Radio astronomy signal search method, device and medium based on YOLO

Abstract

The present application belongs to the technical field of radio astronomy, and particularly relates to a radio astronomy signal search method, device and medium based on YOLO, which can improve the detection capability of weak signals, and improve the overall operation speed and model training speed of the algorithm. The present application generates a dispersion quantity-time graph through non-coherent dispersion processing, and uses an improved YOLO model for target detection. The improved model introduces a windmill-shaped convolution, a differential attention mechanism and a content-guided attention fusion module on the basis of YOLO11, effectively improves the perception ability and anti-interference ability of the model to weak signals, and realizes efficient and high-precision search of radio astronomy signals.

Description

Technical Field

[0001] This invention belongs to the field of radio astronomy technology, specifically relating to a radio astronomy signal search method, device, and medium based on YOLO. Background Technology

[0002] Existing methods for detecting radio astronomy signals can be broadly categorized into two types. The first type is the traditional search method, which typically begins by mitigating radio interference using thresholding methods, then sets up a series of dispersion search grids to predict and compensate for the dispersion in the data. Next, it integrates along the frequency direction to obtain the dispersion-time series, calculates the signal-to-noise ratio (SNR) using matched filtering, and finally selects signals exceeding the SNR threshold as candidate signals. Typical tools include pulsar search software SIGPROC and Presto, and fast radio burst search software Heimdall. This type of method requires extensive manual checking, is inefficient, and prone to missed or false detections. It suffers from limitations in search speed and efficiency when dealing with massive amounts of observational data.

[0003] The second category comprises machine learning and deep learning methods. These methods extract various factors from raw data, such as dispersion-time, frequency-time, and pulse curves, and then use deep learning algorithms for feature extraction and detection. For example, the Drafts algorithm, currently used for fast radio burst detection by the FAST telescope, involves the following steps: first, performing incoherent dedispersion to generate a dispersion-time curve; then, using a CenterNet network to detect and identify the curve, extracting candidate signals; finally, substituting the candidate signals into the original signal for separate dedispersion processing, and then inputting it into a ResNet18 network for judgment. However, these machine learning algorithms rely on expert experience, require manually defined target features, have poor universality, and still have relatively long detection times. Some deep learning algorithms perform poorly in low signal-to-noise ratio environments, have long training times, and still have room for improvement in detection speed. Summary of the Invention

[0004] In view of this, the present invention provides a YOLO-based radio astronomy signal search method, device and medium, which can improve the detection capability of weak signals, while improving the overall running speed of the algorithm and the training speed of the model.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] A YOLO-based radio astronomy signal search method includes: Step 1: Perform incoherent dedispersion processing on the original radio astronomy signal data to obtain multiple dispersion quantity processing results; Step 2: Superimpose the processing results of each dispersion quantity along the frequency axis, and stitch together the superimposed results of all dispersion quantities to form a dispersion quantity-time diagram; Step 3: Input the dispersion-time map into the improved YOLO model for processing, and output the target confidence and bounding box position information to obtain the target dispersion and time delay information, thereby completing the signal detection. The improved YOLO model is based on the YOLO11 model and integrates a windmill-shaped convolution module, a differential attention module, and a content-guided attention fusion module.

[0007] In step one, the incoherent dedispersion processing is accelerated using the Numba.CUDA library and multiple dispersion results are processed in parallel using four threads.

[0008] The windmill-shaped convolution module adopts a modular design of parallel branch convolution and feature concatenation. Its convolution kernel has an outward diffusion trend and is used to replace the shallow convolution of the detect layer in the YOLO11 model to expand the receptive field of the model.

[0009] The calculation process of the windmill-shaped convolution module includes: performing parallel convolution operations in four directions on the input tensor to obtain four feature maps; concatenating the four feature maps; and normalizing the concatenated tensor using a 2×2 convolution kernel to obtain the output feature map.

[0010] The differential attention module replaces the original attention mechanism of the C2PSA module in the YOLO11 model. It splits the query vector and key vector into two sub-vectors, calculates the attention distribution for each, and subtracts the two to cancel out noise, thereby achieving feature focusing.

[0011] The output of the differential attention module is obtained by subtracting the second attention distribution scaled by the learnable parameter from the first attention distribution and then multiplying it by the value vector.

[0012] The content-guided attention fusion module is used to perform feature fusion in the Neck part of the YOLO model. It generates channel attention and spatial attention, and assigns a dedicated spatial importance map to each feature channel to highlight effective information and suppress interference.

[0013] The content-guided attention fusion module's processing steps include: calculating channel attention weights and spatial attention weights; adding the two together to obtain a coarse-grained spatial importance map; performing channel rearrangement and grouped convolution on the input features and the coarse-grained spatial importance map; and obtaining the final refined spatial importance map through the Sigmoid activation function.

[0014] The present invention also proposes a YOLO-based radio astronomy signal search device, including a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described in the present invention.

[0015] The present invention also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the present invention.

[0016] Beneficial effects (1) In the method of this invention, in terms of detection speed and training speed, since the model adopts the YOLO model, which has the characteristics of fast detection speed, fast training speed and suitability for real-time search, this invention adds some modules to the YOLO model, which basically does not affect its detection and training speed. In terms of dispersion reduction method, multi-threading is used for dispersion reduction processing, which also improves the overall algorithm running speed.

[0017] (2) In terms of detection accuracy, if the original YOLO11 model is used for detection, its detection accuracy is not high, and it is difficult to detect weak targets. By improving the YOLO model, its detection effect on weak targets can be effectively improved. Compared with other technologies mentioned above, this algorithm can still maintain high detection accuracy in low signal-to-noise ratio environments. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the overall algorithm of an embodiment of the present invention.

[0019] Figure 2 This is a schematic diagram of the complete structure of the improved YOLO model in an embodiment of the present invention.

[0020] Figure 3 This is a schematic diagram comparing the original time-frequency diagram, the detection results of the original YOLO11 algorithm, and the detection results of the improved algorithm in an embodiment of the present invention.

[0021] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0023] This invention provides a YOLO-based method, device, and medium for searching radio astronomical signals, aiming to address the problems of slow speed and low accuracy in detecting weak signals when processing massive amounts of data in existing radio signal search methods. This invention generates a dispersion-time map through incoherent dedispersion processing and utilizes an improved YOLO model for target detection. Based on YOLO11, this improved model introduces windmill-shaped convolution, differential attention mechanism, and content-guided attention fusion module, effectively enhancing the model's ability to perceive weak signals and its anti-interference capability, achieving efficient and high-precision search for radio astronomical signals.

[0024] This invention provides a YOLO-based radio astronomy signal search method that can efficiently and accurately search for radio astronomy signals. The algorithm implementation process is as follows: Figure 1 As shown, it includes the following steps: Step 1: Perform incoherent dedispersion processing on the original radio astronomy signal data to obtain multiple dispersion quantity processing results; Step 2: Superimpose the processing results of each dispersion quantity along the frequency axis, and stitch together the superimposed results of all dispersion quantities to form a dispersion quantity-time diagram; Step 3: Input the dispersion-time map into the improved YOLO model for processing, and output the target confidence and bounding box position information to obtain the target dispersion and time delay information, thereby completing the signal detection. The improved YOLO model is based on the YOLO11 model and integrates a windmill-shaped convolution module, a differential attention module, and a content-guided attention fusion module.

[0025] In step one, the dispersion value ranges from 10 to 1000 pc cm. - ³, step size is 1 pc cm - ³, a total of 990 dispersion quantity processing results were obtained.

[0026] In step one, the incoherent dedispersion processing is accelerated using the Numba.CUDA library and uses four threads to process multiple dispersion results in parallel, which reduces the processing time by about 40% compared to single-thread processing.

[0027] In step two, before inputting the dispersion-time map into the model, the image size is first converted to the YOLO standard of 640×640 pixels.

[0028] The windmill-shaped convolution module adopts a modular design of parallel branch convolution and feature splicing. Its convolution kernel has an outward diffusion trend and is used to replace the shallow convolution of the detect layer in the YOLO11 model to expand the model's receptive field and improve the ability to observe weak astronomical signals.

[0029] The calculation process of the windmill-shaped convolution module includes: performing parallel convolution operations in four directions on the input tensor to obtain four feature maps; concatenating the four feature maps; and normalizing the concatenated tensor using a 2×2 convolution kernel to obtain the output feature map.

[0030] The differential attention module replaces the original attention mechanism of the C2PSA module in the YOLO11 model. It splits the query vector and key vector into two sub-vectors, calculates the attention distribution for each, and subtracts the two to cancel out noise, thereby achieving feature focusing and generating a sparser and more focused attention map.

[0031] The output of the differential attention module is obtained by subtracting the second attention distribution scaled by the learnable parameter from the first attention distribution and then multiplying it by the value vector.

[0032] The content-guided attention fusion module is used to perform feature fusion in the Neck part of the YOLO model. It generates channel attention and spatial attention, and assigns a dedicated spatial importance map to each feature channel to highlight effective information, suppress astronomical signal interference, and reduce false alarms.

[0033] The content-guided attention fusion module's processing includes: calculating channel attention weights and spatial attention weights; adding them to obtain a coarse-grained spatial importance map; performing channel rearrangement and grouped convolution on the input features and the coarse-grained spatial importance map; and obtaining the final refined spatial importance map through a Sigmoid activation function. Specifically, the implementation process of each step in the method of this invention is as follows: Step 1: Incoherent dedispersion processing.

[0034] Dispersion is the most significant factor interfering with the detection of radio astronomy signals. When radio signals travel through the interstellar medium, the differences in group velocities of electromagnetic waves of different frequencies within the plasma cause the signals to be broadened and smoothed, becoming submerged in noise. The signal's time delay Δt exhibits a quadratic relationship with its frequency.

[0035] In this embodiment, an incoherent dedispersion method is employed. First, a step size of 1 pc cm is used. - ³, for values ​​ranging from 10 to 1000 pc cm - The dispersion measure of 3 is taken to obtain 990 dispersion measure processing results. This process is accelerated using the Numba.CUDA library. Since the superposition operation of each dispersion measure is relatively independent, four threads are used to process multiple dispersion measure results simultaneously, which reduces the time consumption by about 40% compared to single-threaded processing.

[0036] Step 2: Create a dispersion-time plot.

[0037] The dispersion processing results obtained in step one are superimposed along the frequency axis, and then all the superimposed dispersion results are stitched together to form a dispersion-time map. Subsequently, each image is converted to the YOLO common image size (640×640) for input into the model for processing.

[0038] Step 3: Improve YOLO model detection.

[0039] The image processed in step two is then fed into the improved YOLO model for further processing to obtain the target's confidence level and bounding box position information, thereby acquiring the target's dispersion and time delay information and completing the signal detection.

[0040] This invention is based on and improves the YOLO11 deep learning model. To enhance its ability to detect weak radio astronomy signals, three modules are introduced into the YOLO11 model: pinwheel-shaped convolution, differential attention, and content-guided attention fusion (CGAFusion). The complete structure of the improved YOLO model is as follows: Figure 2 As shown.

[0041] The windmill-shaped convolution module is a modular design consisting of parallel branch convolution, feature concatenation, and normalization. Unlike conventional convolution, this module uses asymmetric padding, generating convolution kernels in different directions for different regions of the image, with a distribution that spreads outwards. Let the input tensor be X^(h1,w1,c1), where h1, w1, and c1 represent the height, width, and number of channels of the input tensor, respectively. The first layer of the windmill-shaped convolution performs parallel convolution operations, followed by concatenation of the four feature maps. Finally, a 2×2 convolution kernel with c2 channels is used to normalize the concatenated tensor. The receptive field effectiveness of this module decays from the center to the edges, exhibiting a Gaussian distribution, which is more suitable for small target observation tasks. In this invention, the windmill-shaped convolution module replaces the shallow convolution of the YOLO11 model's detect layer to expand the overall receptive field of the model.

[0042] The Differential Attention module addresses the "attention noise" problem in the Transformer attention mechanism. Its core idea is to learn two attention maps and subtract them to cancel out misalignments or noise signals. This module first splits the query vector Q and key vector K of the input x into two sub-vectors Q1, Q2 and K1, K2 respectively. It then calculates the attention distribution for each sub-vector, and the final output is obtained by subtracting the second attention distribution (scaled by the learnable parameter λ) from the first attention distribution.

[0043] The parameter λ is optimized through training. This invention integrates the DiffAttention differential attention mechanism into the C2PSA module of the YOLO11 model to replace its original attention mechanism, achieving feature focusing effect without increasing detection time or the number of parameters.

[0044] The content-guided attention fusion module mainly consists of two parts: channel attention and spatial attention. Channel attention weights Wc and spatial attention weights are then fused using a formula to obtain a coarse-grained spatial importance map.

[0045] Finally, the input features and the coarse-grained spatial importance map are rearranged in an alternating order to obtain the final refined spatial importance map W.

[0046] In this invention, three additional CGAFusion modules are added. Through this method, the model can effectively suppress astronomical signal interference and reduce false alarms.

[0047] In terms of detection and training speed, the method of this invention utilizes the YOLO model, which features fast detection and training speeds and is suitable for real-time search. Adding the aforementioned modules to the YOLO model has minimal impact on its detection and training speed. For the dispersion reduction method, multi-threading is employed for dispersion reduction, further improving the overall algorithm's running speed. Regarding detection accuracy, such as... Figure 3 As shown, the first image is the original time-frequency plot of the signal, the second image is the detection result of the original YOLO11 algorithm on the dispersion-time plot, and the third image is the detection result of the improved algorithm. It can be seen that by improving the YOLO model, its detection performance for weak targets can be effectively improved, and it can still maintain high detection accuracy in low signal-to-noise ratio environments.

[0048] In this embodiment, firstly, a step size of 1pc cm⁻³ is used to take values ​​of dispersion in the range of 10-1000pc cm⁻³. Incoherent dedispersion is performed on the original data according to the dispersion, resulting in 990 dispersion processing results. Each processing result is then superimposed along the frequency axis, and all dispersion values ​​are stitched together to form a dispersion-time map. This process is accelerated using the numba.cuda library [ ]. Since the superposition operation of each dispersion is relatively independent, four threads are used to process multiple dispersion results simultaneously, which reduces the processing time by about 40% compared to single-thread processing. Then, each image is converted to the YOLO common image size (640×640), and finally, the converted image is placed into the improved YOLO model for processing to obtain the target confidence and bounding box position information, thereby obtaining the target dispersion and time delay information. The signal detection work is completed.

[0049] Dispersion is the most significant factor interfering with the detection of radio astronomy signals. Radio signals in the universe undergo some degree of dispersion as they pass through the interstellar medium. This is due to the difference in group velocities of electromagnetic waves of different frequencies in plasma. This phenomenon causes astronomical signals to be broadened and smoothed, becoming submerged in noise and thus difficult to detect. The signal delay Δt has a quadratic function relationship with frequency, as shown in the following formula: (1) Where c is the speed of light, d is the propagation distance, and f is the propagation distance. p Here, f1 represents the plasma frequency, approximately 1500 Hz, and f2 represents the receiver's start and stop frequencies. The dedispersion method used in this invention is a non-coherent dedispersion method based on a series of dispersion search grids. This involves predicting and compensating for the dispersion of the data to restore the target signal's signal-to-noise ratio and pulse shape, thus aiding in subsequent detection. During dedispersion compensation, Numba.CUDA's multi-threaded processing mode is employed to simultaneously process multiple dispersion data points, thereby improving the overall detection speed.

[0050] This invention employs the YOLO11 deep learning model for detection. To enhance its ability to detect weak radio astronomy signals, this invention introduces three modules into the YOLO11 model: windmill convolution, CGAFusion (Content-Guided Attention Fusion), and Diffattention. This improves the model's ability to detect weak targets while maintaining the inherent high detection and training speed of the YOLO11 model. The windmill-shaped convolution module employs a modular design of "parallel branch convolution + feature concatenation + normalization." Unlike conventional convolution, this module uses an asymmetric padding method to generate different features for different regions of the image. Convolutional kernels in the same direction, with an outward diffusion trend, let the input tensor be... , where h1,w 1, c1 represents the height, width, and number of channels of the input tensor, respectively. The first layer of the windmill-shaped convolution performs parallel convolution operations, and its calculation process is as follows: (2) (3) (4) (5) in This represents a 1×3 convolutional kernel with output channel c', where the padding parameter P(1,0,0,3) represents the number of padding pixels in the four directions (up, down, left, right). After the first convolutional layer, the height of the output feature is... ,width Number of channels , This represents the final output feature count of the PConv module. The four feature maps are then concatenated, and the calculation process is as follows: ... (6) Finally, a 2×2 convolutional kernel with c² channels is used to normalize the concatenated tensor, adjusting the height and width of the output feature map to preset values. This allows PConv to be directly interchangeable with standard convolutional layers. This module also serves as a channel attention mechanism to analyze the feature contribution of different convolutional directions. The calculation formula is as follows: (7) Its receptive field effectiveness decays from the center to the edge, exhibiting a Gaussian distribution. The smaller the target size, the more concentrated the feature distribution, making it more suitable for small target observation tasks. Furthermore, through grouped convolution techniques, the number of parameters is effectively controlled while expanding the receptive field. In this invention, a windmill-shaped convolution module replaces the shallow convolution of the YOLO11 model's detect layer. This method expands the overall receptive field of the model, improving its ability to observe weak astronomical signals.

[0051] DiffAttention is an improvement on the "attention noise" problem in the Transformer attention mechanism. Its core idea is to learn two attention maps and subtract them to cancel out misaligned or noisy signals, achieving precise feature focusing. The DiffAttention module first splits the query vector Q and key vector K of the input x into two sub-vectors, denoted as Q1, Q2 and K1, K2 respectively, where Q1, Q2, K1, and K2 all belong to R... N×2d In the space, V is used as a value vector with half the number of heads but double the feature dimension to match the split of Q and K. The attention distribution of the two sub-vectors is calculated separately: (8) (9) In the denominator This is to reduce the variance of the dot product result and avoid numerical explosion caused by excessively high vector dimensions. The final output is obtained by subtracting the second attention distribution, scaled by the learnable parameter λ, from the first attention distribution: (10) The parameter λ is optimized through training to control the subtraction strength of the second attention distribution, and its calculation method is as follows: (11) in , , , For learnable parameters, For hyperparameters, The magnitude of this value directly determines the effectiveness of differential attention, which can be dynamically adjusted. To achieve a balance between noise cancellation and signal preservation. Because and While the attention weights for noise are roughly the same, they differ significantly when dealing with key areas. Subtracting these weights can cancel out the noise's influence, preserving or strengthening the attention weights for key areas. This results in a sparser, more focused attention map, improving model performance in scenarios susceptible to background or redundant signal interference. This invention integrates the DIffAttention differential attention mechanism into the C2PSA module of the YOLO11 model, replacing its original attention mechanism. Its lightweight architecture allows the module to achieve feature focusing without increasing detection time or the number of parameters.

[0052] The CGAFusion module mainly consists of two parts: channel attention and spatial attention. It generates a dedicated spatial importance map for each channel of the input feature in both the channel and spatial dimensions, respectively. The channel attention weight W... c Spatial attention weight W s The calculation formula is as follows: (12) in, Represents the ReLU activation function. This represents the size of the convolutional layer used. , , These represent the features obtained after global average pooling in the spatial dimension, global average pooling in the channel dimension, and global max pooling in the channel dimension, respectively. They are then fused using a formula to obtain a coarse-grained spatial importance map. (13) Finally, the input features and Rearrange them in alternating order to obtain the final refined spatial importance map W. (14) in It is the Sigmoid activation function. ( ) represents a channel rearrangement operation. This represents the use of grouped convolutional layers with a kernel size of k. CGA assigns a dedicated spatial importance map to each feature channel, guiding the model to focus on the key regions of each channel, thereby highlighting the effective information contained in the features and significantly improving dehazing performance. In this invention, three additional CGAFusion modules are added, which differs from the original YOLO model that directly outputs the Neck part through the C3k2 layer. Through this method, the model can effectively suppress astronomical signal interference and reduce false alarms.

[0053] As can be seen, this invention uses the YOLO model as the basic detection framework, inheriting its characteristics of fast detection and training speed, making it suitable for real-time search scenarios. Furthermore, by employing multi-threaded parallel processing during the dispersion elimination process, the overall algorithm's running speed is further improved. By integrating three modules—windmill convolution, differential attention, and content-guided attention fusion—into the YOLO11 model, the model's ability to detect weak signals is significantly enhanced without significantly impacting detection and training speed. In low signal-to-noise ratio environments, compared to the original YOLO11 model and other traditional algorithms, this invention achieves higher detection accuracy and can effectively detect previously difficult-to-detect weak radio astronomy signals.

[0054] This application also provides an electronic device. Figure 4The structure of an electronic device provided in an embodiment of the present invention is illustrated. For example, the electronic device 40 may include a processor 41, a memory 42, and a transmission device 43. The processor 41 is used to execute the YOLO-based radio astronomy signal search method mentioned in the above embodiments. The processor and the memory can be connected via a bus or other means, taking a bus connection as an example. The transmission device can be connected to the processor and the memory via wired or wireless means. The memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the YOLO-based radio astronomy signal search method in the embodiments of this application. The processor executes various functional applications and data processing by running the non-transitory software programs, instructions, and modules stored in the memory, thereby implementing the YOLO-based radio astronomy signal search method in the above method embodiments. The memory may include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a function; the data storage area may store data created by the processor, etc. Furthermore, the memory may include high-speed random access memory and non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory may optionally include memory remotely located relative to the processor, which can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof. The one or more modules are stored in the memory and, when executed by the processor, perform the YOLO-based radio astronomy signal search method in the embodiments.

[0055] In another aspect, this application also provides a computer-readable storage medium, which may be the computer-readable storage medium included in the apparatus described in the above embodiments; or it may be a standalone computer-readable storage medium not assembled into the device. The computer-readable storage medium may be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, register, floppy disk, hard disk, removable storage disk, CD-ROM, or any other form of storage medium known in the art. The computer-readable storage medium stores one or more programs, which are used by one or more processors to execute the YOLO-based radio astronomy signal search method described in this application.

[0056] This invention includes, but is not limited to, the above embodiments. Any equivalent substitutions or partial improvements made under the spirit and principles of this invention shall be considered within the scope of protection of this invention.

Claims

1. A YOLO-based radio astronomy signal search method, characterized in that, include: Step 1: Perform incoherent dedispersion processing on the original radio astronomy signal data to obtain multiple dispersion quantity processing results; Step 2: Superimpose the processing results of each dispersion quantity along the frequency axis, and stitch together the superimposed results of all dispersion quantities to form a dispersion quantity-time diagram; Step 3: Input the dispersion-time map into the improved YOLO model for processing, and output the target confidence and bounding box position information to obtain the target dispersion and time delay information, thereby completing the signal detection; The improved YOLO model is based on the YOLO11 model and integrates a windmill-shaped convolution module, a differential attention module, and a content-guided attention fusion module.

2. The method according to claim 1, characterized in that, The incoherent dedispersion processing in step one is accelerated using the Numba.CUDA library, and multiple dispersion results are processed in parallel using four threads.

3. The method according to claim 1, characterized in that, The windmill-shaped convolution module adopts a modular design of parallel branch convolution and feature concatenation. Its convolution kernel has an outward diffusion trend and is used to replace the shallow convolution of the detect layer in the YOLO11 model to expand the receptive field of the model.

4. The method according to claim 5, characterized in that, The calculation process of the windmill-shaped convolution module includes: performing parallel convolution operations in four directions on the input tensor to obtain four feature maps; concatenating the four feature maps; and normalizing the concatenated tensor using a 2×2 convolution kernel to obtain the output feature map.

5. The method according to claim 1, characterized in that, The differential attention module is used to replace the original attention mechanism of the C2PSA module in the YOLO11 model. It achieves feature focusing by splitting the query vector and key vector into two sub-vectors, calculating the attention distribution separately, and subtracting the two to cancel out noise.

6. The method according to claim 7, characterized in that, The output of the differential attention module is obtained by subtracting the second attention distribution scaled by the learnable parameters from the first attention distribution, and then multiplying it by the value vector.

7. The method according to claim 1, characterized in that, The content-guided attention fusion module is used to perform feature fusion in the Neck part of the YOLO model. It generates channel attention and spatial attention, and assigns a dedicated spatial importance map to each feature channel to highlight effective information and suppress interference.

8. The method according to claim 9, characterized in that, The processing steps of the content-guided attention fusion module include: calculating channel attention weights and spatial attention weights; adding the two together to obtain a coarse-grained spatial importance map; performing channel rearrangement and group convolution on the input features and the coarse-grained spatial importance map, and obtaining the final refined spatial importance map through the Sigmoid activation function.

9. A YOLO-based radio astronomy signal search device, characterized in that, It includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the method of any one of claims 1 to 10.

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

Images