An ISAR target recognition method, system and device based on deep learning

Abstract

The application discloses an ISAR target recognition method, system and device based on deep learning, relates to the technical field of radar target recognition, and comprises the following steps: firstly, a database is constructed through FEKO simulation, and is processed through labeling and data enhancement; secondly, a convolutional neural network with residual connection is built and pre-trained; thirdly, a three-scale feature pyramid YOLOV3 detection network is constructed, and anchor points are clustered by using k-means; then, joint training and optimization are carried out; finally, two types of detection systems are tested and generated, and target recognition is realized. The ISAR target recognition method, system and device based on deep learning adopt a feature pyramid network and k-means clustering anchor points, design nine groups of special anchor boxes according to the width-height distribution characteristics of ISAR targets, combine an up-sampling feature fusion strategy, realize effective detection of multi-pixel scale targets, and compared with a traditional YOLOV3 fixed anchor point scheme, the recall rate of cross-scale targets is significantly improved, and the industry pain points of small target missed detection and large target boundary blur are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of radar target recognition, and particularly relates to an ISAR target recognition method, system and device based on deep learning. BACKGROUND

[0002] In the field of target imaging and recognition, SAR / ISAR imaging technology has advantages such as all-weather and all-day imaging, and has become an important means of target detection and recognition. SAR is often used as a sensor for imaging moving targets, while ISAR is often used as a fixed radar for imaging moving targets.

[0003] With the development of artificial intelligence and machine learning theories and methods, intelligent target recognition technology has excellent recognition ability under the training of a large number of samples. However, in actual application, the sample acquisition in many application scenarios is difficult and requires high cost, resulting in a very small amount of samples that can be used for training, and not enough samples to train the target recognition model. For example, it is usually difficult to obtain a large number of ISAR images of aircraft and their corresponding models, resulting in a lack of ISAR samples of air targets. For air targets, this lack mainly manifests in the lack of samples in the angle dimension.

[0004] However, in the implementation process of the above technical solutions, at least the following technical problems are found:

[0005] With the development of deep learning technology, the end-to-end recognition method based on convolutional neural network (CNN) exhibits excellent performance in optical image target recognition through automatic learning of multiple feature layers. However, there are essential differences between ISAR images and optical images in data characteristics: the gray scale distribution is significantly affected by speckle noise, the target shape presents nonlinear changes with the observation angle, and the size difference of different targets leads to more prominent multi-scale detection requirements. The existing ISAR recognition method based on deep learning mainly faces the following technical bottlenecks:

[0006] 1. Inefficient multi-scale target detection: the traditional detection framework has poor adaptability to targets with large size changes in ISAR images, and lacks effective multi-scale feature fusion mechanism;

[0007] 2. Limited boundary box positioning accuracy: the anchor box prior parameters have low matching degree with the actual width-height ratio of ISAR targets, and the loss function is not optimized for target size difference, resulting in coexistence of small target missed detection and large target positioning deviation problems;

[0008] 3. Weak anti-interference ability in complex background: the existing network has insufficient ability to distinguish between strong scattering background and target in ISAR images, and is easily disturbed by sidelobe noise, leading to high false alarm rate. Therefore, we propose an ISAR target recognition method, system and device based on deep learning. SUMMARY

[0009] (1) Technical problems solved

[0010] In view of the deficiencies of the prior art, the present application provides an ISAR target recognition method, system and device based on deep learning, which solves the technical problems of insufficient multi-scale target detection efficiency and limited boundary box positioning accuracy of the existing ISAR target recognition technology.

[0011] (2) Technical solutions

[0012] To achieve the above purpose, the present application is realized by the following technical solutions:

[0013] An ISAR target recognition method based on deep learning, the recognition method comprising the following steps:

[0014] An electromagnetic simulation of the target model is performed using FEKO, and the electromagnetic simulation results are imaged to construct a target recognition database;

[0015] The images in the database are preprocessed, including target labeling, coordinate normalization and data enhancement operations, to obtain an image set;

[0016] A convolutional neural network model is constructed, and the constructed convolutional neural network model is initialized, after which a deep learning training framework is used to train MSTAR 10-class targets, the network is optimized using the gradient descent method, and the network pre-training result is obtained;

[0017] On the basis of the convolutional neural network model, a YOLOV3 detection network is constructed, the yolo detection layer is replaced with a fully connected layer, a feature pyramid network structure is used, multi-scale is set, local feature interaction of the feature map is performed using a convolution kernel, the detection network model is initialized, then the YOLOV3 detection network is trained, the network is adjusted using the gradient descent method, and the network weight suitable for detection is obtained;

[0018] The trained detection network model is used to detect the test image set, the target detection and recognition result is obtained by setting the parameters, and the corresponding algorithm application mode is established according to the application scene requirements.

[0019] Preferably, the convolutional neural network model initialization is performed by replacing the fully connected layer with a global average pooling layer for prediction classification, using a 1x1 convolution to compress the feature representation between 3x3 convolutions, using skip connections to construct a residual network, and adding BN after the convolution layer;

[0020] After the convolutional neural network model initialization is completed, a deep learning training framework is used to train MSTAR 10-class targets;

[0021] Among them, the network pre-training is trained on the convolutional neural classification network on MSTAR 10 target classes through Darknet as a framework, the training parameters are set, the standard data enhancement method is used, the small batch gradient descent method is used to change the network propagation direction, and the weight after iteration is selected as the network pre-training result.

[0022] Preferably, in the cross-scale prediction step of constructing the detection network:

[0023] First, the feature map is extracted and up-sampled to twice the original size, and the initial network is extracted for feature extraction;

[0024] The up-sampled feature map is combined with the features extracted by the initial network, and then a convolutional layer is added for tensor prediction;

[0025] Repeat the above operation to complete the box scale prediction;

[0026] For different scale prediction processing, set the corresponding feature mapping step, use the k-means clustering method to determine the anchor points corresponding to different scales, and in the radar data set, the 9 anchor point clusters are: (91x81); (120x125); (254x76); (71x295); (328x66); (170x242); (241x174); (306x219); (219x306).

[0027] Preferably, in the loss function step of the detection network, the detection is converted into regression, the mean square error loss function is used, different weight parameters are set according to the relationship between the boundary box and the target, whether the target exists in the boundary box, and the size of the boundary box, the square root of the boundary is predicted, and finally the mean square error is used for numerical calculation, and the loss function expression is:

[0028] ;

[0029] In the formula, is the number of grids, B is the number of boundary boxes predicted for each grid, is whether the jth boundary box in the ith grid is responsible for predicting the target, is whether the jth boundary box in the ith grid is not responsible for predicting the target, is whether the ith grid contains a target, , are the center coordinates, width and height of the real boundary box, , are the center coordinates, width and height of the predicted boundary box, , are the target confidence of the real and predicted targets, respectively, , These represent the true and predicted target class probabilities, respectively. , These are the weight parameters corresponding to the positioning error and the parameter value confidence, respectively. If there is no target in the bounding box, the corresponding weight parameter is... Conversely, the corresponding weight parameters .

[0030] Preferably, in the bounding box prediction step of the detection network, the bounding box coordinates are predicted using anchor points, based on image offset, bounding box width, and height. The prediction formula is as follows:

[0031] ;

[0032] In the formula, These are the center coordinates, width, and height of the predicted bounding box, respectively. The center coordinates, width, and height of the anchor point frame are given respectively. , The coordinates of the center point of the anchor point frame. These are the offset coordinates of the image. This defines the width and height of the anchor point frame.

[0033] Preferably, in the category prediction step of the detection network construction and training phase, a multi-label classification method is adopted, using a logistic classifier and a binary cross-entropy loss method for category prediction.

[0034] Preferably, an algorithm application mode is established based on the scenario, specifically, an algorithm application scenario is established for batch input detection output results or single input detection output results.

[0035] A deep learning-based ISAR target recognition system, comprising:

[0036] The data infrastructure building module is used to build the data infrastructure and obtain a target recognition database through electromagnetic simulation and imaging.

[0037] The image preprocessing module is used to preprocess images, including labeling, coordinate normalization, and data augmentation of images in the database to obtain an image set.

[0038] The convolutional neural network classification module is used to build and train the convolutional neural network. It initializes the network with a specific structure design, trains it on the MSTAR10 target class based on the Darknet framework, sets parameters and optimizes the network to obtain pre-training results.

[0039] The detection network processing module is configured to build a YOLOV3 detection network, replace a fully connected layer with a yolo detection layer, adopt a feature pyramid network structure, set multiple scales, utilize a convolution kernel to interact local features of a feature map, initialize a detection network model, then train the YOLOV3 detection network, adopt a gradient descent method to adjust the network, and obtain network weights suitable for detection.

[0040] The test and application module is configured to detect a test image set by using the trained detection network, set parameters to obtain results, and establish an algorithm application mode according to a scene.

[0041] Preferably, the convolutional neural network processing module comprises:

[0042] The network building submodule is configured to replace a fully connected layer with a global average pooling layer for prediction and classification, use a 1x1 convolution to compress feature representation between 3x3 convolutions, build a residual network by using a skip connection, add BN after a convolution layer to prevent overfitting, and initialize a network model.

[0043] The network training submodule is configured to train the convolutional neural network based on a Darknet framework on MSTAR10 target classes, set an initial learning rate, a decay mode, a weight decay rate, a momentum, and a batch size, use a standard data enhancement method, adopt a small batch gradient descent method to change a network propagation direction, and select weights of a specific iteration number as a network pre-training result.

[0044] An ISAR target recognition device based on deep learning, comprising a processor and a memory, wherein the memory stores a computer program, and the processor implements an ISAR target recognition method based on deep learning when executing the computer program.

[0045] (Three) beneficial effects

[0046] Data-driven high-precision dataset construction and enhancement

[0047] First, relying on the FEKO electromagnetic simulation technology, a multi-view and multi-frequency ISAR image database is built, and through target labeling, coordinate normalization, and rotation / scaling / noise adding data enhancement operations, the problems of difficult real ISAR data acquisition and insufficient sample diversity are effectively solved. Compared with the traditional artificial feature extraction method, the data preprocessing strategy improves the model training sample size and significantly enhances the network's generalization ability to complex poses and noisy environments.

[0048] Secondly, a residual network structure containing a global average pooling layer, a 1x1 convolution compression and a BN layer is constructed, the gradient vanishing problem of a deep network is relieved through a skip connection, the feature extraction capability is maintained while reducing the parameters of a full connection layer, the initial weights of the network are provided with a priori cognition of basic scattering characteristics of a radar target based on pre-training of the Darknet framework on the MSTAR dataset, and the training convergence speed is improved, thereby providing a strong semantic feature basis for a subsequent detection network;

[0049] Thirdly, a feature pyramid network and k-means clustering anchor points are adopted, 9 groups of exclusive anchor boxes are designed according to the width-height distribution characteristics of the ISAR target (such as (71x295) for adapting to a long and narrow target and (328x66) for adapting to a flat target), and an up-sampling feature fusion strategy is combined to realize effective detection of targets with a pixel size of 52x52 to 13x13, compared with a traditional YOLOV3 fixed anchor scheme, the recall rate of the application to cross-scale targets is significantly improved, and the industry pain points of small target missed detection and large target boundary blur are solved;

[0050] Fourthly, an innovative mean square error loss function containing square root prediction of a bounding box is designed, the gradient contribution of large and small targets is balanced by taking the square root of the width and height parameters, and a weight parameter 、 is introduced to differentially constrain coordinate positioning, confidence and class prediction, experiments show that the loss function reduces the positioning error (IOU) of the bounding box and significantly improves the target discrimination ability in a strong scattering noise environment;

[0051] Fifthly, the detection network supports batch input (such as real-time echo data stream processing of a radar) and single input (such as offline image analysis) dual modes, the dynamic balance of detection speed and accuracy is realized through setting a confidence threshold and a non-maximum suppression parameter, in a GPU environment, the batch detection throughput and the single image detection time consumption can be effectively optimized, and the application requirements of multi-scene applications such as real-time identification of a machine-mounted radar and ground intelligence analysis are met. BRIEF DESCRIPTION OF DRAWINGS

[0052] The above description is only a summary of the technical scheme of the application, in order to more clearly understand the technical means of the application and can be implemented according to the content of the description, the following preferred embodiments of the application are described in detail with reference to the drawings.

[0053] Figure 1 The above description is only a summary of the technical scheme of the application, in order to more clearly understand the technical means of the application and can be implemented according to the content of the description, the following preferred embodiments of the application are described in detail with reference to the drawings.

[0054] Figure 2 The above description is only a summary of the technical scheme of the application, in order to more clearly understand the technical means of the application and can be implemented according to the content of the description, the following preferred embodiments of the application are described in detail with reference to the drawings.

[0055] Figure 3 The above description is only a summary of the technical scheme of the application, in order to more clearly understand the technical means of the application and can be implemented according to the content of the description, the following preferred embodiments of the application are described in detail with reference to the drawings.

[0056] Figure 4 A schematic diagram for prediction of a bounding box in an embodiment of the application;

[0057] Figure 5 A schematic diagram for detecting a network in an embodiment of the application;

[0058] Figure 6 A schematic diagram for a loss function in an embodiment of the application;

[0059] Figure 7 A schematic diagram for a detection process in an embodiment of the application. DETAILED DESCRIPTION

[0060] The embodiment of the application provides an ISAR target recognition method, system and device based on deep learning, solves the technical problems of insufficient multi-scale target detection efficiency and limited boundary box positioning precision of the existing ISAR target recognition technology, adopts a feature pyramid network and k-means clustering anchor point, designs 9 groups of exclusive anchor boxes (such as (71x295) for adapting to long and narrow targets and (328x66) for adapting to flat targets) according to the width-height distribution characteristics of ISAR targets, combines an up-sampling feature fusion strategy, realizes effective detection of 52x52 to 13x13 pixel scale targets, compared with a traditional YOLOV3 fixed anchor point scheme, the recall rate of the application for cross-scale targets is significantly improved, and the industry pain points of small target missed detection and large target boundary blur are solved; a mean square error loss function of a square root prediction of a boundary box balances the gradient contribution of large and small targets by taking square roots of width and height parameters, and a weight parameter is introduced 、 differential constraint coordinate positioning, confidence and class prediction, experiments show that the loss function reduces the positioning error (IOU) of the boundary box, and significantly improves the target discrimination ability in a strong scattering noise environment.

[0061] Embodiment: In order to solve the above technical problems, the technical scheme in the embodiment of the application is as follows:

[0062] Target recognition of the project is mainly target recognition of radar imaging target results. First, FEKO is used to perform electromagnetic simulation on the model of the target, and the electromagnetic simulation results are imaged to obtain a database required for target recognition. Part of the geometric model is as shown in Figure 1 , and the radar simulation image is as shown in Figure 1 and Figure 2 .

[0063] After the database is established, the target is identified according to the database, and the idea is to convert the target identification problem into a regression problem, introduce a residual network structure and build a feature pyramid, use a 1x1 convolution kernel to interact with each feature layer, and then cross-scale to determine the target. Through this way, the feature extraction is more convenient, and the target detection process is more rapid and synchronous, and the K-means clustering ensures that the target representation process tends to be simplified.

[0064] In order to facilitate the description of the content of the network, the following terms are defined first:

[0065] Definition 1: Non-maximum suppression (NMS): During the detection process, due to the existence of objective factors, a target will be judged multiple times, which will undoubtedly affect the detection result. How to avoid it? The definition of non-maximum suppression is to keep the result most similar to the target and discard other judgment values. This method saves a lot of manpower and material resources and effectively improves the identification efficiency.

[0066] Definition 2: Target confidence threshold: As the name implies, the threshold means the boundary, and the basis for judging whether it can pass is the target score. The bounding box below the value is discarded.

[0067] Definition 3: Dimension clustering: The manual determination of the region makes the learning complex, and the introduction of dimension clustering solves the problem, which is defined as clustering the markers with an algorithm, that is, using another simpler way to represent the number of values of each type of region.

[0068] Definition 4: Fine-grained features: The features are connected together in this way to increase the number of channels, and the commonly used methods include feature mapping, etc.

[0069] Definition 5: K-means clustering: This method can improve the model expression ability and make the steps more convenient, so it is used to get more matching prior boxes in the training set bbox, and the distance operation is represented by IoU, and the distance formula is:

[0070]

[0071] The whole algorithm idea is: after clustering the width and height of bbox, multiply the output feature size by the obtained ratio value to get the final result. Assuming that the input is 608x608, then the output convolution layer is 19x19.

[0072] Definition 6: IoU, recall rate (R), accuracy (P), mAP

[0073] IoU = intersection of detection value and true value / union of the two;

[0074] Recall rate = number of correctly detected targets / (number of correctly detected targets + number of missed targets)

[0075] Accuracy = Number of correct targets detected / (Number of correct targets detected + Number of incorrect targets detected)

[0076] mAP = Average precision averaged over all classes in the dataset.

[0077] Definition 7: Anchor box, the predicted bounding box is theoretically correct, but some problems will arise after application, so we mostly do not use this method, and the log space transformation is more reasonable for prediction, and the resulting results are applied to the anchor box to achieve the purpose.

[0078] Definition 8: Target score, the definition of target score is the relationship between the resulting results and the bounding box, that is, the probability.

[0079] Definition 9: Class confidence, the definition of class confidence is to judge whether the object belongs to the class of the database, and calculate the probability.

[0080] Definition 10: Dimension of bounding box, the output result is first subjected to spatial transformation, and the product of the resulting value and the anchor point can be predicted, and then normalized.

[0081] 1. Image preprocessing:

[0082] We first manually label the targets such as aircraft, fighter, helicopter, missile, etc. in the radar image obtained by FEKO using the labelme software. Normalize the labeled coordinates to generate the coordinates of the region center and the width and height of the region for subsequent detection. Then rotate, translate, flip and add noise interference data to the image to increase our dataset and finally get the image set.

[0083] 2. Build a convolutional neural network for classification, initialize the network model:

[0084] First we build a classification network as shown in Table 1:

[0085] Table 1 Classification network structure table

[0086]

[0087] The global average pooling layer replaces the fully connected layer for prediction classification; and 1x1 convolution is used between 3x3 convolution to compress feature representation; Some skip connections are used in the network to build a residual network, making the network more powerful and more conducive to feature extraction; In the process of extraction, it is found that overfitting may occur, and the solution is to add BN after the convolution layer. Since this classification network is a fully convolutional network (FCN), it is necessary to integrate the image to process it in batches.

[0088] 3. Train the convolutional neural network for classification:

[0089] The training process uses the Darknet framework to train the MSTAR10 target, and the initial learning rate is set to 0.01, the decay method is a four-level polynomial, the weight decay rate and momentum are 0.0005 and 0.9 respectively, and the standard data enhancement method is used: random cropping, rotation, exposure, hue, saturation, translation, and flipping. In the case of CPU Intel Core-I7, memory 64G, and graphics card NVIDIA GeForce GTX1080Ti, we set the batch size to 128.

[0090] The small batch gradient descent method is used to change the propagation direction of the network, and the loss function curve is observed and the number of iterations is changed. The weight of 3600 iterations is selected as the result of the future network pre-training.

[0091] 4. Build the YOLOV3 detection network and initialize the network model:

[0092] The idea of building is to replace the previous fully connected layer with the yolo detection layer under the premise of the convolutional neural network classification network. The network is built according to the idea of feature pyramid network (FPN). The feature pyramid is divided into three scales: large (52x52), medium (26x26), and small (13x13) (where small is the base network, medium is 2 times up-sampling addition, and large is 2 times up-sampling addition). The method of convolution kernel (3x3 and 1x1) is used to interact with the local features of each feature map. Each different scale and each target are compared in order. After using this method, it plays the same role as the fully connected layer. The structure of FPN is as follows Figure 3 .

[0093] (1) Cross-scale prediction:

[0094] In the initial process, the feature map is first extracted, then it is sampled to become twice the size of the previous one, then the initial network is feature extracted, and finally the two are merged. The advantage of this method is that the information obtained is more complete, which is conducive to future work. On this basis, the feature map needs to be processed, and before that, a convolution layer should be added, then the tensor is predicted, and this step is repeated to complete the box scale prediction work.

[0095] The detection network prediction process corresponds to different scales, and there are three scales. In the prediction process, the corresponding feature mapping step is 32, 16, 8, which means that if the size of the input image is 416x416, the corresponding detection scale is 13x13, 26x26 and 52x52 respectively, so in each different scale, three scales need nine anchor points to predict, that is, the total number of anchor points used for prediction is 9, and nine frames are arranged in descending order: 1, 2, 3 correspond to small scale 13x13 feature map, 4, 5, 6 correspond to medium scale 26x26, and 7, 8, 9 correspond to large scale 52x52. Using the k-means clustering method, 9 clusters and 3 scales are selected, and then they are uniformly divided. In our radar data set, the 9 anchor point clusters are: (91x81); (120x125); (254x76); (71x295); (328x66); (170x242); (241x174); (306x219); (219x306).

[0096] In depth, the obtained feature map contains (N x N x (B x (5 + C))) entries, where B represents the number of boundary box predictions that each unit can make, and each boundary box has the opportunity to detect objects. 5+C represents the attributes of any boundary box, i.e. center coordinate position, dimension size, object score and C-class confidence.

[0097] (2) Loss function:

[0098] The idea of this algorithm is: the process of directly detecting the target is too complex and cannot achieve the expected effect, so we change the method and convert the detection into a regression method. This method uses the mean square error loss function, that is, the corresponding weight value of the target at different positions is different. So how should the specific steps be carried out? The first step is to distinguish between the two concepts, that is, what is the positioning error and what is the classification error. The positioning error is defined as using the coordinate position of the boundary box to predict the error, so the corresponding weight parameter ; The second step is to distinguish whether the obtained boundary box contains the target or not. The corresponding parameter value of the existence and non-existence of the target is how much. If it does not exist, the corresponding weight parameter is ; If it exists, the value is 1. In this process, we must follow a premise, that is, the smaller the corresponding boundary box, the more sensitive the coordinate position error, so the prediction of the square root of the boundary is more effective than the previous method. Change the prediction value to , and finally use the mean square error to calculate the numerical value after the above steps;

[0099] The expression of the loss function is:

[0100] (1a)

[0101] In expression (1a), the meaning of the i-th grid of the j-th region of the bounding box and the target is 1 if it is responsible for the target, and 0 if it is not responsible for the target. The first and second terms in the expression mean the predicted loss in the positioning process; the third term represents the confidence prediction loss of the target region; the fourth term represents the confidence prediction loss of the target region; the fifth term is used to determine the relationship between the target and the center of the grid.

[0102] (3) Prediction of the bounding box:

[0103] The predicted is the specific position of the bounding box, that is, the coordinates. The specific method is to predict the anchor box obtained. The coordinates of each bounding box are predicted as , according to the offset of the image , and the width and height of the bounding box The bounding box is predicted as follows:

[0104] (2a)

[0105] wherein, is the coordinate position of the center point, and how to use is shown below Figure 4 .

[0106] (4) Class prediction:

[0107] The bounding box has many classes, so it needs to be predicted, and the multi-label classification method can be well performed. The logical classifier is one of the methods used. The class prediction in the training process can be solved by the binary cross-entropy loss method.

[0108] The constructed detection network is shown in Figure 5 .

[0109] 5. Training the detection network and adjusting the network parameters:

[0110] The training method is to initialize the weight using the classification network, the total number of iterations is 15,000, and the initial learning rate value is set to 0.01. Every 3000 times, the learning rate is multiplied by 0.1, and the cumulative multiplication is three times.

[0111] The weight decay rate is 0.0005, and the momentum is 0.9. The training loss function image is shown in Figure 6 .

[0112] ​Using standard data enhancement methods: rotation, hue, saturation, exposure, batch normalization, and using data jitter to generate more data, setting network weight decay regularization term to prevent overfitting; in the case of CPU is Intel I7, memory is 64G, graphics card is NVIDIA GeForce GTX1080Ti, we set the batch size to 64, we found that the network is currently forward propagation, so we need to make it change direction, using small batch gradient descent method, observe the loss function image, found that 12000 times of iteration of weight suitable for the late detection network test.

[0113] 6, test detection network:

[0114] The obtained model is used for testing the radar image test set, and after multiple comparison tests, we find that the threshold of 0.25 and NMS of 0.45 in the detection model can ensure high accuracy of detection and prevent false detection, and the recall rate is almost not reduced.

[0115] 7, generate target detection and recognition system:

[0116] According to different scene needs, corresponding algorithm application scenarios can be established, the first kind is batch input detection, and the output result; the second kind is single input detection, and the output result, both methods have their own characteristics, the first kind is suitable for large batch data detection, GPU parallel operation, faster; the second kind is suitable for demonstration, when inputting an image, the output result can be visualized, and the next image is waiting for input, the specific detection flow chart is as shown in Figure 7

[0117] Finally, it should be noted that: obviously, the above embodiments are only examples for clearly illustrating the present application, and are not limited to the implementation. For ordinary skilled in the art, on the basis of the above description, other different forms of changes or variations can also be made. Here, it is not necessary and impossible to enumerate all the implementation. The obvious changes or variations derived therefrom are still within the scope of the present application.​

Claims

1. A deep learning-based ISAR target recognition method, characterized in that, The identification method includes the following steps: Electromagnetic simulation of the target model was performed using FEKO, and the simulation results were imaged to construct a target recognition database. The images in the database are preprocessed, including target annotation, coordinate normalization, and data augmentation operations, to obtain an image set; A convolutional neural network classification model is constructed and initialized. After completion, a deep learning training framework is used to train the model on 10 MSTAR target classes. The gradient descent method is used to optimize the network and obtain the network pre-training results. Specifically, the initialization of the convolutional neural network classification model uses a global average pooling layer instead of a fully connected layer for prediction and classification. 1×1 convolutions are used between 3×3 convolutions to compress feature representations. A residual network is constructed using skip connections, and batch normalization (BN) is added after the convolutional layers. After the convolutional neural classification network model is initialized, it is trained using a deep learning training framework. Among them, the pre-training of the convolutional neural network uses Darknet as a framework to train the convolutional neural network on 10 MSTAR target classes. The training parameters are set, standard data augmentation methods are used, mini-batch gradient descent is used to change the network propagation direction, and the weights after iteration are selected as the network pre-training results. Based on the convolutional neural network classification model, a YOLOv3 detection network is constructed. The YOLO detection layer replaces the fully connected layer, and a feature pyramid network structure is adopted. Multiple scales are set, and convolutional kernels are used to interact with local features of the feature map. The detection network model is initialized, and then trained using the YOLOv3 detection network. Gradient descent is used to adjust the network weights suitable for detection. In the loss function step of the detection network, detection is converted into regression, and the mean squared error loss function is used. Different weight parameters are set according to the relationship between the bounding box and the target, the presence of the target in the bounding box, and the size of the bounding box. The square root of the boundary is predicted, and finally, the mean squared error is used for numerical calculation of the loss function. The expression is: ； In the formula, B represents the number of grid cells, and B represents the number of bounding boxes predicted for each grid cell. To determine whether the j-th bounding box in the i-th grid is responsible for predicting the target, To determine whether the j-th bounding box in the i-th grid is not responsible for predicting the target. To determine whether the i-th grid contains the target, , The center coordinates, width, and height of the true bounding box. , These are the center coordinates, width, and height of the predicted bounding box, respectively. , The confidence levels of the actual and predicted targets are respectively. , These represent the true and predicted target class probabilities, respectively. , These are the weight parameters corresponding to the positioning error and the parameter value confidence, respectively. If there is no target in the bounding box, the corresponding weight parameter is... Conversely, the corresponding weight parameters ; In the bounding box prediction step of the detection network, the bounding box coordinates are predicted using anchor points. The prediction is based on image offset, bounding box width, and height, and the prediction formula is as follows: ； In the formula, These are the center coordinates, width, and height of the predicted bounding box, respectively. The center coordinates, width, and height of the anchor point frame are given respectively. , The coordinates of the center point of the anchor point frame. These are the offset coordinates of the image. Define the width and height of the anchor point frame; The trained detection network model is used to detect the test image set. The target detection and recognition results are obtained by setting parameters, and the corresponding algorithm application mode is established according to the application scenario requirements.

2. The ISAR target recognition method based on deep learning as described in claim 1, characterized in that: In the cross-scale prediction step of constructing the detection network: First, the feature map is extracted and upsampled to double its original size, while the initial network is also used for feature extraction. The sampled feature map is merged with the features extracted by the initial network, and then added to a convolutional layer for tensor prediction. Repeat the above steps to complete the box scale prediction; For prediction processing at different scales, corresponding feature mapping strides are set, and the k-means clustering method is used to determine the anchor points corresponding to different scales. In the radar dataset, a total of 9 anchor points are set, and the clusters of the 9 anchor points are as follows: (91×81); (120×125); (254×76); (71×295); (328×66); (170×242); (241×174); (306×219); (219×306).

3. The ISAR target recognition method based on deep learning as described in claim 1, characterized in that: In the category prediction step of the detection network construction and training phase, a multi-label classification method is adopted, using a logistic classifier and a binary cross-entropy loss method for category prediction.

4. The ISAR target recognition method based on deep learning as described in claim 1, characterized in that: Based on the scenario, establish the algorithm application mode, specifically the algorithm application scenario of batch input detection output result or single input detection output result.

5. A deep learning-based ISAR target recognition system, characterized in that, The system includes: The data infrastructure building module is used to build the data infrastructure and obtain a target recognition database through electromagnetic simulation and imaging. The image preprocessing module is used to preprocess images, including labeling, coordinate normalization, and data augmentation of images in the database to obtain an image set. The convolutional neural network classification module is used to construct and train the convolutional neural network classification network. It initializes the network with a specific structure design, trains it on the MSTAR 10-class target based on the Darknet framework, sets parameters and optimizes the network to obtain pre-training results. The initialization of the convolutional neural network classification model uses global average pooling layers instead of fully connected layers for prediction and classification. It uses 1×1 convolutions between 3×3 convolutions to compress feature representations, constructs a residual network using skip connections, and adds BN after the convolutional layers. After the convolutional neural classification network model is initialized, it is trained using a deep learning training framework. Among them, the pre-training of the convolutional neural network uses Darknet as a framework to train the convolutional neural network on 10 MSTAR target classes. The training parameters are set, standard data augmentation methods are used, mini-batch gradient descent is used to change the network propagation direction, and the weights after iteration are selected as the network pre-training results. The detection network processing module is used to construct the YOLOv3 detection network. It replaces the fully connected layers with YOLO detection layers, adopts a feature pyramid network structure, sets multiple scales, and utilizes convolutional kernels to interact with local features in the feature maps. After initializing the detection network model, the YOLOv3 detection network is trained, using gradient descent to adjust the network weights suitable for detection. In the loss function step of the detection network, detection is converted into regression, using the mean squared error loss function. Different weight parameters are set based on the relationship between the bounding box and the target, the presence of the target in the bounding box, and the size of the bounding box. The loss function is calculated numerically using the mean squared error, and the result is a loss function. The expression is: ； In the formula, B represents the number of grid cells, and B represents the number of bounding boxes predicted for each grid cell. To determine whether the j-th bounding box in the i-th grid is responsible for predicting the target, To determine whether the j-th bounding box in the i-th grid is not responsible for predicting the target. To determine whether the i-th grid contains the target, , The center coordinates, width, and height of the true bounding box. , These are the center coordinates, width, and height of the predicted bounding box, respectively. , The confidence levels of the actual and predicted targets are respectively. , These represent the true and predicted target class probabilities, respectively. , These are the weight parameters corresponding to the positioning error and the parameter value confidence, respectively. If there is no target in the bounding box, the corresponding weight parameter is... Conversely, the corresponding weight parameters ; In the bounding box prediction step of the detection network, the bounding box coordinates are predicted using anchor points. The prediction is based on image offset, bounding box width, and height, and the prediction formula is as follows: ； In the formula, These are the center coordinates, width, and height of the predicted bounding box, respectively. The center coordinates, width, and height of the anchor point frame are given respectively. , The coordinates of the center point of the anchor point frame. These are the offset coordinates of the image. Define the width and height of the anchor point frame; The testing and application module is used to detect a set of test images using a trained detection network, set parameters to obtain results, and establish algorithm application modes based on the scenario.

6. The ISAR target recognition system based on deep learning as described in claim 5, characterized in that: The convolutional neural network processing module includes: The network construction submodule is used to replace fully connected layers with global average pooling layers for prediction and classification, compress feature representations with 1×1 convolutions between 3×3 convolutions, construct residual networks using skip connections, add BN after convolutional layers to prevent overfitting, and initialize the network model. The network training submodule is used to train the convolutional neural classification network on 10 MSTAR target classes based on the Darknet framework. It sets the initial learning rate, decay method, weight decay rate, momentum and batch size, uses standard data augmentation methods, adopts mini-batch gradient descent to change the network propagation direction, and selects weights with a specific number of iterations as the network pre-training result.

7. A deep learning-based ISAR target recognition device, characterized in that, The device includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the deep learning-based ISAR target recognition method as described in any one of claims 1-4.

Images