A method for determining the type of drone based on sound recognition

Abstract

The present invention relates to a method for determining the type of drone based on sound recognition, and to the technical field of drone type identification. In the method, in response to a collected audio signal, a first audio signal feature is obtained after preprocessing, and the first audio signal feature is respectively subjected to MFCC feature extraction to obtain a second audio signal feature as an input feature and a GFCC feature extraction to obtain a third audio signal feature. The input features are adaptively fused in the spatial dimension, and the input features are adaptively adjusted in the channel dimension, so that the synergy of the two improves the feature expression ability and adaptability of drone type identification. In addition, the global information processing ability and feature representation ability are significantly improved through feature enhancement, thereby further improving the ability of drone type identification.

Description

Technical Field

[0001] The present invention relates to the technical field of drone type identification, and in particular to a drone type identification method based on sound recognition. Background Art

[0002] At present, there are many methods for identifying drones, such as radar, radio detection, video image detection, and sound detection. Sound recognition has the advantages of low equipment cost, easy miniaturization and integration, and easy combination with other methods, so it is widely used in the civilian field. There are some drone identification methods based on acoustic signals, such as the traditional acoustic signal recognition methods, such as the use of shallow machine learning models such as Gaussian mixture models and support vector machines. However, the recognition performance of these methods needs to be improved when facing complex acoustic fields such as long-distance propagation. In recent years, deep learning technology has made breakthrough progress in the field of sound recognition. Deep learning models such as convolutional neural networks (CNN) and long short-term memory networks (LSTM) can automatically extract and learn high-level feature representations from raw noise data, significantly improving the performance of tasks such as speech recognition and speaker recognition.

[0003] The principle of drone noise recognition based on deep learning can be summarized as the following steps: first, use the sound sensor to collect drone noise signals and perform preprocessing, then extract acoustic features, such as time-frequency domain features, and then use deep learning models such as CNN and LSTM to perform high-level representation learning and classification decisions on the features, and finally use the recognition results for drone target tracking, threat assessment and other tasks. Compared with traditional recognition methods, deep learning-based methods can fully explore and utilize the rich pattern information of sound signals, have stronger feature representation and modeling capabilities, and thus can achieve higher recognition accuracy and environmental adaptability. However, there is still a problem of low recognition accuracy. Summary of the invention

[0004] The technical problem to be solved by the present application is to provide a method for determining the type of drone based on sound recognition, which has the characteristic of improving the ability to identify the type of drone.

[0005] A method for determining the type of a drone based on sound recognition provided in one embodiment includes:

[0006] Collect the audio signal of the drone;

[0007] In response to the collected audio signal, preprocessing is performed to obtain a first audio signal feature, and MFCC feature extraction and GFCC feature extraction are performed on the first audio signal feature to obtain a second audio signal feature and a third audio signal feature;

[0008] The second audio signal feature and the third audio signal feature are used as inputs of a drone type recognition network to obtain the type of drone to which the collected audio signal belongs, including:

[0009] Performing 7×7 convolution processing on the second audio signal feature to obtain a first feature map, performing 7×7 convolution processing on the third audio signal feature to obtain a second feature map; performing feature fusion on the first feature map and the second feature map to obtain a third feature map;

[0010] Adaptively fusing the features of the third feature map in the spatial dimension to obtain a sixth feature map;

[0011] Adaptively adjusting the features of the sixth feature map in the channel dimension to obtain a seventh feature map;

[0012] Performing feature enhancement processing on the seventh feature map to obtain an eighth feature map;

[0013] Determine the type of the drone based on the eighth characteristic graph;

[0014] The step of adaptively fusing the features of the third feature map in the spatial dimension to obtain the sixth feature map includes:

[0015] Performing a first multi-scale feature extraction on the third feature map to obtain a multi-scale feature map;

[0016] Performing self-correction based on the third feature map and the multi-scale feature map, and obtaining a first fusion weight and a second fusion weight based on the self-correction result;

[0017] Multiplying the first fusion weight by the first feature map to obtain a fourth feature map, and multiplying the second fusion weight by the second feature map to obtain a fifth feature map;

[0018] The fourth feature map and the fifth feature map are feature-fused to obtain a sixth feature map.

[0019] The beneficial effects of the present invention are:

[0020] Since the input features are adaptively fused in the spatial dimension and adaptively adjusted in the channel dimension, the synergy of the two improves the feature expression and adaptive capabilities of drone type recognition. In addition, feature enhancement significantly improves the global information processing and feature representation capabilities, thereby further improving the drone type recognition capability. Among them, in the adaptive fusion of input features in the spatial dimension, self-correction is performed, and fusion weights are obtained based on the self-correction results to achieve feature fusion, which helps to form self-correction weight information, highlight the main feature information, and suppress secondary information, thereby greatly improving the drone type recognition capability. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a schematic diagram of a method for determining the type of a drone according to an embodiment of the present application;

[0022] Figure 2 This application Figure 1 A schematic flow chart of an embodiment of step S30;

[0023] Figure 3 This application Figure 2 A flowchart of an embodiment of step S302;

[0024] Figure 4 This application Figure 3 A flowchart of an embodiment of step S3021 in FIG.

[0025] Figure 5 This application Figure 3 A flowchart of an embodiment of step S3022;

[0026] Figure 6 This application Figure 2 A flowchart of an embodiment of step S303;

[0027] Figure 7 This is a schematic diagram of a process flow of any level feature enhancement processing in one embodiment of the present application;

[0028] Figure 8 This application Figure 7 A flowchart of an embodiment of step S100 in FIG. DETAILED DESCRIPTION

[0029] The present invention is further described in detail below by specific embodiments in conjunction with the accompanying drawings. Wherein similar elements in different embodiments adopt associated similar element numbers. In the following embodiments, many detailed descriptions are for making the present application better understood. However, those skilled in the art can easily recognize that some features can be omitted in different situations, or can be replaced by other elements, materials, methods. In some cases, some operations related to the present application are not shown or described in the specification, this is to avoid the core part of the present application being overwhelmed by too much description, and for those skilled in the art, it is not necessary to describe these related operations in detail, and they can fully understand the related operations according to the description in the specification and the general technical knowledge in the art.

[0030] In addition, the features, operations or characteristics described in the specification can be combined in any appropriate manner to form various implementations. At the same time, the steps or actions in the method description can also be interchanged or adjusted in a manner that is obvious to those skilled in the art. Therefore, the various sequences in the specification and the drawings are only for the purpose of clearly describing a certain embodiment and are not meant to be a required sequence, unless otherwise specified that a certain sequence must be followed.

[0031] The serial numbers assigned to the components herein, such as "first", "second", etc., are only used to distinguish the objects described and do not have any order or technical meaning. The "connection" and "coupling" mentioned in this application, unless otherwise specified, include direct and indirect connections (couplings).

[0032] To facilitate the description of the inventive concept of the present application, the following is a brief description of drone sound recognition technology based on deep learning.

[0033] In the current scheme, the method of identifying drones based on sound signals extracts the Log-Mel features of drone sound signals as the noise features of noise samples. For the deep learning model, the IRBNet model with IRTBlock-A module and IRTBlock-B module is selected. This model has higher recognition accuracy than the traditional network model.

[0034] However, the applicant found in the research that although the above scheme can effectively identify drones, it is difficult to adaptively adjust the weights of various features, so that it can only focus on a single feature, resulting in a low ability to identify the type of drone.

[0035] In view of this, an embodiment of the present application provides a method for determining the type of drone based on sound recognition. In this method, on the one hand, the input features are adaptively fused in the spatial dimension, and on the other hand, the input features are adaptively adjusted in the channel dimension. The synergistic effect of the two improves the feature expression ability and adaptive ability of drone type identification. In addition, the global information processing ability and feature representation ability are significantly improved through feature enhancement, thereby further improving the ability of drone type identification. Among them, in the adaptive fusion of input features in the spatial dimension, the fusion weight is obtained through self-correction to achieve feature fusion, which helps to form self-correction weight information, highlight the main feature information, and suppress secondary information, thereby greatly improving the drone type identification ability.

[0036] A method for determining the type of drone based on sound recognition is provided in one embodiment of the present application. Please refer to Figure 1 ,include:

[0037] Step S10, collecting the audio signal of the drone.

[0038] The sound signal containing the drone audio signal is collected, and the type of the drone is identified based on the sound signal. The collected sound signal is filtered to obtain the denoised drone audio signal.

[0039] Step S20, in response to the collected audio signal, preprocessing is performed to obtain a first audio signal feature, and MFCC (Mel Frequency Cepstral Coefficient) feature extraction is performed on the first audio signal feature to obtain a second audio signal feature and GFCC (Gamma Frequency Cepstral Coefficient) feature extraction is performed to obtain a third audio signal feature.

[0040] In one embodiment, preprocessing the collected audio signal to obtain the first audio signal feature includes:

[0041] First, the collected audio signal is pre-emphasized through a high-pass filter.

[0042] Secondly, the audio signal after pre-emphasis processing is divided into frames so that a section of audio signal can be divided into multiple frames. Among them, since the audio signal is constantly changing with time, it is not convenient to study, so a section of audio is divided into multiple frames, and the signal after the framing operation can be considered to be stable. In one embodiment, in order to make the junction between two adjacent frames not abrupt, frame shift is also introduced. In one embodiment, the frame length is 2048 and the frame shift is 512. In this way, when the sampling rate of the sound is 44.1Khz, the duration of each frame is about 0.046 seconds.

[0043] Thirdly, the framed audio signal is subjected to windowing processing. The purpose of windowing processing is to reduce the spectrum leakage caused by signal truncation, improve the accuracy of spectrum analysis, and make the signal transition smoothly at the frame boundary to reduce the discontinuity caused in the frame division process.

[0044] Finally, the audio signal after windowing is subjected to a fast Fourier transform to obtain the first audio signal feature after preprocessing. The purpose of the Fourier transform is to transform the windowed signal into the frequency domain, because it is difficult to see the audio signal feature in the time domain. After the Fourier transform, more abundant information about the audio signal can be obtained. The absolute value of the Fourier transformed signal is taken and squared to obtain the power spectrum of the signal, which is the first audio signal feature.

[0045] In one embodiment, MFCC feature extraction is performed on the first audio signal feature to obtain the second audio signal feature , extract the first audio signal feature by GFCC feature to obtain the third audio signal feature .

[0046] Since the Mel-frequency cepstral coefficient features and the gamma-frequency cepstral coefficient features extract different aspects of the sound signal through the Mel filter bank and the gamma-frequency filter bank respectively, the fused features can more comprehensively characterize the pattern of the drone sound, thereby improving the recognition performance.

[0047] Step S30: Using the second audio signal feature and the third audio signal feature as inputs to a drone type recognition network to obtain the type of drone to which the collected audio signal belongs.

[0048] In one embodiment, please refer to Figure 2 , step S30 may include:

[0049] Step S301, performing 7×7 convolution processing on the second audio signal feature to obtain a first feature map, performing 7×7 convolution processing on the third audio signal feature to obtain a second feature map; and performing feature fusion on the first feature map and the second feature map to obtain a third feature map.

[0050] In one embodiment, the input second audio signal feature and the third audio signal feature are respectively subjected to preliminary feature extraction by 7×7 convolution processing with a step size of 2 to obtain a first feature map And the second feature map Then the first feature map And the second feature map Perform feature fusion to obtain the third feature map ,Right now .

[0051] Through step S301, the advantages of the two feature maps, MFCC and GFCC, obtained after the convolution operation are effectively integrated.

[0052] Step S302, adaptively fuse the features of the third feature map in the spatial dimension to obtain a sixth feature map.

[0053] In one implementation, please refer to Figure 3 , step S302 may include:

[0054] Step S3021: Perform a first multi-scale feature extraction on the third feature map to obtain a multi-scale feature map.

[0055] In one embodiment, please refer to Figure 4 , step S3021 may include:

[0056] Step S30211, after average pooling is performed on the third feature map, the feature map after average pooling is mapped to the original feature space by upsampling to obtain a first multi-scale feature map.

[0057] In one embodiment, a filter with a step size of 2 and a size of 2×2 is used for average pooling, and then a 3×3 bilinear interpolation upsampling operation is performed, and then a normalization operation is performed to obtain a first multi-scale feature map. Therefore, step S30211 can be expressed as:

[0058] ,

[0059] in, represents the first multi-scale feature map, represents the normalization operation, represents bilinear interpolation upsampling, represents average pooling, Represents a predefined 2×2 filter.

[0060] Step S30212, after performing convolution processing on the third feature map, normalization operation is performed to obtain a second multi-scale feature map.

[0061] In one embodiment, after performing 3×3 convolution processing on the third feature map, a normalization operation is performed to obtain a second multi-scale feature map of the same size as the first multi-scale feature map. .

[0062] Through step S3021, based on average pooling and convolution operations, two feature maps of different scales are generated. Together with the original input features, three feature maps of different scales are included. In this way, features are extracted from different receptive fields and resolutions to capture different levels of feature information.

[0063] Step S3022: Perform self-correction based on the third feature map and the multi-scale feature map, and obtain a first fusion weight and a second fusion weight based on the self-correction result.

[0064] In one embodiment, please refer to Figure 5 , step S3022 includes:

[0065] Step S30221, performing a self-correction operation on the third feature map, the first multi-scale feature map, and the second multi-scale feature map to obtain a self-correction result, which can be expressed as:

[0066] ,

[0067] in, represents the self-calibration result, represents element-wise multiplication, represents element-by-element addition, Represents the activation function.

[0068] In the embodiment of the present application, the third characteristic diagram As a residual, it helps to form self-correction weight information. and the first multi-scale feature map after resizing Add them together, and then generate a gating signal through a sigmoid activation function. This gating signal plays a role in adjusting the degree of fusion of features of different scales, which is equivalent to an adaptive switch. The sigmoid activation function maps the addition result to between 0 and 1 as the fusion weight coefficient.

[0069] The fuser combines the gated signal with the second multi-scale feature map Perform element-by-element multiplication to obtain the self-corrected result Through this operation, the gating signal selectively modulates the second multi-scale feature map The role of the characteristics.

[0070] Through step S30221, the gated signal is applied to the third feature map Adaptive feature selection and modulation are performed to obtain a new feature map , the third feature map The important features are highlighted and the minor features are suppressed.

[0071] Step S30222, the self-correction result is processed by convolution and activation function in sequence to obtain the total fusion weight, which can be expressed as:

[0072] ,

[0073] in, represents the total fusion weight, Represents the filter bank of the convolution process.

[0074] In order to further optimize the fusion process, the self-correction results Further convolution and sigmoid activation operations give the total fusion weight The total fusion weight It is equivalent to an adaptive fusion strategy, which guides the weight distribution of different features during fusion.

[0075] Step S30223, mapping the total fusion weight to the fusion weight dimension, and calculating the first fusion weight and the second fusion weight.

[0076] In one embodiment, step S30223 may include:

[0077] ,

[0078] in, represents the first fusion weight, Represents the second fusion weight.

[0079] Step S3023: multiply the first fusion weight by the first feature map to obtain a fourth feature map, and multiply the second fusion weight by the second feature map to obtain a fifth feature map.

[0080] In one embodiment, step S304 includes:

[0081] ,

[0082] in, represents the fourth feature map, represents the fifth feature map.

[0083] Step S3024: perform feature fusion on the fourth feature map and the fifth feature map to obtain a sixth feature map.

[0084] In one embodiment, step S3024 can be expressed as:

[0085] ,

[0086] in, represents the sixth feature map.

[0087] Through step S3023 and step S3024, the original feature map is fused using the fusion weights. and The weighted fusion is performed to obtain the final fused sixth feature map. Here, based on the fusion weight, the features from different sources are adaptively fused according to their importance, and a more discriminative and complementary feature representation is obtained.

[0088] Through step S302, the fusion weights are adaptively fused on the features in the spatial dimension, and the fusion weights of different feature maps can be automatically learned and adjusted to highlight the features with strong expressiveness and suppress the redundant or noisy features. This adaptive weight strategy makes the fusion features more refined and effective, reduces the interference of noise, and thus improves the discrimination of the fusion features. High-quality fusion features are good inputs for subsequent recognition links, which can ease the recognition difficulty and improve the recognition accuracy.

[0089] Step S303: adaptively adjust the features of the sixth feature map in the channel dimension to obtain a seventh feature map.

[0090] In one embodiment, please refer to Figure 6 , step S303 may include:

[0091] Step S3031, perform global average pooling on the sixth feature map. Through global average pooling, the spatial dimension information of each channel of the sixth feature map is compressed into a scalar value with a width and a height of 1. For example, assuming that the shape of the sixth feature map is (batch_size, channels, height, width), where batch_size represents the number of elements, channels represents the number of channels, height represents the height, and width represents the width, then after global average pooling, the shape of the sixth feature map becomes (batch_size, channels, 1, 1).

[0092] Step S3032, performing a first dimension adjustment on the sixth feature map after pooling, including first removing the spatial dimension, and then exchanging the channel dimension and the last dimension.

[0093] Then through step S3032, (batch_size, channels, 1, 1) can be adjusted to (batch_size, 1, channels).

[0094] Step S3033: Perform one-dimensional channel convolution on the sixth feature map after the first dimension adjustment, so that the convolution kernel slides on the channel dimension.

[0095] The applicant found in the research that in order to flexibly capture the dependencies between channels, the computational efficiency will decrease.

[0096] In view of this, in order to flexibly capture the dependencies between channels while taking into account computational efficiency, the size of the convolution kernel is set based on the number of channels in the feature map and combined with hyperparameters.

[0097] In one embodiment, the calculation formula for the convolution kernel size for performing one-dimensional channel convolution can be expressed as:

[0098] ,

[0099] Among them, k represents the convolution kernel size, c represents the number of channels, represents the logarithm calculation with base 2, and b represent hyperparameters, and odd represents rounding down.

[0100] In one embodiment, a can be set to 2 and b can be set to 1.

[0101] The convolution kernel size calculated based on the above formula can enable the attention mechanism to play a role in different network structures and tasks, thereby improving the network's representation ability and performance.

[0102] Step S3034, performing a second dimension adjustment on the sixth feature map after the one-dimensional channel convolution so that the dimension is adjusted to the dimension of the sixth feature map after pooling.

[0103] Based on step S3032, the second dimension adjustment includes: firstly exchanging the channel dimension and the last dimension, and the shape becomes (batch_size, channels, 1). Then adding a new dimension to the last dimension, and the shape becomes (batch_size, channels, 1, 1).

[0104] Step S3035, perform sigmoid activation processing on the sixth feature map after the second dimension adjustment to obtain the attention weight of each channel.

[0105] Step S3036, multiply the obtained attention weight of each channel by the sixth feature map to obtain the seventh feature map.

[0106] Through step S303, the importance of the channel can be adaptively adjusted to enhance the expression of the feature channel with strong judgment ability, so that the optimized features are more refined in the channel dimension, which helps to highlight the discriminative feature patterns, thereby improving the recognition accuracy.

[0107] Step S304: perform feature enhancement processing on the seventh feature map to obtain an eighth feature map.

[0108] The applicant found in the research that the current feature enhancement processing focuses on local features and lacks explicit modeling of the global context, resulting in low recognition ability of drone types.

[0109] In view of this, in one embodiment of the present application, in step S304, the seventh feature map is used as input, and multi-level feature enhancement processing is performed to obtain an eighth feature map; wherein, except for the first-level feature enhancement processing, the output of the previous-level feature enhancement processing is used as the input of the current-level feature enhancement processing.

[0110] In one embodiment, please refer to Figure 7 , any level of feature enhancement processing includes:

[0111] Step S100, performing second multi-scale feature extraction on the input feature map, and splicing the extracted multi-scale feature maps.

[0112] In one embodiment, please refer to Figure 8 , step S100 includes:

[0113] Step S1001, performing 1×1 convolution processing on the input feature map to obtain a third multi-scale feature map.

[0114] In step S1001, a 1×1 convolution is performed on the input feature map to reduce the number of channels, thereby obtaining a third multi-scale feature map.

[0115] Step S1002: Perform 1×1 convolution processing and then 3×3 convolution processing on the input feature map to obtain a fourth multi-scale feature map.

[0116] In step S1002, a 1×1 convolution is performed on the input feature map to reduce the number of channels, and then a 3×3 convolution is performed on the result of the dimensionality reduction to increase the number of channels, thereby obtaining a fourth multi-scale feature map.

[0117] Step S1003: Perform 1×1 convolution processing and then 5×5 convolution processing on the input feature map to obtain a fifth multi-scale feature map.

[0118] In step S1003, a 1×1 convolution is performed on the input feature map to reduce the number of channels, and then a 5×5 convolution is performed on the result of the dimensionality reduction to increase the number of channels, thereby obtaining a fifth multi-scale feature map.

[0119] Step S1004, performing maximum pooling processing and then 1×1 convolution processing on the input feature map to obtain a sixth multi-scale feature map.

[0120] In step S1004, the input feature map is first subjected to maximum pooling processing, and then a 1×1 convolution is performed on the result of the maximum pooling processing to reduce the number of channels, thereby obtaining a sixth multi-scale feature map.

[0121] Step S1005 , performing channel stitching on the obtained third multi-scale feature map, fourth multi-scale feature map, fifth multi-scale feature map and sixth multi-scale feature map.

[0122] Through step S100, convolution and pooling operations of different sizes are used in parallel. This design achieves efficient feature learning through multi-scale feature extraction and dimensionality reduction operations. The outputs of each branch are spliced ​​in the channel dimension, so that the network can simultaneously capture frequency domain features of different scales, providing rich and efficient feature representation for sound identification tasks.

[0123] However, the applicant found in the study that the process of step S100 mainly focuses on local features and lacks explicit modeling of the global context. Therefore, any level of feature enhancement processing also includes:

[0124] Step S200, performing cross attention processing on the spliced ​​feature map.

[0125] In one implementation, the cross-attention processing in step S200 adopts CCA cross-attention processing, which can not only improve the speed and accuracy of effective information transmission of internal modules in the network, but also reduce the amount of calculation and improve operating efficiency on this basis.

[0126] The drone type recognition network in the embodiment of the present application is a neural network obtained by training and learning based on the collected drone sound samples. In one embodiment, sound signals of four different types of drones in various flight states within a range of 1 to 50 meters are collected. Among them, the sound signals within the range of 1 to 10 meters are used as the sample library of training data, the sound signals within the range of 11 to 20 meters are used as the sample library of the first test data, the sound signals within the range of 21 to 30 meters are used as the sample library of the second test data, the sound signals within the range of 31 to 40 meters are used as the sample library of the third test data, and the sound signals within the range of 41 to 50 meters are used as the sample library of the fourth test data. After training and testing, the drone type recognition network used in the embodiment of the present application has the characteristic of still having a high recognition ability at a long distance.

[0127] In one embodiment of the present application, a computer-readable storage medium is provided, on which a program is stored. The stored program includes a method that can be loaded by a processor and process any one of the above embodiments.

[0128] Those skilled in the art will appreciate that all or part of the functions of the various methods in the above-mentioned embodiments can be implemented by hardware or by computer programs. When all or part of the functions in the above-mentioned embodiments are implemented by computer programs, the program can be stored in a computer-readable storage medium, and the storage medium can include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the program is executed by a computer to implement the above-mentioned functions. For example, the program is stored in the memory of the device, and when the program in the memory is executed by the processor, all or part of the above-mentioned functions can be implemented. In addition, when all or part of the functions in the above-mentioned embodiments are implemented by computer programs, the program can also be stored in a storage medium such as a server, another computer, disk, optical disk, flash disk or mobile hard disk, and can be downloaded or copied and saved in the memory of the local device, or the system of the local device is updated, and when the program in the memory is executed by the processor, all or part of the functions in the above-mentioned embodiments can be implemented.

[0129] The above specific examples are used to illustrate the present invention, which is only used to help understand the present invention and is not intended to limit the present invention. For those skilled in the art, according to the concept of the present invention, some simple deductions, modifications or substitutions can be made.

Claims

1. A method for determining the type of drone based on sound recognition, characterized in that: include: Collect the audio signal of the drone; In response to the collected audio signal, preprocessing is performed to obtain a first audio signal feature, and MFCC feature extraction and GFCC feature extraction are performed on the first audio signal feature to obtain a second audio signal feature and a third audio signal feature; The second audio signal feature and the third audio signal feature are used as inputs of a drone type recognition network to obtain the type of drone to which the collected audio signal belongs, including: Performing 7×7 convolution processing on the second audio signal feature to obtain a first feature map, performing 7×7 convolution processing on the third audio signal feature to obtain a second feature map; performing feature fusion on the first feature map and the second feature map to obtain a third feature map; Adaptively fusing the features of the third feature map in the spatial dimension to obtain a sixth feature map; Adaptively adjusting the features of the sixth feature map in the channel dimension to obtain the seventh feature map includes: Adaptively adjusting the features of the sixth feature map in the channel dimension to obtain the seventh feature map includes: Perform global average pooling on the sixth feature map; wherein the shape of the sixth feature map is (batch_size, channels, height, width), wherein batch_size represents the number of elements, channels represents the number of channels, height represents the height, and width represents the width. After global average pooling, the shape of the sixth feature map becomes (batch_size, channels, 1, 1); The first dimension of the sixth feature map after pooling is adjusted, including removing one spatial dimension, then exchanging the channel dimension and the last dimension, so (batch_size, channels, 1, 1) is adjusted to (batch_size, 1, channels); Perform one-dimensional channel convolution on the sixth feature map after the first dimension adjustment, so that the convolution kernel slides on the channel dimension; Performing a second dimension adjustment on the sixth feature map after the one-dimensional channel convolution so that the dimension is adjusted to the dimension of the sixth feature map after pooling; Perform sigmoid activation processing on the sixth feature map after the second dimension adjustment to obtain the attention weight of each channel; Multiply the obtained attention weight of each channel by the sixth feature map to obtain the seventh feature map; Performing feature enhancement processing on the seventh feature map to obtain an eighth feature map; Determine the type of the drone based on the eighth characteristic graph; The step of adaptively fusing the features of the third feature map in the spatial dimension to obtain the sixth feature map includes: Performing a first multi-scale feature extraction on the third feature map to obtain a multi-scale feature map; Performing self-correction based on the third feature map and the multi-scale feature map, and obtaining a first fusion weight and a second fusion weight based on the self-correction result; Multiplying the first fusion weight by the first feature map to obtain a fourth feature map, and multiplying the second fusion weight by the second feature map to obtain a fifth feature map; The fourth feature map and the fifth feature map are feature-fused to obtain a sixth feature map.

2. The method for determining the type of drone according to claim 1, wherein: The step of performing a first multi-scale feature extraction on the third feature map to obtain a multi-scale feature map includes: After average pooling the third feature map, the average pooled feature map is mapped to the original feature space by upsampling to obtain a first multi-scale feature map; After the third feature map is convolved, it is normalized to obtain a second multi-scale feature map.

3. The method for determining the type of drone according to claim 2, wherein: After the third feature map is average pooled, the feature map after average pooling is mapped to the original feature space by upsampling to obtain a first multi-scale feature map, including: A filter with a step size of 2 and a size of 2×2 is used for average pooling, and then a 3×3 bilinear interpolation upsampling operation is performed, and then a normalization operation is performed to obtain the first multi-scale feature map, including: ， in, represents the first multi-scale feature map, represents the normalization operation, represents bilinear interpolation upsampling, represents average pooling, represents a predefined 2×2 filter, represents the third feature map; After the third feature map is convolved, a normalization operation is performed to obtain a second multi-scale feature map, including: After performing 3×3 convolution on the third feature map, normalization is performed to obtain a second multi-scale feature map of the same size as the first multi-scale feature map. .

4. The method for determining the type of drone according to claim 1, wherein: The self-correction based on the third feature map and the multi-scale feature map, and obtaining the first fusion weight and the second fusion weight based on the self-correction result, includes: Performing a self-correction operation on the third feature map, the first multi-scale feature map, and the second multi-scale feature map to obtain a self-correction result; The self-correction results are processed by convolution and activation functions in sequence to obtain the total fusion weight; The total fusion weight is mapped to the fusion weight dimension, and the first fusion weight and the second fusion weight are calculated.

5. The method for determining the type of drone according to claim 4, wherein: The self-correction operation is performed on the third feature map, the first multi-scale feature map, and the second multi-scale feature map to obtain a self-correction result, including: ， in, represents the self-calibration result, represents element-wise multiplication, represents element-by-element addition, represents the activation function, represents the third feature map, represents the first multi-scale feature map, represents the second multi-scale feature map; The self-correction result is processed by convolution and activation function in sequence to obtain the total fusion weight, including: ， in, represents the total fusion weight, represents the filter bank of the convolution process, Represents a normalization operation; The mapping of the total fusion weight to the fusion weight dimension to calculate the first fusion weight and the second fusion weight includes: ， ， in, represents the first fusion weight, Represents the second fusion weight.

6. The method for determining the type of drone according to claim 5, wherein: The method of multiplying the first fusion weight by the first feature map to obtain the fourth feature map, and multiplying the second fusion weight by the second feature map to obtain the fifth feature map includes: ， ， in, represents the fourth feature map, represents the fifth feature map, represents the first feature map, represents the second feature map; The step of fusing the fourth feature map and the fifth feature map to obtain the sixth feature map includes: ， in, represents the sixth feature map.

7. The method for determining the type of drone according to claim 1, wherein: The one-dimensional channel convolution of the sixth feature map after the first dimension adjustment includes: setting the size of the convolution kernel based on the number of channels of the feature map and in combination with the hyperparameters, including: ， Among them, k represents the convolution kernel size, c represents the number of channels, represents the logarithm calculation with base 2, and b represent hyperparameters, and odd represents rounding down.

8. The method for determining the type of drone according to claim 1, wherein: The step of performing feature enhancement processing on the seventh feature map to obtain the eighth feature map includes: taking the seventh feature map as input, performing multi-level feature enhancement processing to obtain a ninth feature map; wherein, except for the first-level feature enhancement processing, the output of the previous-level feature enhancement processing is used as the input of the current-level feature enhancement processing; any first-level feature enhancement processing includes: Performing second multi-scale feature extraction on the input feature map, and splicing the extracted multi-scale feature maps; The concatenated feature maps are subjected to cross attention processing.

9. The method for determining the type of drone according to claim 8, wherein: The step of performing second multi-scale feature extraction on the input feature map and splicing the extracted multi-scale feature maps includes: Perform 1×1 convolution on the input feature map to obtain a third multi-scale feature map; Perform 1×1 convolution on the input feature map and then perform 3×3 convolution on it to obtain a fourth multi-scale feature map; The input feature map is subjected to 1×1 convolution and then to 5×5 convolution to obtain the fifth multi-scale feature map; Perform maximum pooling on the input feature map and then perform 1×1 convolution to obtain the sixth multi-scale feature map; Channel-joining the obtained third multi-scale feature map, fourth multi-scale feature map, fifth multi-scale feature map and sixth multi-scale feature map; The cross-attention processing of the spliced ​​feature map includes: performing CCA cross-attention processing on the spliced ​​feature map.

Images