the current detection box, jump to S311, and continue to process the next detection box.

[0149] S314, determine the intersection-union ratio (IU) of the current detection box with each retained detection box in the retained set, and determine whether there is a retained detection box that satisfies condition two, wherein condition two includes: the IU corresponding to the retained detection box is greater than a preset IU threshold; if there is a retained detection box that satisfies condition two, then jump to S315 and execute S315; if there is no retained detection box that satisfies condition two, then jump to S316 and execute S316.

[0150] S315, determine that the overlap between the current detection box and the retained detection boxes exceeds the preset overlap threshold, complete the duplicate detection of the current detection box, jump to S311, and continue to process the next detection box; S316, add the current detection box to the retained set, and jump to S311 to continue to process the next detection box.

[0151] The deduplication process, from S311 to S316, determines whether to add the current bounding box to the retention set. This deduplication is divided into two layers: the first layer (S311-S313) checks the inclusion relationship; if the current bounding box is completely contained by a larger box in the retention set, the current bounding box is discarded directly; the second layer (S314-S316) checks the overlap relationship; if the overlap (IOU) between the current bounding box and a box in the retention set exceeds a preset threshold, the current bounding box is also discarded; only when the overlap between the current bounding box and all the retained bounding boxes is very low is it added to the retention set.

[0152] Finally, the regions bounded by all detection boxes in the retained set are determined as the first target regions. S264 solves the problem of region overlap caused by high duty cycle signals or morphological processing through a hierarchical deduplication algorithm, ensuring that each first target region in the final output is unique and valid, providing an accurate target for subsequent reduction mapping and feature recognition.

[0153] Specifically, S270, reduction mapping: The resolution of multiple first target regions is reduced according to a preset magnification factor to obtain multiple second target regions, so as to identify the target information of the target UAV based on the multiple second target regions. The implementation of S270 is the same as that of S140 above, and the specific implementation of S140 can be referred to, and will not be repeated here.

[0154] After completing contour recognition on a high-resolution image, the recognition results need to be accurately mapped back from the magnified virtual coordinate system to the real physical coordinate system of the original signal.

[0155] Coordinate inverse scaling transforms the geometric parameters of the first target region from the magnified image scale back to the original image scale. Specifically, each first target region can be represented by a bounding box, whose coordinates and dimensions on the magnified image are (x, y, y). s , y s , w s , h s To obtain the second target region at the corresponding original resolution, a division operation is performed on each parameter of the bounding box using a preset magnification factor, for example, x = x s / scale_factor, y = y s / scale_factor, and so on.

[0156] Since pixel coordinates must be integers, this operation usually involves rounding. However, simple scaling may cause the calculated coordinates to exceed the boundaries of the original image, resulting in negative values ​​or values ​​greater than the image width. Therefore, boundary correction is further performed using a clamping function to restrict the calculated coordinate values ​​to within the valid range of the original image, such as x = clamp(x, 0, original_width - 1), to ensure the validity and accuracy of the coordinates of each secondary target region and avoid program errors or data misalignment caused by coordinates exceeding the limits.

[0157] During coordinate inverse scaling, division and rounding operations inevitably introduce minute pixel-level errors, which may cause slight deviations between the boundary of the second target region and the actual energy boundary of the original signal. To compensate for this error, a small-scale fine-tuning can be performed on the local area covered by each second target region on the original target signal image. For example, a small-scale morphological closing operation or local binarization can be performed on the local area again to correct the boundary offset caused by scaling, so that the region boundary perfectly matches the signal energy distribution.

[0158] In addition, a confidence score can be calculated for each second target region. This score can be based on indicators such as the average signal energy in the region, the integrity of the edge contour, or the degree of matching with typical signal models, such as confidence, to provide a quantitative basis for subsequent identification decisions.

[0159] When multiple potential targets are detected or results need to be fused from multiple frames of data, the system can also filter or weight the data based on confidence scores to select reliable target information.

[0160] Figure 7 shows a schematic diagram of the structure of the drone signal identification device provided in an embodiment of this application. As shown in Figure 7, the identification device 400 includes: an acquisition module 410, a magnification module 420, an identification module 430, and a reduction module 440.

[0161] The acquisition module 410 is used to acquire the target signal image obtained by detecting the target UAV. The target signal image includes the frequency distribution information of the target signal and the duty cycle of the target signal is greater than a preset duty cycle threshold.

[0162] The amplification module 420 is used to amplify the resolution of the target signal image according to a preset amplification factor to obtain an amplified signal image.

[0163] The recognition module 430 is used to recognize multiple edge contours of the target signal in the magnified signal image to obtain multiple first target regions on the magnified signal image; the reduction module 440 is used to reduce the resolution of the multiple first target regions according to a preset magnification factor to obtain multiple second target regions, so as to identify the target information of the target UAV based on the multiple second target regions.

[0164] The drone signal identification device 400 of this application embodiment also includes other modules for performing the steps of the above method embodiments, which will not be described in detail here.

[0165] Figure 8 shows a schematic diagram of the structure of the electronic device provided in the embodiment of this application. The specific implementation of the electronic device 500 is not limited by the specific embodiment of this application.

[0166] As shown in Figure 8, the electronic device 500 may include a processor 502 and a memory 504.

[0167] The memory 504 is used to store the computer program 506. The memory 504 may include high-speed RAM, and may also include non-volatile memory, such as at least one disk drive. The computer program 506 may include computer-executable instructions.

[0168] The processor 502 is used to execute the computer program 506 to implement the above-described embodiment of the drone signal recognition method.

[0169] Processor 502 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application. Electronic device 500 may include one or more processors of the same type, such as one or more CPUs; or it may include processors of different types, such as one or more CPUs and one or more ASICs.

[0170] This application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for identifying drone signals.

[0171] This application provides a computer program that can be executed by a processor to implement the above-described method for identifying drone signals.

[0172] This application provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described method for identifying drone signals.

[0173] In the several embodiments provided in this application, any function, if implemented as a software functional module / unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the technical solution of this application can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or other electronic device) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing computer program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0174] The algorithms or displays provided herein are not inherently related to any particular computer, virtual system, or other device. Various general-purpose systems can also be used in conjunction with the teachings herein. The required structure for constructing such systems is apparent from the above description. Furthermore, the embodiments of this application are not directed to any particular programming language. It should be understood that the content of this application described herein can be implemented using various programming languages, and the above description of specific languages ​​is for the purpose of disclosing the best mode of implementation of this application.

[0175] It should be noted that the above embodiments are illustrative of this application and not restrictive, and those skilled in the art can devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses should not be construed as limiting the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. This application can be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In claims enumerating several means, several units or modules of these means may be embodied by the same item of hardware. The use of the words first, second, and third, etc., does not indicate any order. These words can be interpreted as names. The steps in the above embodiments, unless otherwise specified, should not be construed as limiting the order of execution.

[0176] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for identifying unmanned aerial vehicle (UAV) signals, characterized in that, The method includes: acquiring a target signal image obtained by detecting a target drone, wherein the target signal image includes frequency distribution information of the target signal, and the duty cycle of the target signal is greater than a preset duty cycle threshold; magnifying the resolution of the target signal image according to a preset magnification factor to obtain a magnified signal image; identifying multiple edge contours of the target signal in the magnified signal image to obtain multiple first target regions on the magnified signal image; reducing the resolution of the multiple first target regions according to the preset magnification factor to obtain multiple second target regions, so as to identify the target information of the target drone based on the multiple second target regions.

2. The method according to claim 1, characterized in that, Before enlarging the target signal image to a preset magnification factor to obtain an enlarged signal image, the method further includes: identifying interference information in the target signal image to obtain the interference region where the interference information is located; generating a binary mask image based on the image size of the target signal image and the interference region in the target signal image, wherein the image size of the binary mask image is consistent with the image size of the target signal image, the binary mask image includes a mask region, and the mask region corresponds to the interference region in the target signal image; repairing the pixel values ​​of multiple pixels in the target signal image that correspond to the mask region in the binary mask image based on the pixel values ​​of multiple pixels in the target signal image to obtain a target repaired image; the step of enlarging the target signal image to a preset magnification factor to obtain an enlarged signal image includes: enlarging the target repaired image to a preset magnification factor to obtain an enlarged signal image.

3. The method according to claim 2, characterized in that, The step of identifying interference information in the target signal image and obtaining the interference region where the interference information is located in the target signal image includes: identifying single-tone interference information and constant-frequency interference information in the target signal image; determining the region where the single-tone interference information or constant-frequency interference information is located in the target signal image as the interference region; wherein, the method for identifying the single-tone interference information includes: in time-frequency analysis, if an energy concentration region with a frequency bandwidth lower than a preset bandwidth threshold and a signal duration exceeding a preset stability threshold is identified from the target signal image, then it is determined that single-tone interference exists in the region, and single-tone interference information is obtained; or, for time... Morphological feature detection is performed on the target signal image in the form of a frequency spectrum. If a straight line is detected in the target signal image, it is determined that the region where the straight line is located has single-tone interference, and single-tone interference information is obtained. The method for identifying constant-frequency interference information includes: performing power spectral density statistics on the target signal image within a preset time window to obtain a power spectrum; determining a background noise energy reference value based on the power spectrum; traversing the frequency points in the power spectrum, if there is a frequency point whose average power value within the preset time window is continuously higher than the background noise energy reference value, it is determined that constant-frequency interference exists at that frequency point, and constant-frequency interference information is obtained.

4. The method according to claim 2, characterized in that, The target signal image includes multiple interference regions; generating a binary mask image based on the image size of the target signal image and the interference regions in the target signal image includes: determining interference regions with a spatial distance less than a preset distance threshold as neighboring regions; merging the neighboring regions; performing a slight dilation process on the merged interference regions according to a preset dilation parameter to obtain an expanded interference region, wherein the expanded interference region is used to cover the edge of the interference region; and generating a binary mask image based on the image size of the target signal image and the expanded interference region in the target signal image.

5. The method according to claim 3, characterized in that, Before identifying multiple edge contours of the target signal in the magnified signal image and obtaining multiple first target regions on the magnified signal image, the method further includes: performing grayscale processing on the magnified signal image to obtain a grayscale magnified signal image; performing noise reduction processing on the grayscale magnified signal image; determining a first structuring element parameter based on the morphological features corresponding to the single-tone interference information; performing an opening operation on the noise-reduced magnified signal image based on the first structuring element parameter; determining a second structuring element parameter based on the preset magnification factor; and performing a closing operation on the opening magnified signal image based on the second structuring element parameter.

6. The method according to claim 5, characterized in that, Before identifying multiple edge contours of the target signal in the magnified signal image and obtaining multiple first target regions on the magnified signal image, the method further includes: determining a grayscale histogram of the magnified signal image after closing operation processing, wherein the horizontal axis of the grayscale histogram is the grayscale value of a pixel, and the vertical axis of the grayscale histogram is the number of pixels; determining multiple grayscale coordinate points based on the vertices of each vertical cylinder in the grayscale histogram, wherein the grayscale coordinate points correspond one-to-one with the vertices of the vertical cylinders; determining the maximum value point among the multiple grayscale coordinate points as the peak coordinate point; determining the peak coordinate point with the largest number of pixels among the multiple peak coordinate points as the background peak point, and further determining the multiple... The point with the largest grayscale value among the peak coordinate points is determined as the signal peak point; the line connecting the background peak point and the signal peak point is determined as the energy trend line; the grayscale coordinate points located between the grayscale values ​​of the background peak point and the signal peak point are determined as valid coordinate points; the valid coordinate point farthest from the energy trend line is determined as the target coordinate point; the grayscale value of the target coordinate point is determined as the grayscale threshold; the amplified signal image is converted from grayscale form to binary form according to the grayscale threshold, wherein the amplified signal image in binary form includes background points with pixel values ​​of the first type and signal points with pixel values ​​of the second type.

7. The method according to claim 6, characterized in that, The step of identifying multiple edge contours of the target signal in the magnified signal image to obtain multiple first target regions on the magnified signal image includes: performing a closing operation on the binary form of the magnified signal image according to the second structuring element parameters; extracting the signal contours from the binary form of the magnified signal image after the closing operation to obtain the contour information of the target signal, wherein each contour information corresponds to a detection box; selecting detection boxes from the multiple detection boxes whose aspect ratio is within a preset aspect ratio range and whose area is within a preset area range; and determining the region defined by the selected detection boxes in the magnified signal image as the first target region.

8. The method according to claim 7, characterized in that, The step of determining the region bounded by the selected detection boxes in the amplified signal image as the first target region includes: sorting the selected detection boxes in descending order according to their area to obtain the detection box arrangement order; initializing a retention set, wherein the retention set is initially empty; performing repeated detection on each detection box in the order of the detection boxes to obtain an updated retention set; and determining the region bounded by the detection boxes in the updated retention set in the amplified signal image as the first target region. The process of repeatedly detecting the current detection box includes: determining the intersection area between the current detection box and each retained detection box in the retention set; determining whether there is a retained detection box that satisfies condition one, wherein condition one includes: the intersection area between the retained detection box and the current detection box is equal to the area of ​​the current detection box, and the area of ​​the current detection box is less than or equal to the area of ​​the current detection box. The area of ​​the current detection frame is equal to the area of ​​the retained detection frame. If a retained detection frame satisfies condition one, the current detection frame is determined to be contained by the retained detection frames, the repeated detection of the current detection frame is completed, and the next detection frame is processed. If no retained detection frame satisfies condition one, the intersection-union ratio (IUR) of the current detection frame with each retained detection frame in the retained set is determined, and it is determined whether a retained detection frame satisfies condition two, wherein condition two includes: the IUR corresponding to the retained detection frame is greater than a preset IUR threshold. If a retained detection frame satisfies condition two, the overlap between the current detection frame and the retained detection frames exceeds a preset overlap threshold, the repeated detection of the current detection frame is completed, and the next detection frame is processed. If no retained detection frame satisfies condition two, the current detection frame is added to the retained set, and the next detection frame is processed.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the method described in any one of claims 1 to 8.

Images