SAR Network Target Recognition Method Based on Xception

Abstract

This invention relates to the field of target recognition technology, and more particularly to a target recognition method based on X-ception SAR networks. This invention combines 2D SAR images acquired by an unmanned aerial vehicle (UAV) platform with multiple convolutional neural network optimization strategies, including improved activation functions, channel-spatial attention mechanisms, and improved loss functions, to achieve accurate extraction of key target features and robust recognition under small sample conditions. It possesses advantages such as lightweight structure, strong adaptability, and ease of deployment, making it more suitable for complex traffic monitoring and safety identification scenarios with limited resources or high requirements for deployment flexibility. Furthermore, by incorporating transfer learning strategies, parameters are fine-tuned using a small amount of measured data on a pre-trained model, enabling rapid transfer of the model from the source domain to the target domain. The recognition state is determined by the predicted values ​​output by the model, and corresponding parameters are adjusted based on the recognition state. This invention improves the target recognition accuracy in complex SAR images.

Description

Technical Field

[0001] This invention relates to the field of target recognition technology, and in particular to a target recognition method for SAR networks based on Xception. Background Technology

[0002] Synthetic Aperture Radar (SAR), as an active imaging sensor, possesses the ability to operate independently of natural lighting conditions and penetrate complex environments such as rain, fog, and smoke. Therefore, it has been widely used in all-weather, all-time traffic monitoring, battlefield target identification, and unmanned system perception. Compared to traditional optical images, SAR images record the electromagnetic scattering characteristics of targets. Their imaging results are significantly affected by various factors such as the target's material, geometry, and relative orientation, leading to highly nonlinear image features, complex and unstable spatial structures, and greatly increasing the difficulty of target identification.

[0003] Traditional SAR target recognition methods typically rely on manually extracting low-level physical features such as scattering center distribution and edge contours, and then combining them with classification algorithms such as support vector machines, principal component analysis, or kernel methods for identification. While these methods have achieved certain results under specific conditions, their performance is highly dependent on the quality of feature engineering, lacks generalization ability, and is difficult to adapt to real-world application scenarios with large target pose variations, difficult sample annotation, or low image signal-to-noise ratio. Furthermore, due to the characteristics of SAR images, such as sensitivity to target deformation, sparse texture information, and significant background interference, existing recognition methods often suffer from high false positive rates and unstable recognition when handling complex scenes, hindering their widespread adoption and application in practical engineering.

[0004] In recent years, the rapid development of deep learning technology has provided new solutions for SAR image recognition. Convolutional Neural Networks (CNNs), through multi-level automatic feature extraction mechanisms, have achieved remarkable results in the field of natural image recognition and are gradually being introduced into SAR target recognition tasks. CNN-based recognition methods can, to some extent, reduce reliance on hand-crafted features and improve the autonomous learning and adaptability of recognition models. However, standard CNN structures still have many limitations in SAR image recognition, such as large network parameters, large training sample requirements, susceptibility to overfitting, and sensitivity to changes in target geometry, especially under small sample conditions.

[0005] Chinese Patent Publication No. CN115909086B discloses a SAR target detection and recognition method based on a multi-level enhancement network. The implementation scheme is as follows: SAR measured data is labeled and divided to obtain training and test sets; a multi-level enhancement network is constructed, consisting of a data-level enhancement module, a feature-level enhancement module, a region proposal module, and a decision-level enhancement module cascaded together; the multi-level enhancement network is trained using the training set based on the stochastic gradient descent algorithm; and test set images are input into the trained multi-level enhancement network to obtain the SAR target detection and recognition results.

[0006] It is evident that the existing technology has the following problems: the target recognition has poor adaptability in real-world environments such as low signal-to-noise ratio and complex attitude changes, resulting in low target recognition accuracy in complex SAR images. Summary of the Invention

[0007] To address this issue, the present invention provides a SAR network target recognition method based on Xception, which overcomes the problem that the target recognition in the prior art has poor adaptability in real-world environments such as small sample sizes, low signal-to-noise ratios, and complex attitude changes, resulting in low target recognition accuracy in complex SAR images.

[0008] To achieve the above objectives, this invention provides a target recognition method for SAR networks based on Xception, comprising:

[0009] Acquire SAR images generated by electromagnetic simulation and SAR target data measured by millimeter-wave radar;

[0010] The SAR image and the SAR target data are preprocessed, wherein the preprocessing includes at least data cleaning and data augmentation;

[0011] The pre-processed SAR image is used to pre-train a recognition model containing the Dropout mechanism to obtain a first recognition model;

[0012] The network layer parameters of the first recognition model are adjusted using the training set divided from the preprocessed SAR target data, and the adjusted first recognition model is trained with training parameters to obtain the second recognition model. The training parameters include at least the batch size.

[0013] The test set divided from the preprocessed SAR target data is input into the second recognition model, and the recognition accuracy of the second recognition model is output.

[0014] The recognition state is determined based on the recognition accuracy, the batch size is adjusted based on the recognition state, and the dropout rate of the Dropout mechanism during pre-training is adjusted based on the recognition state after adjusting the batch size.

[0015] Further, the recognition model includes an improved Xception network, a global average pooling layer, and a fully connected layer. The process of pre-training the recognition model containing the Dropout mechanism using the preprocessed SAR image to obtain the first recognition model includes: dividing the preprocessed SAR image into a training set and a validation set according to a preset ratio; inputting the training set into the improved Xception network to output a feature map, wherein the improved Xception network adds a channel-spatial attention mechanism after the last layer of the Xception network, and replaces the ReLU activation function of the preset convolutional layer in the Xception network with a LeakyReLU activation function; inputting the feature map into the global average pooling layer to calculate the average value of each channel in the feature map to obtain a feature vector; using a fully connected layer to reduce the feature dimension of the feature vector and outputting the original score of the target category corresponding to the classification task; using a Softmax classifier to normalize the multiple original scores to obtain a predicted probability, and iterating the recognition model based on the loss function calculated based on the predicted probability to obtain the first recognition model.

[0016] Furthermore, the process of pre-training the recognition model containing the Dropout mechanism using the preprocessed SAR image to obtain the first recognition model also includes: freezing several layers in the Xception network during the initial pre-training period; calculating the validation parameters of the pre-trained recognition model based on the validation set; unfreezing the preset layers at least once based on the validation parameters, and retraining after each unfreezing; calculating the loss function after each training session, and if the loss function is less than or equal to the preset loss function, the trained recognition model is recorded as the first recognition model.

[0017] Furthermore, the process of reducing the feature dimension of the feature vector using a fully connected layer and outputting the original score of the target category for the corresponding classification task includes: inputting the feature vector into a fully connected layer to obtain a first feature vector with reduced feature dimension; optimizing the first feature vector based on the ReLU activation function, L2 regularization, and Dropout mechanism to obtain a second feature vector; re-inputting the second feature vector into a fully connected layer to obtain a third feature vector with reduced feature dimension; and using the ReLU activation function, L2 regularization, and Dropout mechanism in synergy on the third feature vector to output the original score of the target category for the corresponding classification task at the fully connected layer.

[0018] Further, the process of adjusting the network layer parameters of the first identification model using the training set divided from the preprocessed SAR target data to obtain the second identification model includes: replacing the classification layer of the first identification model with a fully connected layer and a Softmax activation layer according to the number of target categories; inputting the training set divided from the SAR target data into the improved first identification model for training; calculating the gradient of the loss function with respect to the network parameters layer by layer based on the chain rule, and updating the network layer parameters iteratively using the gradient descent algorithm according to the gradient until the absolute value of the slope of the loss function curve output by the consecutive preset number of iterations is less than a preset threshold, stopping the model training, and recording the trained model as the second identification model.

[0019] Furthermore, the process of determining the recognition status based on the recognition accuracy and adjusting the batch size based on the recognition status includes: when the recognition status is determined to be unqualified based on the recognition accuracy, obtaining the recognition accuracy of each round during the training iteration; plotting the iteration round-recognition accuracy curve and calculating the average value of the absolute value of the slope of the curve; if the average value is greater than the preset average value, adjusting the batch size based on the difference between the average value and the preset average value; wherein, the recognition status is unqualified when the recognition accuracy is greater than the preset accuracy.

[0020] Furthermore, the process of adjusting the batch size based on the difference between the average value and the preset average value includes: increasing the batch size based on the difference between the average value and the preset average value, and the increase in batch size is proportional to the difference.

[0021] Furthermore, the method also includes: retraining the second recognition model after adjusting the batch size; if the recognition status after training is unqualified, obtaining the recognition accuracy of the training set and the test set in the SAR target data respectively; calculating the absolute value of the difference between the recognition accuracy of the training set and the test set; if the absolute value is greater than a preset difference, adjusting the dropout rate in the Dropout mechanism based on the ratio of the absolute value to the preset difference.

[0022] Furthermore, the process of adjusting the dropout rate based on the ratio of the absolute value to the preset difference includes: increasing the dropout rate based on the ratio of the absolute value to the preset difference, and the increase in the dropout rate is proportional to the ratio.

[0023] Furthermore, the method also includes: re-detecting the recognition status after adjusting the loss rate; if the recognition status is unqualified, re-acquiring the Loss curve; calculating the integral of the Loss curve; if the integral is greater than a preset integral, adjusting the learning rate period of the gradient descent algorithm based on the ratio of the integral to the preset integral, and the reduction of the period is proportional to the ratio.

[0024] Compared with existing technologies, the advantages of this invention lie in its ability to combine 2D SAR images acquired by an unmanned aerial vehicle (UAV) platform with multiple convolutional neural network optimization strategies, including improved activation functions, channel-spatial attention mechanisms, and improved loss functions. This enables accurate extraction of key target features and robust identification under small sample conditions, offering advantages such as lightweight structure, strong adaptability, and ease of deployment. It is particularly suitable for complex traffic monitoring and safety identification scenarios with limited resources or high deployment flexibility requirements. Furthermore, by incorporating transfer learning strategies, parameters are fine-tuned using a small amount of real-world data on a pre-trained model, enabling rapid model transfer from the source domain to the target domain. The identification state is determined by the model's predicted output, and corresponding parameters are adjusted based on this state. This invention improves the accuracy of target identification in complex SAR images.

[0025] Furthermore, by adding a channel-spatial attention mechanism and replacing the activation function to improve the Xception network, this invention can enhance the model's ability to focus on the target region and suppress background noise interference. Thus, it can improve recognition accuracy and model stability under actual environmental conditions such as changes in viewpoint, pose deformation, and low signal-to-noise ratio.

[0026] Furthermore, by initially freezing several layers in the Xception network, this invention can ensure the feature extraction capability during the pre-training stage, prevent the distortion of low-level features caused by small datasets, reduce the number of training parameters, optimize time costs, and improve the model's ability to recognize features at different levels in SAR images by gradually unfreezing preset layers in the network, thereby further improving the target recognition accuracy in complex SAR images.

[0027] Furthermore, by repeatedly inputting the feature vector into the fully connected layer and repeatedly optimizing the feature vector through the ReLU activation function, L2 regularization, and Dropout mechanism, this invention can reduce the number of model parameters while reducing the risk of model overfitting, thereby further improving the target recognition accuracy in complex SAR images.

[0028] Furthermore, this invention utilizes a transfer learning strategy to fine-tune parameters based on a small amount of measured data on a pre-trained model, enabling rapid transfer of the model from the source domain to the target domain, thereby improving the target recognition efficiency in complex SAR images.

[0029] Furthermore, this invention determines whether to adjust the batch size based on the average absolute value of the slope of the iteration-recognition accuracy curve when the recognition status is unqualified. This allows for more effective adjustment based on the cause, thereby further improving the target recognition accuracy in complex SAR images.

[0030] Furthermore, the present invention increases the batch size based on the difference between the average value and the preset average value, which can more accurately adjust the batch size, thereby solving the problem of training convergence being blocked, and further improving the target recognition accuracy in complex SAR images.

[0031] Furthermore, this invention determines the reasons for unqualified recognition status based on the absolute value of the difference between the recognition accuracy of the training set and the test set, and can adjust the model more effectively based on the reasons, thereby further improving the recognition accuracy of targets in complex SAR images.

[0032] Furthermore, this invention increases the dropout rate in the Dropout mechanism based on the ratio of the absolute value to the preset difference, which can more accurately adjust the dropout rate in the Dropout mechanism, thereby matching the network complexity of the model with the data scale, and further improving the target recognition accuracy in complex SAR images.

[0033] Furthermore, if the re-detection of the identification status is still unqualified, the present invention determines whether the Loss curve has rebounded based on the integral of the Loss curve, so that the model "skips" the optimal solution during the optimization process, and adjusts the learning rate period of the gradient descent algorithm based on this reason, thereby further improving the identification accuracy of targets in complex SAR images. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the structure of the SAR network target recognition system based on Xception according to an embodiment of the present invention;

[0035] Figure 2 This is a flowchart illustrating the steps of the SAR network target recognition method based on Xception according to an embodiment of the present invention.

[0036] Figure 3 This is a flowchart illustrating the pre-training steps of the SAR network target recognition model based on Xception according to an embodiment of the present invention.

[0037] Figure 4 This is a partial structural diagram of the Xception network in the SAR network target recognition model based on Xception according to an embodiment of the present invention;

[0038] Figure 5 This is a flowchart illustrating the steps of generating a second recognition model using the SAR network target recognition method based on Xception according to an embodiment of the present invention.

[0039] Figure 6 This is a flowchart illustrating the steps of determining the accuracy of an identification based on a comparison between the recognition accuracy and a preset accuracy, according to an embodiment of the present invention.

[0040] Figure 7Training curves for the model using ReLU and LeakyReLU activation functions respectively;

[0041] Figure 8 Training curves for the model with spatial-channel attention and the original model;

[0042] Figure 9 Training curves for the model using Focal Loss and cross-entropy loss functions respectively;

[0043] Figure 10 The average recognition rate of each model;

[0044] Figure 11 The average recognition rates of the model were obtained by applying the Adam gradient optimizer, SGD, and MBGD respectively. Detailed Implementation

[0045] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0046] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0047] It should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0048] Please see Figure 1 As shown, it is a schematic diagram of the structure of the SAR network target recognition system based on Xception according to an embodiment of the present invention.

[0049] The system includes an acquisition unit, a preprocessing unit, a first recognition model generation unit, a second recognition model generation unit, an output unit, and an analysis unit.

[0050] The acquisition unit is used to acquire SAR images generated by electromagnetic simulation and SAR target data measured by millimeter-wave radar.

[0051] The preprocessing unit is connected to the acquisition unit and is used to preprocess the SAR image and the SAR target data, wherein the preprocessing includes at least data cleaning and data augmentation.

[0052] The first recognition model generation unit is connected to the preprocessing unit, and is used to pre-train the recognition model containing the Dropout mechanism using the preprocessed SAR image to obtain the first recognition model.

[0053] The second identification model generation unit is connected to the first identification model generation unit. It uses the training set divided in the preprocessed SAR target data to adjust the network layer parameters of the first identification model and sets training parameters to train the adjusted first identification model to obtain the second identification model. The training parameters include at least the batch size.

[0054] The output unit is connected to the second identification model generation unit, and is used to input the test set divided in the preprocessed SAR target data into the second identification model and output the identification accuracy of the second identification model.

[0055] The analysis unit is connected to the output unit and is used to determine the recognition status based on the recognition accuracy, adjust the batch size based on the recognition status, and adjust the dropout rate of the Dropout mechanism during pre-training based on the recognition status after adjusting the batch size.

[0056] Specifically, the processing steps of the acquisition unit and the data preprocessing unit are as follows: Since the sizes of all target images in the SAR dataset are inconsistent, to adapt to the network structure centered on Xception, the size of all SAR images used for network training and testing is first unified to 299×299 to ensure successful input into the Xception network. Next, diversity processing is performed on all SAR image datasets. The SAR images generated by electromagnetic simulation are divided, with an average of 80% selected from each class as the training set and the remaining 20% ​​as the validation set. The purpose of the validation set here is to monitor the validation loss of the model during pre-training. If the loss of the training set continues to decrease while the loss of the validation set continues to increase, it indicates that the model is overfitting to training set noise or specific sample features, requiring early stopping for readjustment. Typically, when the loss of the validation set stabilizes at its lowest point, it represents that the model has reached its optimal generalization state. Then, an average of 80% of the radar-measured SAR target data from each class is selected as the training set for fine-tuning the network model in the second stage, with the remaining 20% ​​used as the test set. The selected test set samples are always kept in their original, untrained state and must not be disclosed. They are only used at the end to test the model's recognition performance.

[0057] The input image information is sequentially stored into the corresponding channels to form a complete data path. Finally, the path is read into the algorithm's runtime environment to form the training set, validation set, and test set. Whether it's the training set, validation set, or test set, all are labeled image samples. It's necessary to properly categorize these datasets into folders, ensuring that SAR images within the same folder represent the same type of target, thus preparing for target category determination after the algorithm's output.

[0058] Meanwhile, since modifying different imaging parameters can generate various SAR images with different attitudes and deformations through electromagnetic simulation, data augmentation has been performed, making the quantity sufficient to meet the basic training of the model. However, the amount of high-quality image data obtained from actual radar measurements is relatively small, which can be used for final testing. But before entering the fine-tuning stage, the training images obtained from these images need to be augmented to a certain extent, such as by rotation, flipping, or translation.

[0059] Please see Figure 2 The diagram shown is a flowchart of the steps of the SAR network target recognition method based on Xception according to an embodiment of the present invention.

[0060] The steps of the SAR network target recognition method based on Xception include:

[0061] S1, acquires SAR images generated by electromagnetic simulation and SAR target data measured by millimeter-wave radar through the acquisition unit;

[0062] S2, the SAR image and the SAR target data are preprocessed by a preprocessing unit connected to the acquisition unit, wherein the preprocessing includes at least data cleaning and data augmentation;

[0063] S3, the first recognition model generation unit connected to the preprocessing unit uses the preprocessed SAR image to pre-train the recognition model containing the Dropout mechanism to obtain the first recognition model;

[0064] S4, the second identification model generation unit connected to the first identification model generation unit adjusts the network layer parameters of the first identification model using the training set divided in the preprocessed SAR target data, and sets training parameters to train the adjusted first identification model to obtain the second identification model, wherein the training parameters include at least the batch size.

[0065] S5, the test set divided in the preprocessed SAR target data is input into the second identification model through the output unit connected to the second identification model generation unit, and the identification accuracy of the second identification model is output.

[0066] S6, the analysis unit connected to the output unit determines the recognition status based on the recognition accuracy, adjusts the batch size based on the recognition status, and adjusts the dropout rate of the Dropout mechanism during pre-training based on the recognition status after adjusting the batch size.

[0067] Please see Figure 3 The diagram shown is a flowchart illustrating the pre-training steps of the SAR network target recognition model based on Xception according to an embodiment of the present invention.

[0068] Specifically, Xception comprises over 300 layers (including sub-layers), employing a strategy of freezing the first two-thirds of the base network, including the first nine Inception sub-modules. These sub-modules contain all the basic convolutional kernels, such as 3×3 and 5×5. The reason for freezing these network layers is that they largely capture more basic features, such as the edge contours and textures of SAR images, while deeper network layers largely contain more advanced image features. This approach ensures the feature extraction capability during pre-training, prevents low-level feature distortion caused by small datasets, reduces the number of training parameters, and optimizes time costs. Based on the performance on the validation set, the higher-level convolutional layers are gradually unfrozen, improving the model's ability to recognize features at different levels in SAR images.

[0069] The training set of SAR images generated by electromagnetic simulation is input into the network after data augmentation. After the input images pass through the last layer of the Xception network module, an 8×8×2048 feature map is output and then standardized.

[0070] The output feature map is fed into a global average pooling layer, where the mean value of each channel in the feature map is calculated, transforming the 8×8×2048 vector into a 1×1×2048 vector, significantly reducing the number of parameters. At this layer, the most important features have already been extracted.

[0071] The output feature vector is sent to a fully connected layer and its dimension is reduced to 1024 to further reduce the number of model parameters and reduce the risk of overfitting; the ReLU non-linear activation function is introduced to enhance the non-linear expressive power; L2 regularization is introduced with an intensity coefficient of 0.01, the core function of which is to prevent the model from overfitting by constraining the size of the weights; at the same time, the Dropout mechanism is introduced to immediately block 50% of the neuron output.

[0072] The output vector from the upper layer is fed back into the fully connected layer, where a progressive dimensionality reduction is used to gradually reduce the feature dimension to 512. A ReLU non-linear activation function is introduced, along with L2 regularization, and combined with the Dropout mechanism.

[0073] The fully connected layer outputs num_classes, which corresponds to the number of target classes in the classification task. The probability is normalized using a softmax classifier to obtain the prediction result and loss function.

[0074] The backpropagation process involves calculating the gradient of the loss function with respect to the network parameters layer by layer using the chain rule. The network is then iterated using the gradient descent algorithm until the loss function is minimized. This experiment utilizes the Adam gradient descent algorithm with an adaptive learning rate, which has low memory consumption, avoids the tediousness of manually adjusting the learning rate, and is flexible and convenient, making it suitable for model training.

[0075] After continuous network iteration and parameter updates, the pre-trained model finally acquired the ability to extract basic general features from the source domain samples, providing parameter weights for subsequent learning of new tasks in the target domain and laying a solid foundation. The pre-training phase thus concludes.

[0076] Please see Figure 4 As shown, it is a partial structural diagram of the Xception network in the SAR network target recognition model based on Xception according to an embodiment of the present invention.

[0077] Specifically, the Xception network employs a modular structure, also known as GoogLeNet. It uses global average pooling layers instead of fully connected layers, thus reducing spatial dimensionality and significantly decreasing the number of parameters. This avoids the overfitting risk caused by fully connected layers while preserving important information for each channel. To reduce computational cost, additional 1×1 convolutional layers are added within nested modules, limiting the number of input channels.

[0078] Please see Figure 5 The diagram shown is a flowchart illustrating the steps of generating a second recognition model using the SAR network target recognition method based on Xception according to an embodiment of the present invention.

[0079] Specifically, the pre-trained model replaces the original classification layer while retaining the convolutional layers. A new classification layer, including a Dense layer and a Softmax activation layer, is added at the top based on the number of target categories in the actual test. The training set, divided from the SAR dataset of the actual radar targets, is augmented and fed into the modified network model. Similarly, some convolutional layers are initially frozen, and new fully connected layers are trained primarily using target domain SAR data, continuously optimizing their parameters to fully utilize the general features obtained in the pre-training phase. Then, higher-level convolutional layers are gradually unfrozen to balance feature generality and domain adaptability. Backpropagation and gradient updates are used iteratively to update network layer parameters throughout the process. Training stops when the curves for training loss and accuracy stabilize. At this point, the model has transitioned from general features to target domain-specific features, enabling it to adapt to target classification tasks.

[0080] Please see Figure 6 The diagram shows a flowchart illustrating the steps of determining the recognition status based on a comparison between the recognition accuracy and a preset accuracy in an embodiment of the present invention. The process of determining the recognition status based on the recognition accuracy and adjusting the batch size based on the recognition status in this embodiment includes: when the recognition status is deemed unqualified based on the recognition accuracy, obtaining the recognition accuracy for each round during the training iteration; plotting the iteration round-recognition accuracy curve and calculating the average value of the absolute value of the curve's slope; if the average value is greater than a preset average value, adjusting the batch size based on the difference between the average value and the preset average value; wherein, the recognition status is deemed unqualified when the recognition accuracy is greater than the preset accuracy.

[0081] Specifically, taking the SAR image recognition of target vehicles measured by radar as an example, and based on the performance limits of the model and the fault tolerance requirements of the actual environment, as well as some historical data obtained through statistics and analysis during the historical training process, the numerical settings of subsequent corresponding preset or critical parameters are made.

[0082] Specifically, if the preset accuracy L0 = 85%, the comparison process between the recognition accuracy L and the preset accuracy L0 is as follows:

[0083] If the recognition accuracy L is greater than or equal to the preset accuracy L0, it indicates that the recognition status is qualified.

[0084] If the recognition accuracy L is less than the preset accuracy L0, it means that the recognition status is not qualified. Then, the recognition accuracy of each iteration in the iteration process is obtained, and the average slope of the plotted iteration round-recognition accuracy curve is calculated.

[0085] Specifically, if the average value is greater than the preset average value, it indicates that the model still experiences fluctuations in recognition accuracy during the final iteration stage. This is mainly due to an unreasonable batch size setting in the training parameter configuration, which leads to a significant increase in gradient noise and ultimately hinders the training convergence process. Therefore, the batch size is adjusted based on the difference between the average value and the preset average value, where the preset difference between the average value and the preset average value is P0 = 2%. The comparison process between the difference P and the preset difference P0 is as follows:

[0086] If the difference P is less than or equal to the preset difference P0, the batch size will be adjusted to twice the original batch size.

[0087] If the difference P is greater than the preset difference P0, the batch size will be adjusted to 4 times the original batch size.

[0088] Specifically, after retraining the second recognition model with adjusted batch size, if the recognition performance is still unsatisfactory, the recognition accuracy of the training and test sets in the SAR target data is obtained respectively, and the absolute value of the difference between the recognition accuracy of the training set and the test set is calculated. If the absolute value is greater than a preset difference, it indicates that the network complexity of the recognition model is high and does not match the data scale. In this case, the dropout rate in the Dropout mechanism needs to be adjusted based on the ratio of the absolute value to the preset difference. The preset ratio of the absolute value to the preset difference is Q0 = 1.5. The comparison process between the ratio Q0 and the preset difference is as follows:

[0089] If the ratio Q is less than or equal to the preset ratio Q0, the dropout rate in the Dropout mechanism will be adjusted to 1.1 times the original dropout rate.

[0090] If the ratio Q is greater than the preset ratio Q0, the dropout rate in the Dropout mechanism will be adjusted to 1.3 times the original dropout rate, that is, the step-wise adjustment is expressed as the adjustment step size.

[0091] Specifically, after adjusting the loss rate, the recognition status of the second recognition model is re-checked. If the recognition status is still unsatisfactory, the integral of the obtained loss curve is recalculated. If the integral is greater than the preset integral, it indicates that the loss curve has rebounded. This is because the learning rate is too large, causing the model to "skip" the optimal solution during the optimization process, thus forming oscillations. Therefore, the learning rate period is adjusted by reducing the ratio of the integral to the preset integral using the gradient descent algorithm. The preset ratio of the integral to the preset integral is R0 = 1.5. The comparison process between the ratio R of the integral to the preset integral and the preset ratio R0 is as follows:

[0092] If the ratio R is less than or equal to the preset ratio R0, the period for adjusting the learning rate of the gradient descent algorithm will be adjusted to 0.9 times the original period.

[0093] If the ratio R is greater than the preset ratio R0, the period for adjusting the learning rate of the gradient descent algorithm will be adjusted to 0.83 times the original period.

[0094] Specifically, the recognition status after adjusting the learning rate is re-examined. If the recognition status is not satisfactory, the learning rate adjustment cycle of the gradient descent algorithm is repeated at least once until the number of adjustments is less than the preset number and the recognition status is satisfactory, or the number of adjustments is equal to the preset number and the adjustment stops.

[0095] If the recognition status is still unqualified after stopping the adjustment, obtain the recognition accuracy corresponding to multiple preset iteration rounds after adjustment, and calculate the average of multiple recognition accuracy rounds. The preset iteration rounds are the rounds corresponding to the later stage of the iteration when it tends to be stable.

[0096] If the average value is less than the preset average value, it indicates that the recognition status after training is unqualified due to the small amount of data in the collected dataset and uneven category classification. By adjusting the adjustment step size of the geometric transformation parameters in image enhancement, the minority samples can be diversified and the dataset size can be increased. The adjustment step size of the geometric transformation parameters in data enhancement is adjusted based on the ratio of the average value to the preset average value, where the preset ratio T0 = 0.92. The specific comparison process between the ratio T0 and the preset average value is as follows:

[0097] If the ratio T is greater than the preset ratio T0, the adjustment step size of the geometric change parameter in data augmentation will be adjusted to 0.91 times the original adjustment step size;

[0098] If the ratio T is less than or equal to the preset ratio T0, the adjustment step size of the geometric transformation parameters in data augmentation will be adjusted to 0.86 times the original adjustment step size. Among them, the geometric transformation parameters with adjustable adjustment step size include flipping and rotation. For example, the initial probability of horizontal flipping is 0.5, and it is adjusted accordingly based on the ratio T.

[0099] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0100] Example 1

[0101] This embodiment identifies three types of vehicles, and the experimental design is as follows:

[0102] I. Experimental conditions and parameter design

[0103] The experimental environment setup, training, and testing in this paper were all conducted under the Windows 11 64-bit operating system. The relevant hardware devices and software environment platform selections are shown in Table 1:

[0104] Table 1 Experimental Environment

[0105] hardware software Processor: Intel Core i9 Compiler: PyCharm Memory: 64GB Language: Python 3.6 Graphics card: NVIDIA GeForce X4060 Ti Deep learning framework: Keras Video memory: 8GB

[0106] The Keras framework used in this embodiment is a high-level deep learning framework. Essentially, it is a high-level neural network library written in Python that seamlessly integrates with various low-level deep learning libraries, such as TensorFlow. Keras provides a simpler and easier-to-use interface and offers flexible high-level APIs. Users can directly add various active layers (such as convolutional layers, pooling layers, and fully connected layers) to build deep network models.

[0107] Finally, the recognition accuracy of the three types of vehicles was tested using a test set derived from the measured images, in order to evaluate the model's performance. The division of the SAR image training and test sets in the experiment is shown in Table 2.

[0108] Table 2 Details of Dataset Partitioning

[0109]

[0110] To verify the effectiveness of the improvement strategy, the initial network model needs to be adjusted according to the different optimization strategies mentioned above. Before that, the basic model without any optimization strategy is defined as GLN-1. This model uses ReLU as the non-linear activation function for both convolutional and fully connected layers, without any attention mechanism or improvement to the classic cross-entropy loss function. Subsequent optimization strategies are all based on this model, but the initial parameter settings of all models remain unchanged. The constant term φ is set to 1e-6, the learning rate decay coefficient α is set to 0.9, the weight decay coefficient in the loss function is set to 1e-4, the total number of training epochs is set to 100, and the batch size is set to 32. One iteration is equivalent to training once with batch size samples, and one epoch is equivalent to training once with all samples in the training set. The Dropout overfitting control parameter is set to 0.5, and the Adam optimizer is used when performing gradient descent (the initial learning rate is set to 0.0005).

[0111] Table 3 Model Structure and Parameter Settings

[0112]

[0113] The detailed structure and parameters of the model are shown in Table 3. After determining the experimental environment and parameter settings, this paper conducts relevant comparative experiments on existing SAR image datasets based on the GLN-1 basic network model to verify the positive impact of the optimization strategy on the final recognition performance of the model. Since the experimental results obtained from a small number of experiments may contain experimental errors and randomness, each different comparative experiment is performed 1000 times to reduce errors and randomness, making the experimental results more realistic and reliable.

[0114] II. Comparative Experiments and Result Analysis of Algorithm and Structure Optimization

[0115] (1) Introducing the LeakyReLU nonlinear activation function

[0116] Although the ReLU nonlinear activation function has gained favor among researchers in most deep learning tasks, it still carries the risk of "neuron death," where negative inputs lead to zero gradients and prevent timely weight updates. LeakyReLU, however, provides a small gradient slope in the negative range, thus preserving some information in this region and improving the stability of network iterations. In our experiments, we replaced the ReLU activation function in the convolutional module of the Xception backbone in the original GLN-1 model with LeakyReLU, while keeping the other network layers and parameters unchanged. This modified network model was named GLN-2. This allows us to verify whether the improved activation function optimizes the training of the network model. Table 4 shows the experimental results of applying the two nonlinear activation functions to the SAR target recognition process.

[0117] Table 4 shows the experimental results of the models using ReLU and LeakyReLU activation functions, respectively.

[0118] Model GLN-1 GLN-2 Recognition accuracy (%) 75.38 79.97

[0119] As shown in the table above, using LeakyReLU as the nonlinear activation function of the model designed in this paper results in a higher recognition rate than using the ReLU function, indicating that the network iteration is more stable, thereby improving the model's recognition accuracy. Figure 7 As shown, the training curves are for models using ReLU and LeakyReLU activation functions respectively. Although LeakyReLU requires more time to calculate gradients in the negative interval during training, both methods achieve relatively stable accuracy in almost the same training batches, with no significant difference in training time. Overall, LeakyReLU performs more stably and reliably in SAR target recognition.

[0120] (2) Introducing a channel-space fusion attention mechanism

[0121] As mentioned above, the fusion mechanism of channel attention and spatial attention can theoretically play a crucial role in training the network model. Because SAR images contain speckle noise caused by the superposition of scatterer phases, channel attention is first used to filter out noisy channels, followed by spatial attention to locate key feature regions in the SAR image, forming a dual visual focus of "channel-space". This method also improves the network model's adaptability to complex scenes. The channel attention module uses the classic SE module, which allows the model to adaptively learn the importance weights of each channel, making the network model pay more attention to the responses of key target features, suppressing interference from irrelevant or redundant channels, and improving computational efficiency.

[0122] Based on the GLN-1 basic model, a channel-spatial fusion attention mechanism is added. The impact of this approach on the final model performance is investigated. Specifically, after the output of the last convolutional layer of the Xception network and before global average pooling, a channel-spatial attention mechanism is added. All other layers and initial parameters remain unchanged, and this model is defined as GLN-3. SAR target recognition experiments are conducted on both the GLN-1 and GLN-3 models. The experimental results are shown in Table 5.

[0123] Table 5. Experimental results of the model incorporating spatial-channel attention and the original model.

[0124] Model GLN-1 GLN-3 Recognition accuracy (%) 75.38 85.35

[0125] As shown in the table above, the initial network model GLN-1, without the attention mechanism, has a significantly lower SAR target recognition rate than the GLN-3 model, which incorporates the spatial-channel attention mechanism. This is because the channel-spatial attention mechanism simultaneously enhances the model's ability to extract key features in both the channel and spatial dimensions of SAR images. This allows the improved model to better focus on and lock onto the detailed features of the target itself in SAR images generated in complex scenes. Figure 8 As shown, the training curves of the model with spatial-channel attention and the original model are displayed. During the training phase, the model with spatial-channel attention mechanism has a slightly better training accuracy than the original model, indicating that this optimization strategy improves the robustness of the model.

[0126] (3) Improve the loss function

[0127] Specifically, regarding the mathematical expression of the Focal Loss function mentioned above, we set the sample adjustment parameter γ to 2 and the dynamic adjustment factor α... t The value was set to 0.25. Then, the traditional cross-entropy loss function in GLN-1 was replaced with the Focal Loss loss function, while the rest of the network and parameter settings remained unchanged. This model was named GLN-4. Equal numbers and sizes of SAR data samples were then fed into both the GLN-1 and GLN-4 models for thorough training. The final experimental results are shown in Table 6.

[0128] Table 6 shows the experimental results of the models using Focal Loss and cross-entropy loss functions, respectively.

[0129] Model GLN-1 GLN-4 Recognition accuracy (%) 75.38 82.07

[0130] As shown in the table above, the Focal Loss function dynamically adjusts the input SAR image samples, ultimately significantly enhancing the improved model's ability to identify SAR targets, specifically resulting in a marked increase in the model's target recognition success rate. Furthermore, by focusing more on difficult-to-classify SAR samples, the Focal Loss function reduces ineffective, iterative training on simple samples, thereby improving the learning efficiency of all key image features. Figure 9 As shown, the training curves of the model using Focal Loss and cross-entropy loss functions are respectively. The GLN-4 model using Focal Loss speeds up the convergence, indicating that this method can guide the model parameters to update in an effective direction and has a greater advantage in overall training efficiency.

[0131] (4) Combining the three optimization strategies

[0132] The experimental results above show that the three optimization strategies described in this paper can ultimately enable the SAR target recognition model based on convolutional neural networks to achieve higher recognition efficiency and stronger classification performance. This experiment combines the three optimization strategies in the network model to evaluate their impact on the final recognition results. The LeakyReLU activation function, channel-space hybrid attention mechanism, and Focal Loss loss function are all introduced into the initial basic model GLN-1, while other parameters and network layers remain unchanged, defining this model as GLN-5. This model is the final optimized model, and the models corresponding to different optimization strategies are shown in Table 7.

[0133] Table 7 Models corresponding to different optimization strategies

[0134]

[0135] Five convolutional neural network models (GLN-1, GLN-2, GLN-3, GLN-4, and GLN-5) were subjected to SAR image target recognition experiments with identical steps, each run 1000 times. The final experimental results are as follows: Figure 10 and Figure 11 As shown, these are the average recognition rates of each model and the average recognition rates of the models using the Adam gradient optimizer, SGD, and MBGD, respectively:

[0136] Depend on Figure 10 The experimental results show that when the LeakyReLU activation function, channel-space hybrid attention mechanism, and Focal Loss loss function are all introduced into the original network model, the model has the best recognition effect on the measured SAR targets. The initial model without any optimization strategy has the lowest recognition success rate among the five models, and its recognition performance is poor. Figure 11 This paper demonstrates the impact of using the Adam optimizer and two other traditional gradient descent algorithms on SAR target recognition performance across all models used in the experiments. Multiple experiments show that the Adam optimizer contributes the most to improving the recognition performance of the proposed model.

[0137] III. Summary of Experimental Results of the Algorithm

[0138] (I) SAR Image Recognition Results of Three Types of Target Vehicles Measured by Radar

[0139] Based on the SAR target recognition algorithm designed in this paper using a convolutional neural network, and by integrating the optimization strategies of the aforementioned algorithms and structures, and after continuous verification, the optimal model for identifying and classifying three types of vehicle targets—school buses, Tesla Model 3s, and Volkswagen Santanas—was obtained in a real radar test environment; this model is the GLN-5 model. This model integrates three improvement strategies to achieve the best recognition performance. A total of 1000 experiments were conducted. The experimental results show that the average recognition rate for the three types of vehicle targets was 87.63%. School buses, due to their significantly larger size and prominent outline features, were easier to identify, achieving an accuracy rate of 89.78%. Although vans are similar in length, their outlines and structural details differ, resulting in a good recognition rate. The experiments verify that the network model built in this paper can effectively identify SAR targets.

[0140] (II) Comparison of recognition performance of different methods

[0141] To further verify the superiority of the model designed in this paper, which uses Xception as the core network and combines various optimization strategies, for SAR target identification and classification, the test results obtained in this paper are compared with those obtained by Support Vector Machine (SVM), Monogellic Sparse Representation Classification (MSRC), Kernel Linear Representation (KLR), and traditional convolutional neural networks. Specific results are shown in Table 8.

[0142] Table 8 Results of different methods for SAR target identification

[0143]

[0144] As shown in the table above, the SAR target recognition model based on the Xception network designed in this paper has the highest average recognition rate and the best recognition effect, which once again verifies the reliability and stability of the method.

[0145] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0146] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for SAR network target recognition based on Xception, characterized in that, The method comprises the following steps: obtaining SAR images generated by electromagnetic simulation and SAR target data measured by millimeter wave radar; preprocessing the SAR images and the SAR target data, wherein the preprocessing at least includes data cleaning and data enhancement; pre-training an identification model containing a Dropout mechanism using the preprocessed SAR images to obtain a first identification model; adjusting network layer parameters of the first identification model using a training set divided in the preprocessed SAR target data, and training the adjusted first identification model by setting training parameters to obtain a second identification model, wherein the training parameters at least include batch size; inputting a test set divided in the preprocessed SAR target data into the second identification model, and outputting an identification accuracy of the second identification model; determining an identification state based on the identification accuracy, adjusting the batch size based on the identification state, and adjusting a dropout rate of the Dropout mechanism in the pre-training process based on the identification state after adjusting the batch size; wherein the identification model comprises an improved Xception network, a global average pooling layer and a fully connected layer; the process of pre-training the identification model containing the Dropout mechanism using the preprocessed SAR images to obtain the first identification model comprises: dividing the preprocessed SAR images into a training set and a validation set according to a preset proportion; inputting the training set into the improved Xception network to output a feature map, wherein the improved Xception network is obtained by adding a channel-spatial attention mechanism after the last layer of the Xception network, and replacing the Relu activation function of a preset convolution layer in the Xception network with a LeakyReLU activation function; inputting the feature map into the global average pooling layer to calculate the average value of each channel in the feature map to obtain a feature vector; reducing the feature dimension of the feature vector using the fully connected layer, and outputting an original score of a target class corresponding to a classification task; normalizing a plurality of the original scores respectively using a Softmax classifier to obtain a prediction probability, and iterating the identification model based on a loss function calculated based on the prediction probability to obtain the first identification model.

2. The Xception-based SAR network target recognition method according to claim 1, characterized in that, The process of pre-training the identification model containing the Dropout mechanism using the preprocessed SAR images to obtain the first identification model further comprises: freezing a plurality of layers in the Xception network in an initial pre-training period; calculating a validation parameter of the identification model after pre-training based on the validation set; unfreezing the preset layers at least once based on the validation parameter, and retraining after each unfreezing; calculating a loss function after each training, and recording the identification model after training as the first identification model if the loss function is less than or equal to a preset loss function.

3. The Xception-based SAR network target recognition method according to claim 1, characterized in that, The process of reducing the feature dimension of the feature vector using the fully connected layer and outputting an original score of a target class corresponding to a classification task comprises: inputting the feature vector into the fully connected layer to obtain a first feature vector after reducing the feature dimension; The first feature vector is optimized based on a ReLU activation function, L2 regularization, and a Dropout mechanism to obtain a second feature vector; The second feature vector is re-input into a fully connected layer to obtain a third feature vector with reduced feature dimension; The third feature vector is cooperatively acted on by the ReLU activation function, L2 regularization, and Dropout mechanism to output an original score of a target category of a corresponding classification task in the fully connected layer.

4. The Xception-based SAR network target recognition method according to claim 1, characterized in that, The process of adjusting network layer parameters of the first recognition model by using the preprocessed training set divided in the SAR target data to obtain a second recognition model includes: replacing the classification layer of the first recognition model with a fully connected layer and a Softmax activation layer according to the number of categories of the target; inputting the training set divided in the SAR target data into the improved first recognition model for training; calculating the gradient of the loss function on the network parameters layer by layer based on the chain rule, and updating the iterative network layer parameters according to the gradient using the gradient descent algorithm until the absolute value of the slope of the loss function curve output in the continuous preset number of iterations is less than the preset threshold, stopping the model training, and recording the trained model as the second recognition model.

5. The Xception-based SAR network target recognition method according to claim 4, characterized in that, The process of determining the recognition state based on the recognition accuracy and adjusting the batch size based on the recognition state includes: when the recognition state is unqualified based on the recognition accuracy, obtaining the recognition accuracy of each round in the training iteration process; drawing an iteration round-recognition accuracy curve and calculating the average of the absolute value of the slope of the curve; if the average is greater than a preset average, adjusting the batch size based on the difference between the average and the preset average; wherein the recognition state is unqualified when the recognition accuracy is greater than a preset accuracy.

6. The Xception-based SAR network target recognition method according to claim 5, characterized in that, The process of adjusting the batch size based on the difference between the average and the preset average includes: increasing the batch size based on the difference between the average and the preset average, and the increase amplitude of the batch size is proportional to the difference.

7. The Xception-based SAR network target recognition method according to claim 6, characterized in that, The method further includes: retraining the second recognition model after adjusting the batch size, and if the trained recognition state is unqualified, obtaining the recognition accuracy of the training set and the test set in the SAR target data, respectively; calculating the absolute value of the difference between the recognition accuracy of the training set and the test set; if the absolute value is greater than a preset difference, adjusting the dropout rate in the Dropout mechanism based on the ratio of the absolute value to the preset difference.

8. The Xception-based SAR network target recognition method according to claim 6, characterized in that, The process of adjusting the dropout rate in the Dropout mechanism based on the ratio of the absolute value to the preset difference includes: increasing the dropout rate in the Dropout mechanism based on the ratio of the absolute value to the preset difference, and the increase amplitude of the dropout rate in the Dropout mechanism is proportional to the ratio.

9. The Xception-based SAR network target recognition method according to claim 8, characterized in that, The method further includes: re-detecting the recognition state after adjusting the dropout rate, and if the recognition state is unqualified, re-obtaining the Loss curve; calculating the integral of the Loss curve; if the integral is greater than a preset integral, reducing the period of adjusting the learning rate of the gradient descent algorithm based on the ratio of the integral to the preset integral, and the reduction amplitude of the period is proportional to the ratio.

Images