A method for intelligent identification of a bee colony target based on SAR imaging

Abstract

The application discloses a kind of swarm target intelligent identification methods based on SAR imaging, first to the SAR image of different formation of swarm unmanned aerial vehicle and other targets obtained is preprocessed, it is converted into binary diagram and pixel value is normalized;IMG-DCNN deep convolutional neural network model is constructed, the model is composed of input layer, three layers of convolution-pooling feature extraction layer and classification layer, introduce Dropout technology and L2 regularized dual mechanism to inhibit overfitting;Using AdaBound adaptive gradient descent algorithm optimizes model training, combined with the cross-entropy loss function of L2 regularization term as training target, the target class determination is completed by Softmax classifier.The method improves the recognition accuracy of swarm unmanned aerial vehicle, and the recognition accuracy of different swarm formation, the training convergence speed is faster, and the recognition precision is higher, provides an efficient solution for SAR image recognition of unmanned aerial vehicle swarm target.

Description

Technical Field

[0001] This invention relates to the field of radar target intelligent recognition technology, specifically to a swarm target intelligent recognition method based on SAR imaging. Background Technology

[0002] With the rapid development of UAV swarm technology, its clustering, low-altitude, slow, small size, and multi-formation coordination characteristics pose a severe challenge to traditional air defense detection systems. Synthetic Aperture Radar (SAR), with its advantages of all-weather, all-time, long-range imaging, and strong anti-jamming capabilities, has become a core technology for UAV swarm target detection. However, traditional SAR target identification methods largely rely on manually designed feature extraction operators, which are poorly adaptable to target morphology changes in complex electromagnetic environments, making it difficult to effectively distinguish UAV swarms from other aerial targets, and even more difficult to accurately identify different swarm formations.

[0003] In recent years, convolutional neural networks (CNNs) have made breakthrough progress in the field of image recognition, but their direct application to bee swarm SAR image recognition still has significant shortcomings:

[0004] First, the deep model has a large number of parameters, which can easily lead to overfitting problems in scenarios where the SAR image dataset is limited in size.

[0005] Secondly, traditional optimization algorithms such as stochastic gradient descent (SGD) and AdaGrad have drawbacks such as difficulty in dynamically adapting the learning rate, slow convergence speed, and easy getting trapped in local optima, resulting in low model training efficiency and recognition accuracy that is difficult to meet the needs of practical application.

[0006] Currently, there is a lack of dedicated SAR image intelligent recognition technology for drone swarm targets and multi-formation patterns, and there is an urgent need to develop a high-precision, fast-convergence, and highly generalizable solution. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention discloses a swarm target intelligent identification method based on SAR imaging, in order to solve the problems mentioned in the background section.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a method for intelligent identification of bee swarm targets based on SAR imaging, comprising the following steps:

[0009] S1. Acquire SAR image data of swarm drones and other targets through radar SAR imaging, perform preprocessing, and construct training and test sets;

[0010] S2. Construct the IMG-DCNN network model, set the cross-entropy loss function, and combine Dropout technology with L2 regularization to construct the structural risk loss function to prevent overfitting.

[0011] S3. Use the AdaBound adaptive gradient descent algorithm to train and optimize the DCNN, iteratively update the network parameters, and save the model with the highest recognition rate on the test set.

[0012] S4. Use the trained DCNN to extract features, and use the Softmax classifier to predict the target category in the SAR image and output the recognition result.

[0013] Preferably, in step S1, the SAR image preprocessing step specifically includes:

[0014] Image transformation converts the acquired SAR image into an independent 0-1 binary image, where 0 represents a black background and 1 represents a white foreground;

[0015] Normalization processing normalizes the value of each range cell in the SAR image to the range of [0,1]. Specifically, the normalization method is to divide each pixel value by the maximum pixel value of the image.

[0016] Image cropping and adjustment: The effective area of ​​the SAR image is cropped and the image size is uniformly adjusted to 256×256 pixels, that is, the height H1=256, width W1=256, and number of channels D1=1 (single-channel grayscale image) of the input feature map.

[0017] The purpose of binary representation is to convert SAR images into 0-1 binary images, thereby eliminating the interference of irrelevant information such as noise and clutter in SAR images on target identification and highlighting the target's outline and morphological features.

[0018] Preferably, in step S2, the IMG-DCNN network model includes an input layer, a feature extraction layer, and a classification layer connected in sequence, wherein:

[0019] The input layer receives the preprocessed SAR image;

[0020] The feature extraction layer, constructed from three stacked convolutional-pooling layer modules, is used to identify useful features from the input image;

[0021] The classification layer consists of a fully connected layer and a Softmax classifier. The fully connected layer flattens the feature map output by the feature extraction layer into a one-dimensional vector before performing classification.

[0022] Each convolutional layer and fully connected layer uses ReLU as the activation function.

[0023] Preferably, step S2 further includes:

[0024] Given that too many model parameters can lead to overfitting, Dropout is used to reduce the number of parameters; simultaneously, a Softmax regression classifier is used to estimate the probability of each classification result for the input image, specifically including:

[0025] Let's define a dataset. It contains N samples, where each sample is an input image. and actual label The set; in order to estimate The probability of each classification result is determined using a Softmax regression classifier, with the function expression as follows:

[0026] ,

[0027] Softmax regression is used to estimate the classification probability of the output feature vector, thus providing a basis for subsequent classification decisions. In the formula:

[0028] ,

[0029] θ represents all trainable parameters in IMG-DCNN. The function evaluates the difference between the model's predictions and the actual labels, improving model performance by reducing the loss function. The cross-entropy empirical loss function is expressed as:

[0030] ,

[0031] In the above formula, Representative indicator function, i.e. Otherwise, the value is 0, where k represents the k-th target category.

[0032] Preferably, by fully utilizing Dropout technology, the L2 norm is incorporated into the penalty term of the empirical risk loss function to address the overfitting problem caused by the size of network parameters. After constraining all weights and biases in the network, the formula for the structural risk loss function is obtained:

[0033] ,

[0034] In the above formula, Indicates the risk factor based on experience; This represents the model's learnable parameter vector; Represents the regularization factor;

[0035] The training process for minimizing structural risk is represented as follows:

[0036] ,

[0037] in, This represents the optimal set of model parameters; during the prediction process, the classification probability is maximized to determine the category to which the sample belongs; based on the model's output feature vector, the most probable classification result is provided for the input sample.

[0038] Preferably, in step S3, the specific steps for training the model using the AdaBound adaptive gradient descent algorithm include:

[0039] 1) Model building and hyperparameter configuration: Build a convolutional neural network according to the preset network structure, and configure the training cycle, the number of input samples per batch, and all hyperparameters required for training;

[0040] 2) Initialization operation: Initialize all trainable parameters of the model to random values ​​close to zero, and initialize the first moment estimate and second moment estimate of the gradient;

[0041] 3) Iterative training: Perform a preset number of iterative training iterations, with each iteration executing the following sequentially:

[0042] 4) After the iteration is completed, output the optimal training model with the highest recognition rate on the test set.

[0043] Preferably, iterative training specifically includes the following:

[0044] 1) Forward propagation calculation: Based on the cross-entropy structure risk loss function with regularization term, calculate the forward propagation error of the current model;

[0045] 2) Gradient calculation: Solve for the gradient of the loss function with respect to the trainable parameters of the model in the current iteration;

[0046] 3) Dynamic learning rate constraint: Update the adaptive learning rate based on the gradient moment estimate, and simultaneously calculate the lower and upper bounds of the learning rate for this iteration;

[0047] 4) Gradient moment update: Based on the exponential moving average method, update the first moment estimate and the second moment estimate of the gradient respectively;

[0048] 5) Parameter update: After pruning the upper and lower bounds of the adaptive learning rate, update all trainable parameters of the model;

[0049] 6) Performance verification and optimal saving: Calculate the loss value and recognition accuracy of the current model on the training set and the test set respectively. If the accuracy of the test set is better than the historical best value, save the current model parameters and update the optimal recognition rate record.

[0050] After a total of T iterations, the model with optimal performance is obtained:

[0051] ,

[0052] In the formula, N represents the number of samples in the dataset, and B represents the number of samples in each batch. This indicates the number of training cycles.

[0053] Compared with existing technologies, the advantages of this invention are as follows: This invention generates SAR images of single swarm drones and other targets, as well as drone swarm formations, through extensive simulations, and preprocesses the SAR images. Furthermore, it employs a DCNN visual perception mechanism to analyze local characteristics step-by-step, enabling the presentation of rich morphological information of targets in the SAR images. To avoid excessive model complexity and overfitting, Dropout and L2 regularization are applied to the DCNN, limiting both the number and scale of model parameters. Finally, the AdaBound algorithm is used to optimize network training. Compared to methods such as SGD, SGD (Nestrov), and AdaGrad, this invention can more effectively accelerate the convergence speed of training and improve the accuracy of model recognition. Moreover, this invention outperforms other methods in both recognition accuracy and multi-target recognition stability, providing an effective solution for drone swarm target recognition. It achieves a recognition accuracy of over 96.28% for single swarm drones and over 95.06% for different swarm formations. Attached Figure Description

[0054] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0055] In the attached diagram:

[0056] Figure 1 This is an overall framework diagram of the intelligent target recognition method for bee colonies based on SAR imaging according to the present invention;

[0057] Figure 2 These are schematic diagrams of three common activation functions: (a) Sigmoid, (b) ReLU, and (c) Mish.

[0058] Figure 3 This is a schematic diagram of the model structure of the IMG-DCNN of this invention;

[0059] Figure 4 This is a schematic diagram illustrating the training effect of the IMG-DCNN of the present invention;

[0060] Figure 5 This is a schematic diagram illustrating the impact of the parameter configuration of this invention on IMG-DCNN; where (a) shows the effect of the p value on the model, and (b) shows the effect of the p value on the model. The impact of the value on the model; (c) The impact of image size on the model;

[0061] Figure 6The Conv2d-3 layer weights of this invention are variable. A schematic diagram illustrating the changes; where (a) (b) (c) , (d) (e) . Detailed Implementation

[0062] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0063] Example: This invention proposes a method for intelligent target recognition of bee swarms based on SAR imaging.

[0064] (I) Overall Methodological Framework

[0065] The main implementation steps of this method are described in detail as follows: Figure 1 As shown, the model comprises two main components: the SAR image visualization step and the training and testing of the DCNN. In the visualization step, the n×n SAR image is converted into independent 0-1 binary images, where 0 represents a black background and 1 represents a white foreground. To ensure model stability, the value of each distance unit in the SAR image has been normalized to the range [0,1]. In the subsequent training step, this embodiment sets a cross-entropy loss function and uses gradient descent to train the model, thereby improving its performance. In the final testing phase, this application uses the trained DCNN to extract features and applies a Softmax classifier for classification to determine the target category in the SAR image.

[0066] (II) DCNN Structure

[0067] In a convolutional neural network (CNN) consisting of alternating convolutional layers, nonlinear activation layers, and pooling layers, the l-th ( The initial feature maps of the modules are as follows: , , and In this context, these represent the height, width, and number of channels of the feature map. If the input to the first layer of the network is a 256×256 color image, since this color image consists of three channels: R, G, and B, then for this image... =256、 =256、 =3. In the convolutional layer of the l-th module, use and To represent the weights and bias parameters, where , and These represent the height, width, and number of convolutional kernels, respectively. All convolutional kernels in a single layer have the same size. After the convolution operation is completed, the k-th ( ) convolution kernels and The generated feature map is defined as:

[0068] , (1)

[0069] Where * denotes convolution operation, For the first The bias term of each convolution kernel, This represents a non-linear activation function.

[0070] Feature map The value of each node in the process is calculated by a convolutional neuron. In other words, if the convolutional kernel of the l-th hidden layer is... Use step size and Input feature map If convolution is performed on the height and width, then at the node... The output of the convolutional neuron at point 1, after being processed by the non-linear activation function, is:

[0071] (2)

[0072] in, For Hadamard products.

[0073] By using convolutional kernels to perform dot product operations on multi-channel input feature maps in specific regions, convolutional neurons can obtain the sum of results across various dimensions. After adding bias terms, the nodes of the output feature map can be calculated. This enables local perception of the feature map while sharing the weight parameter W and bias parameter b during the convolution operation.

[0074] When the convolution kernel of the l-th hidden layer by and The step size in the input feature map If convolution is performed on the height and width, then the feature map output by this layer... At height ,width Number of channels The number of parameters in each dimension and training parameters All are represented as:

[0075] , (3)

[0076] Before performing convolution calculations, the number of zero-padding points at the edges of the input feature map in both the height and width directions is calculated using... and To express.

[0077] As can be seen from formula (3), the size of the feature map generated by the convolutional layer decreases as the pooling window size increases. Therefore, the pooling layer may also perform dimensionality reduction on the input feature map. It is worth mentioning that when... , , When this occurs, the corresponding convolution is called pointwise convolution or 1×1 convolution.

[0078] Following max pooling, the convolutional output feature map of layer l is then processed. become The height and width dimensions of the pooling window are determined by... and The size step in these two directions is determined by... and Therefore, given the convolutional layer output and pooling parameters, the corresponding feature map is obtained through max pooling.

[0079] After max pooling, the values ​​in the resulting feature map are the largest pixel values ​​within the range covered by the pooling window. Specifically, the k-th... Each node of the feature map output by the pooling layer The values ​​are all derived from a corresponding input feature map. × The maximum pixel value within the pooling window of the specified size:

[0080] (4)

[0081] in, , .

[0082] If average pooling is used, equation (4) is rewritten as:

[0083] , (5)

[0084] The pooling layer output feature map is obtained by calculating using formulas (4) and (5). height ,width Number of channels and parameter quantity The representation is as follows:

[0085] , (6)

[0086] According to formula (6), the last pooling layer will refine the original one-dimensional feature map. Compression is performed, transforming it into a lower-dimensional representation, i.e. This compression process reduces the dimension of the feature map, resulting in a dimension of:

[0087] , (7)

[0088] Compression helps extract more compact and abstract features, providing more effective input for subsequent classification or regression tasks.

[0089] If the weights and biases of the l-th fully connected layer are set to... and Then the i-th ( The output signal value of 10 neurons after processing by the activation function will be:

[0090] , (8)

[0091] This output signal will serve as the input signal for the next layer, participating in subsequent calculations. Through this computational method, the fully connected layer can perform non-linear transformations and feature extraction on the input data, providing support for the training and optimization of deep learning models. Meanwhile, the formula for calculating the number of parameters in a fully connected layer is:

[0092] , (9)

[0093] As can be seen from formulas (1) and (8), the output values ​​of neurons in convolutional and fully connected layers are obtained by linearly combining the parameters of that layer with the output values ​​of neurons in the previous layer. This method can be regarded as a linear transformation of the output features of the previous layer. However, nonlinear features usually exist in actual input data. Therefore, in this embodiment, a nonlinear activation function is introduced at the output end of each convolutional and fully connected layer to enhance the model's ability to represent nonlinear features. The selection of a suitable activation function has a profound impact on the convergence speed and performance of network training. For example, Sigmoid, ReLU, and Mish are common activation functions, and their function graphs are shown in... Figure 2 It is displayed in the middle.

[0094] The Sigmoid function is represented as follows:

[0095] , (10)

[0096] This activation function plays a fundamental role in neural networks, but its value range is limited to (0,1). The ReLU function can be expressed as follows:

[0097] , (11)

[0098] When using the ReLU function, its gradient remains constant when x>0, effectively avoiding the vanishing gradient problem. It also boasts high derivative computation efficiency, making the ReLU parameters converge much faster than the Sigmoid function. However, ReLU also has limitations: when x<0, the gradient of the loss function is 0, preventing parameter updates and resulting in a "training death" situation. To address this issue, several superior activation functions have been proposed, such as Leaky ReLU, PReLU, and RReLU. These activation functions still possess a gradient when x<0, allowing parameters to continue updating and avoiding the training "death" problem. The Mish function can be represented as follows:

[0099] , (12)

[0100] In this embodiment, the ReLU model is selected, which has higher formula fluency, making gradient calculation and updates simpler, while also accelerating network training and greatly improving computational efficiency.

[0101] When initializing parameters, this embodiment follows two important principles to ensure the stability and effectiveness of model training.

[0102] (1) The output value of a nonlinear activation function has a zero-mean property, that is, the average value of its output value is close to zero:

[0103] , (13)

[0104] (2) The output value of each nonlinear activation function has a stable variance, that is, the variance remains unchanged:

[0105] , (14)

[0106] To obtain more accurate weight values, the bias parameters in the network are initialized using both the Xavier and Kaiming methods. Properly setting the bias parameters can help improve the convergence speed and stability of the model, thereby obtaining more accurate network weights.

[0107] The Xavier initialization strategy ensures that the weight parameters of the convolutional layer follow the following average distribution:

[0108] , (15)

[0109] In the formula, This represents the number of neurons in layer l. This represents the number of neurons in layer l+1. This initialization method, by setting the weight parameters appropriately, helps the model converge better during training, improving its accuracy and generalization ability.

[0110] Using the Kaiming initialization method, the weight parameters of the convolutional layer can conform to the following Gaussian distribution, thereby optimizing the training effect of the model:

[0111] , (16)

[0112] in 'a' is the slope parameter of PReLU, which is set to 0 in the ReLU function. This initialization method, through specific settings of the weight parameters, allows the model to better adapt to the data distribution during training, improving the model's convergence speed and accuracy.

[0113] Introducing random variables into the weighting parameters In this embodiment, the original formula (11) is improved to better address the overfitting problem and enhance the model's generalization ability, thus yielding:

[0114] , (17)

[0115] During training, some weight parameters are selectively retained using formula (17) and updated using the backpropagation algorithm. This allows the model to better utilize data features during learning, improving training effectiveness. In the testing phase, Dropout technology can be used to retain all key parameters, and forward propagation can be used to optimize the model, further improving its performance and generalization ability.

[0116] This application sets the output of the convolutional layer to... Therefore, the output of each dimension needs to be standardized, which can be described as follows:

[0117] , (19)

[0118] In the above formula, and It is calculated based on each mini-batch in the training dataset. To prevent the denominator from reaching a value of 0, the value is approached to near zero. Add hyperparameters. and The normalized values ​​are scaled and offset, and these two parameters are adjusted to control the range of the normalized values.

[0119] , (20)

[0120] like Figure 3 The model structure of IMG-DCNN is shown, and Table 1 lists the detailed parameter settings of the network. The network model mainly consists of three parts: an input layer, a feature extraction layer, and a classification layer. The feature extraction layer is mainly responsible for identifying useful features from the input image; this layer is constructed from three stacked convolutional-pooling layer modules. Subsequently, the classification layer uses these features to classify the target. This enables IMG-DCNN to efficiently perform SAR image recognition tasks.

[0121] Table 1 Parameter Configuration of IMG-DCNN

[0122]

[0123] The model uses 3x3 convolutional kernels and 4x4 pooling windows, with a stride of 1x1. To prevent information loss, "valid" is used to pad with zeros. ReLU is used as the activation function for each layer. Given that too many model parameters may lead to overfitting, this application employs Dropout to reduce the number of parameters. Specific parameter settings are shown in formulas (3) and (9). Calculations show that the network contains a total of 132,304 trainable parameters.

[0124] Suppose a dataset It contains N samples, all of which are input images. and actual label The set. In order to estimate For each classification result probability, this application chooses to use a Softmax regression classifier. The following is the function assumed in this embodiment.

[0125] , (twenty one)

[0126] Using this hypothesis function, softmax regression can be used to estimate the classification probability of the output feature vector, thus providing a basis for subsequent classification decisions. Where:

[0127] ,

[0128] This represents all trainable parameters in IMG-DCNN. This function evaluates the difference between the model's predictions and the actual labels, improving model performance by reducing the loss function. The cross-entropy empirical loss function can be expressed as:

[0129] ,(twenty two)

[0130] In the above formula, Representative indicator function, i.e. Otherwise, the value is 0.

[0131] By fully utilizing Dropout technology, the L2 norm is incorporated into the penalty term of the empirical risk loss function to address the overfitting problem caused by the size of network parameters. After constraining all weights and biases in the network, the formula for the structural risk loss function is obtained:

[0132] , (twenty three)

[0133] In the above formula, This represents the regularization factor. The training process for minimizing structural risk can be represented as:

[0134] ,(twenty four)

[0135] In the prediction process, this application determines the category of sample x by maximizing the classification probability. Specifically, the prediction process can be expressed as:

[0136] , (25)

[0137] in This represents the predicted label. In this way, this application can provide the most probable classification result for the input sample based on the model's output feature vector.

[0138] (III) DCNN Training Optimization

[0139] Stochastic Gradient Descent (SGD) is an algorithm that, in each round of repeated training, randomly selects some training samples from the training set to update the parameters. Specifically, in the... During repeated training, the formula for parameter updates is:

[0140] , (26)

[0141] In the parameter update expression of formula (26), Represents the learning rate, while The gradient of the loss function represents the degree of sensitivity of the loss function to the model parameters. The learning rate determines the step size for each parameter update, while the gradient of the loss function indicates how sensitive the loss function is to the model parameters. It can be represented as:

[0142] , (27)

[0143] During network training, stochastic gradient descent (SGD) uses the same learning rate to adjust parameters, which makes finding an appropriate learning rate challenging. This presents a challenge. If the learning speed is too high, the training becomes unstable; conversely, if the learning speed is too slow, the convergence efficiency decreases, and it may even lead to getting trapped in local optima or saddle points. To address these challenges, this application employs the adaptive gradient descent algorithm AdaBound to improve the parameter tuning process. AdaBound considers previous gradient effects in all parameter update steps. It adjusts Equation (27) by using gradient estimates of first and second moments respectively. and At the same time, regarding Set a dynamic boundary to ensure that in each iteration, It can be set to a stable interval, and this interval will gradually decrease as the number of iterations increases. The training optimization algorithm 1 for deep convolutional neural networks (DCNN) based on AdaBound is shown below.

[0144]

[0145] After a total After several iterations, the model with the best performance was obtained. The number of iterations is related to the number of samples in the dataset. The number of samples (B) in each batch and the number of training cycles. Related, that is:

[0146] (28)

[0147] Specific experiments and analysis

[0148] (a) IMG-DCNN recognition performance

[0149] This embodiment will use the Python 3.7 programming language and the Keras framework with Tensorflow 1.14 as the backend to develop all the convolutional neural network models. The hardware foundation is an Intel Core i5 9600K and a Windows 10 64-bit operating system, and a control experiment will be conducted under the same hardware and software conditions.

[0150] In the initial stage, SAR images of bee swarm targets and various other targets were obtained through extensive simulation experiments. The effective regions of the images were then cropped, and the image size was adjusted to 256 x 256 (w = h = 256) to obtain the final dataset. The IMG-DCNN was validated using the test dataset. Furthermore, the accuracy variation curves of the test set were obtained by comparing SGD, Nesterov-based SGD, and AdaGrad training algorithms with the AdaBound algorithm used in this paper.

[0151] Depend on Figure 4It can be seen that, on the test dataset, compared with other algorithms, the AdaBound algorithm for training IMG-DCNN has a faster convergence speed and better recognition effect, and has a higher accuracy asymptote. Therefore, the AdaBound algorithm plays a certain role in optimizing network training.

[0152] Table 2 shows the recognition accuracy of the training algorithm for different targets. It can be seen that, using the AdaBound training algorithm, the recognition accuracy for swarm drones can reach over 96.28%. Table 3 shows the recognition accuracy of the training algorithm for different swarm formations. It can be seen that, using the AdaBound training algorithm, the recognition accuracy for swarm drones can reach over 95.06%.

[0153] Table 2. IMG-DCNN's accuracy rate (%) in identifying swarm drones and other targets.

[0154] Training Algorithm Swarm drones AIM missile F-117 Tomahawk missile SGD 94.73 91.45 94.78 94.73 SGD (Nesterov) 94.78 92.17 95.39 94.67 AdaGrad 95.61 92.23 95.56 95.45 AdaBound 96.28 93.56 96.17 95.95

[0155] Table 3. IMG-DCNN's accuracy rate (%) in recognizing different drone swarm formations

[0156] Training Algorithm Square formation Trapezoidal formation Horizontal parallel formation Parallel formation AdaBound 95.34 95.06 97.12 96.33

[0157] (II) The impact of parameter configuration on IMG-DCNN

[0158] Figure 5 This demonstrates the Dropout parameters during training. L2 regularization parameters There is also a curve showing the impact of image size on recognition accuracy. Figure 5 (a) indicates that when the Dropout parameter The model performs optimally at that time, and at the same time, as Performance initially improves but then deteriorates; Figure 5 (b) Display the regularization parameters The effect is best at this time; Figure 5 (c) shows that the larger the image, the better the recognition effect.

[0159] Figure 6 For different This diagram illustrates the change of the weight histogram of the third convolutional layer in a deep convolutional neural network over training cycles. The horizontal axis represents the parameter values, and the vertical axis represents the frequency of those values. Figure 6 It can be seen that the weight value is close to 0, and as... As the value increases, the radius of the neighborhood gradually decreases.

[0160] The above results indicate that the relevant parameters in the experiment and They should be set in the intervals [0, 0.2] and [0, 10] respectively. -3 Within the specified range; experimental data show that the SAR image recognition technology based on DCNN proposed in this invention outperforms other methods in terms of recognition accuracy and multi-target recognition stability, providing an effective solution for target recognition in UAV swarms. Ultimately, the recognition accuracy for a single UAV swarm reached over 96.28%, while the recognition accuracy for different swarm formations reached over 95.06%.

[0161] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent target recognition of bee swarms based on SAR imaging, characterized in that: Includes the following steps: S1. Acquire SAR image data of swarm drones and other targets through radar SAR imaging, perform preprocessing, and construct training and test sets; S2. Construct the IMG-DCNN network model, set the cross-entropy loss function, and combine Dropout technology with L2 regularization to construct the structural risk loss function; S3. Use the AdaBound adaptive gradient descent algorithm to train and optimize the DCNN, iteratively update the network parameters, and save the model with the highest recognition rate on the test set. S4. Use the trained DCNN to extract features, and use the Softmax classifier to predict the target category in the SAR image and output the recognition result.

2. The method for intelligent identification of bee swarm targets based on SAR imaging according to claim 1, characterized in that: In step S1, the SAR image preprocessing steps specifically include: Image transformation converts the acquired SAR image into an independent 0-1 binary image, where 0 represents a black background and 1 represents a white foreground; Normalization processing normalizes the value of each range cell in the SAR image to the range of [0,1]. Image cropping and adjustment: cropping the effective area of ​​SAR images.

3. The method for intelligent identification of bee swarm targets based on SAR imaging according to claim 1, characterized in that: In step S2, the IMG-DCNN network model includes an input layer, a feature extraction layer, and a classification layer connected in sequence, wherein: The input layer receives the preprocessed SAR image; The feature extraction layer, constructed from three stacked convolutional-pooling layer modules, is used to identify useful features from the input image; The classification layer consists of a fully connected layer and a Softmax classifier. The fully connected layer flattens the feature map output by the feature extraction layer into a one-dimensional vector before performing classification. Each convolutional layer and fully connected layer uses ReLU as the activation function.

4. The method for intelligent identification of bee swarm targets based on SAR imaging according to claim 3, characterized in that: Step S2 also includes: Dropout is used to reduce the number of parameters; a Softmax regression classifier is used to estimate the probability of each classification result for the input image, specifically including: Let's define a dataset. It contains N samples, where each sample is an input image. and actual label The set; in order to estimate The probability of each classification result is determined using a Softmax regression classifier, with the function expression as follows: ， Softmax regression is used to estimate the classification probability of the output feature vector, thus providing a basis for subsequent classification decisions. In the formula: ， θ represents all trainable parameters in IMG-DCNN, and the performance of the model is improved by reducing the loss function; while the cross-entropy empirical loss function is expressed as: ， In the above formula, Representative indicator function, i.e. Otherwise, the value is 0, where k represents the k-th target category.

5. The method for intelligent identification of bee swarm targets based on SAR imaging according to claim 4, characterized in that: Based on Dropout technology, the L2 norm is incorporated into the penalty term of the empirical risk loss function. After constraining all weights and biases in the network, the formula for the structural risk loss function is obtained: ， In the above formula, Indicates the risk factor based on experience; This represents a learnable parameter vector of the model; Represents the regularization factor; The training process for minimizing structural risk is represented as follows: ， in, This represents the optimal set of model parameters; during the prediction process, the classification probability is maximized to determine the category to which the sample belongs; based on the model's output feature vector, the most probable classification result is provided for the input sample.

6. The method for intelligent identification of bee swarm targets based on SAR imaging according to claim 1, characterized in that: In step S3, the specific steps for training the model using the AdaBound adaptive gradient descent algorithm include: 1) Model building and hyperparameter configuration: Convolutional neural network is built according to the preset network structure, and the training cycle, the number of input samples per batch and all hyperparameters required for training are configured; 2) Initialization operation: Initialize all trainable parameters of the model to random values ​​close to zero, and initialize the first moment estimate and second moment estimate of the gradient; 3) Iterative training: Perform a preset number of iterative training iterations, with each iteration executing the following sequentially: 4) After the iteration is completed, output the optimal training model with the highest recognition rate on the test set.

7. The method for intelligent identification of bee swarm targets based on SAR imaging according to claim 6, characterized in that: Iterative training specifically includes the following: 1) Forward propagation calculation: Based on the cross-entropy structure risk loss function with regularization term, calculate the forward propagation error of the current model; 2) Gradient calculation: Solve for the gradient of the loss function with respect to the trainable parameters of the model in the current iteration; 3) Dynamic learning rate constraint: Update the adaptive learning rate based on the gradient moment estimate, and simultaneously calculate the lower and upper bounds of the learning rate for this iteration; 4) Gradient moment update: Based on the exponential moving average method, update the first moment estimate and the second moment estimate of the gradient respectively; 5) Parameter update: After pruning the upper and lower bounds of the adaptive learning rate, update all trainable parameters of the model; 6) Performance verification and optimal saving: Calculate the loss value and recognition accuracy of the current model on the training set and the test set respectively. If the accuracy of the test set is better than the historical best value, save the current model parameters and update the optimal recognition rate record. After a total of T iterations, the model with optimal performance is obtained: ， In the formula, N represents the number of samples in the dataset, and B represents the number of samples in each batch. This indicates the number of training cycles.

Images