Classification and recognition method of low, slow and small targets

Abstract

The present invention discloses a classification and recognition method for low, slow and small targets, which has accurate recognition and short recognition time. The present invention is implemented by the following technical scheme: based on the PPI image formed by the original point track, different target track data are used as training samples of the data preprocessing module; the data preprocessing module predicts and preprocesses the law of track-related point track data, generates a training set, and constructs a deep learning network model and a network optimization module that sequentially uses a deep convolutional network DCNN and a long short-term memory network LSTM, extracts two sets of features from the frame data, and splices them to obtain joint features, inputs the deep learning and image recognition target track data into the weighted calculation module in real time, compresses and accelerates the trained network model structure, and realizes model compression and acceleration through weighted calculation and model pruning, realizes the deep features required for accurate classification, and completes the accurate recognition of the classification and recognition of small flying birds and drones.

Description

Technical Field

[0001] The present invention relates to radar detection and identification technology for targets such as flying birds and unmanned aerial vehicles, and mainly to a classification and identification method for "low, slow and small" targets with a flight altitude below 1,000 meters. In particular, when a low-altitude, low-power search radar detects small targets, a method for classifying and identifying detected small birds and small unmanned aerial vehicles is used. Technical Background

[0002] The target characteristics and flight characteristics of birds are similar to those of small drones. They belong to the "low, small, and slow" category of targets, which will interfere with the identification of small drone targets. Low, slow, and small targets, the full name of which is "low-altitude, slow, and small flying targets", have low observability. Effective monitoring and identification of both have become a difficult problem that urgently needs to be solved to ensure the safety of air routes and urban security. Birds and unmanned aerial vehicles (UAVs) are typical "low, slow, and small" targets with low observability. Birds and drones have many types of targets, low flight altitudes, and their flight altitudes are generally below 1,000 meters. They are slow, highly maneuverable, and have many background interferences. They are low, small, and slow, and the Doppler shift is not obvious. The radar scattering area of ​​aerial targets is small. In addition, the detection environment is complex, which brings great trouble to target detection. With a slow speed and a small radar reflection area, they are "difficult to find, difficult to capture, difficult to deal with, and difficult to deal with". "Low, slow, and small" targets pose a great threat to the air defense security of important targets. The unlicensed and random flight of drones seriously affects and threatens public safety. Since it is difficult to distinguish between light and small drones and flying birds, the detection and identification of "low, small and slow" targets has become a global problem. Modern radars require accurate threat level assessment and elimination of threats, and targets need to be identified and classified. The ordinary MTI algorithm is very effective in suppressing ground clutter, but it cannot eliminate motion clutter; the target tracking algorithm can eliminate some false alarm trajectories, but the targets with a large number of unstable trajectories formed by complex environments such as geographic environment information (mainly habitats and foraging grounds), urban vehicles, and airflow in space cannot be completely distinguished by the trajectories formed by the target tracking algorithm. For the detection and identification technology of flying birds and drones in complex scenes in the future, the detection and identification technology of rotorcraft drones and flying birds in complex scenes in recent years has been the main means of detecting flying birds and drones. Commonly used detection methods mainly provide detection and identification methods from aspects such as echo modeling and micro-motion characteristic recognition, maneuvering feature enhancement and extraction in pan-exploration mode, distributed multi-view feature fusion, motion trajectory differences, and deep learning intelligent classification. Because the threat levels and threat directions of birds and small drones to low-altitude airspace are different, the countermeasures against birds and small drones are also different. Therefore, being able to distinguish between small birds and small drone types is crucial to maintaining the safety of low-altitude airspace.

[0003] In order to distinguish between light and small UAV targets and flying bird targets, the low-altitude airspace radar surveillance system uses radar data to extract target motion features to improve the recognition ability of the low-altitude airspace radar surveillance system. The traditional recognition method is to use signal processing, through data analysis, artificially design formulas to extract several features such as speed, acceleration, curvature, RCS variance, RCS range, etc., and then use classification strategies such as support vector machine, K nearest neighbor, Markov chain, decision tree, etc. to achieve target recognition. This type of method is limited by the field of human knowledge. The shallow features of the target are extracted from the input data. The feature types are limited and the representation ability is insufficient, and the accuracy and generalization ability of classification recognition are limited. At the same time, for the two types of targets, small flying birds and small UAVs, from the perspective of conventional motion characteristics and RCS changes, there is no significant difference between these two types of targets, and traditional recognition methods are difficult to accurately identify. At present, target recognition technology based on deep learning can be separated from traditional feature engineering to achieve automatic perception and recognition of features. However, relying on manual and shallow feature learning structures is not only complex but also leads to a decrease in recognition accuracy. There is still an overfitting problem when using convolutional neural networks to process limited label data. Because the categorical dependent variable in the classification problem may be seriously biased, that is, the ratio between categories is seriously unbalanced. If the data is seriously unbalanced, the conclusion drawn from the prediction is often biased, that is, the classification result will be biased towards the class with more observations. To deal with this problem, either cut off a part of the class with more observations (i.e. undersampling), or perform Bootstrap sampling on the class with less observations (i.e. oversampling). However, there are problems with doing so. For the first method, the cut data will lead to the loss of some implicit information; and in the second method, the simple replication formed by sampling with replacement will cause the model to overfit. Summary of the invention

[0004] In order to overcome the problem of difficulty in classifying low, slow and small targets, the present invention provides a classification and recognition method for low, slow and small targets, which has low implementation cost, accurate recognition, high efficiency, high real-time performance and can shorten the recognition time.

[0005] The above-mentioned purpose of the present invention can be achieved by the following measures: a classification and recognition method for low, slow and small targets, characterized in that it includes the following steps: based on surrounding geographical environment information, historical track data, collection of relevant point track data, and pixel density PPI (Pixels Per Inch) images formed by original point tracks, different target track data are used as training samples for a data preprocessing module; the data preprocessing module first preprocesses the original data, predicts unknown data using the target track data law, automatically analyzes the data to obtain the track-related point track data law, performs data prediction and preprocessing, generates a training set, and constructs a deep learning network model and a network optimization module. The deep learning network model uses deep learning algorithms and image recognition related algorithms to classify and recognize two types of PPI images formed by low, slow and small targets, and performs frame processing, data feature extraction and classification recognition on two groups of feature data time series of flying birds and drones, and successively uses deep convolutional network DCNN and long short-term memory network LSTM to extract the above two groups of features from the frame data, and splices the two groups of features to obtain joint features; the network optimization module constructs a network self-learning based on the DCNN_LSTM architecture recognition network Learning mechanism, incremental learning training of the joint features of the deep learning network model DCNN_LSTM recognition network, recognition of new targets that have not been learned, training the deep learning network model with the training set, realizing the fast incremental learning of the deep learning network model for new data, and inputting the deep learning and image recognition target track data into the weighted calculation module in real time, compressing and accelerating the trained network model structure, and calculating the channel entropy value of the DCNN part of the DCNN_LSTM recognition network based on the model pruning method of the channel entropy value. Model compression and acceleration are realized through weighted calculation and model pruning, and automatic learning is carried out from the original flight data of the target and the flight trajectory graphics to realize the deep features required for accurate classification, so as to perceive the subtle differences in the target movement, and weighting is performed on the basis of the two recognitions to obtain the final recognition result, so as to complete the accurate recognition of the classification and recognition of small flying birds and small drones.

[0006] Compared with the prior art, the present invention has the following beneficial effects.

[0007] 1) The present invention aims at feature extraction, which is the key to all recognition tasks. Based on the surrounding geographical environment information, historical track data, collected relevant point track data, and PPI images formed by the original point tracks, different target track data are used as training samples for the data preprocessing module; the data preprocessing module first preprocesses the original data, predicts the unknown data using the target track data law, automatically analyzes the data to obtain the track-related point track data law, performs data prediction and data preprocessing, generates a training set, and constructs a deep learning network model and a network optimization module; the deep learning network model is based on the DCNN_LSTM architecture recognition network to perceive the subtle differences in target motion and realize the deep features required for accurate classification, which greatly improves the network classification and recognition capabilities and can save a lot of manpower and material resources.

[0008] 2) The present invention adopts deep learning algorithms and image recognition related algorithms to perform two types of classification and recognition calculations on low, slow and small targets; it automatically learns from the target flight original data and flight trajectory graphics, and classifies and recognizes small birds and small drones in combination with trajectory graphics and data points, thereby achieving accurate recognition of small flying birds and small drones, avoiding the traditional recognition method using signal processing, artificially designing formulas to extract several features such as speed, acceleration, curvature, RCS variance, RCS range, etc., and relying on classification strategies such as support vector machines, K nearest neighbors, Markov chains, and decision trees to achieve target recognition, which is limited by the limitations of artificial knowledge fields, has limited feature types, insufficient representation capabilities, and limited accuracy and generalization capabilities of classification and recognition.

[0009] 3) The present invention comprehensively identifies the short-term dependent features and long-term dependent features of network target motion, adopts deep learning algorithms and image recognition related algorithms, performs frame processing, data feature extraction and classification and recognition on the two sets of feature data time series of flying birds and drones, and successively adopts DCNN and LSTM networks to extract the above two sets of features from the frame data, and splices the two sets of features to obtain joint features, which are used for subsequent classification and recognition, thereby greatly improving the efficiency of network classification and recognition.

[0010] 4) The generalization ability of the model for new samples is an important indicator for determining whether the system can be practical. In order to achieve model compression and acceleration, a network self-learning mechanism is constructed based on the DCNN_LSTM architecture recognition network, and the joint features of the deep learning network model DCNN_LSTM recognition network are incrementally learned and trained to identify new targets that have not been learned. The deep learning network model is trained with a training set, and the learned features have stronger discrimination. Compared with independent time series signals or statistical features, the deep learning network model is used to achieve rapid incremental learning of new data, and the deep learning and image recognition target track data are input into the weighted calculation module in real time, and the trained network model structure is compressed and accelerated, and the model pruning method based on the channel entropy value is used. By calculating the channel entropy value of the DCNN part of the DCNN_LSTM recognition network, deleting the channels with small entropy values, and fine-tuning the parameters to achieve the purpose of model compression and acceleration, the model recognition output time is shortened, and the recognition time can be shortened.

[0011] 5) The present invention is based on the idea of ​​transfer learning, combined with the method of joint feature data review, based on the incremental learning method of the model of data review + transfer learning, to perform incremental learning model training on the DCNN_LSTM recognition network, and use the sequentiality between data to smooth the predicted label sequence to improve its performance, and use model pruning to achieve model compression acceleration. Select unimportant parameters in the model, reduce network parameters, and reduce the amount of calculation. Eliminating unimportant parameters or channels will not have a significant impact on the effect of the model, and the compressed model can also restore some network performance through the process of retraining. The channel entropy value is used to evaluate the importance of the channel, and the channels in the deep recognition network are pruned. The adaptability of the recognition system to new samples is improved, laying a solid foundation for the extensive engineering application of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a schematic diagram of the processing flow of the classification and recognition of low, slow and small targets of the present invention;

[0013] Figure 2 yes Figure 1 Schematic diagram of data preprocessing flow of data preprocessing module;

[0014] Figure 3 It is a flow chart of the deep learning classification and recognition network based on DCNN_LSTM to construct a deep learning network model in the present invention;

[0015] Figure 4 It is a flow chart of the model pruning method based on channel entropy value of the present invention. DETAILED DESCRIPTION

[0016] See also Figure 1. According to the present invention, based on the surrounding geographical environment information, historical track data, collected relevant point track data, and the PPI image formed by the original point track, different target track data are used as training samples for the data preprocessing module; the data preprocessing module first preprocesses the original data, uses the target track data law to predict the unknown data, automatically analyzes the data to obtain the track-related point track data law, performs data prediction and data preprocessing, generates a training set, and constructs a deep learning network model and a network optimization module; the deep learning network model uses deep learning algorithms and image recognition related algorithms to classify and identify the PPI images and data formed by low, slow and small targets, performs frame processing, data feature extraction and classification recognition on the two groups of feature frame data time series of flying birds and drones, and successively uses DCNN and LSTM networks to extract the above two groups of features from the frame data, and splices the two groups of features to obtain joint features; the network optimization module is based on DCNN_ The LSTM architecture recognition network constructs a network self-learning mechanism, performs incremental learning training on the joint features of the deep learning network model DCNN_LSTM recognition network, recognizes new targets that have not been learned, and uses the training set to train the deep learning network model to achieve rapid incremental learning of new data by the deep learning network model. The deep learning and image recognition target track data are input into the weighted calculation module in real time, and the trained network model structure is compressed and accelerated. The channel entropy value of the DCNN part of the DCNN_LSTM recognition network is calculated based on the model pruning method of the channel entropy value. Model compression and acceleration are achieved through weighted calculation and model pruning. Automatic learning is performed from the target flight raw data and flight trajectory graphics to achieve the deep features required for accurate classification, so as to perceive the subtle differences in target movement, and weighting is performed on the basis of the two recognitions to obtain the final recognition result, completing the accurate recognition of the classification and recognition of small flying birds and small drones.

[0017] Data preprocessing

[0018] See also Figure 2 The data preprocessing module pre-trains the original data set in the data preprocessing stage, fuses the obtained data with the original data, pre-processes the obtained data, and performs the following steps on the target track sequence data input in real time: coordinate conversion, abnormal data cleaning, and sample equalization processing. In the coordinate conversion, the radar coordinates of the target track are converted into three-dimensional coordinates (x i ,y i , z i )(x i ,y i ,z i ), compare the echo amplitude value with the calibration signal amplitude for the input target track sequence and echo amplitude value change sequence to obtain the equivalent target RCS value.

[0019] In the abnormal data cleaning, the data preprocessing module performs anomaly detection on the sequence data according to the sampling time, deletes irrelevant data and duplicate data in the original data, and processes missing values ​​and abnormal values; adds a label containing at least one complete target track to the data of the corresponding target, and arranges the historical point vector of the target in chronological order for each label and each track; if the data change of adjacent sampling points exceeds the predetermined threshold, it is confirmed that the current sampling value is abnormal, and the average of the two sampling values ​​before and after the rectangular window with a width of L is taken to replace the current abnormal sampling value, and slides on each training data. The number of target historical points of each track does not exceed the width of the rectangular window, and the track data with a length less than L is padded to make each label reach the same length.

[0020] If the number of samples in each category in the sample set used for training is unevenly distributed, the features contained in the category with a small sample size cannot be well obtained, thereby over-relying on limited data samples and causing overfitting, and the model classification accuracy deteriorates. Therefore, before model training, it is necessary to balance the samples in the training set. During the model training phase, the data preprocessing module balances the samples in the training set. This embodiment uses the SMOTE (Synthetic Minority Oversampling Technique) oversampling algorithm to alleviate data imbalance. During the model training phase, the data preprocessing module balances the samples in the training set, uses the SMOTE oversampling algorithm to analyze and simulate minority class samples, adds artificially simulated new samples to the data set, randomly selects N samples from the K nearest neighbors for random linear interpolation to create new minority class samples, and synthesizes the new samples with the original data to generate a new training set; the sample balance is achieved by increasing the number of minority class samples in the classification, and unified in multiple channels, and the joint feature vectors and category labels of all samples in the previous network training are saved as historical data in advance, and the data cleaned from abnormal data is stored in the database. Then, when parameter fine-tuning is required, the historical feature data is exported and used together with the feature data of the new samples for parameter fine-tuning of the subsequent layers of the network.

[0021] Sample balance is achieved by increasing the number of minority class samples in the classification, unifying them in multiple channels, preserving the joint feature vectors and category labels of all samples in the previous network training as historical data, and storing the data cleaned from abnormal data. Then, when parameter fine-tuning is required, the historical feature data is exported and used together with the feature data of the new samples for parameter fine-tuning of the next few layers of the network.

[0022] The input data includes the target track sequence and echo amplitude value change sequence, wherein the track information includes the radial distance R between the target and the radar at each sampling moment. i、Azimuth angle θ i , pitch angle φ i and radial velocity V i , as well as radar cross-section azimuth and differential amplitude, pitch and differential amplitude.

[0023] Building a Deep Learning Network

[0024] See also Figure 3 In the construction of the deep learning network model, for the sequence data input by the data preprocessing module, the deep network is first used to perform sliding window data framing and short-term dependency feature extraction, and a deep convolutional neural network DCNN short-term feature extraction network and a long short-term memory network LSTM long-term feature extraction network are constructed. The fused data and the corresponding labels are sent to the supervised deep convolutional neural network (DCNN), and then LSTM and recurrent neural network RNN ​​are used to obtain the long-term dependency feature privileges of the target track feature extraction results. Subsequently, the fused features extracted by DCNN are spliced ​​with the short-term and long-term dependency features of LSTM. Finally, two fully connected layers and a softmax layer are used to complete the target recognition and classification tasks, and the classification task results are output.

[0025] The data preprocessed by the data preprocessing module is the spatial coordinates of the target (x i ,y i , z i ), radial velocity V i , RCS and other information, that is, multi-channel time series. Since the separability of the target is reflected in the motion characteristics of the target, the sampling value of the target at a single moment has no classification significance. The data of multiple consecutive moments need to be processed as a whole to obtain the target motion characteristics. Therefore, sliding window processing is performed on the input sequence data. The data preprocessing module sets the sliding window size to m, and the sliding step size S=m, which means that the data of consecutive sampling points is taken as a frame of data at a time, and this frame of data is input into the subsequent classification network to obtain a classification result. The sliding step size of the sliding window can be set to S=m, then there is no overlap between adjacent frames. If the sliding step size is less than m, there is overlap between adjacent frames.

[0026] The data preprocessing module constructs a deep convolutional neural network DCNN for deep feature extraction for a frame of data intercepted by the sliding window. For each frame of data, which is a multi-channel time series, the convolution and pooling kernels are constructed according to the time dimension, and the maximum value of the pooling window is taken as the output of the current window. The output data of the convolution layer is subjected to maximum pooling. The training set data is input into the convolutional neural network, and after the input layer, convolution layer, pooling layer, fully connected layer and regression layer, the target category representing the track dimension is obtained.

[0027] For each frame of input data, the deep convolutional neural network DCNN can extract a set of deep feature vectors, which reflect the short-term features of the target in m time steps. However, the long-term dependency features of the target during flight cannot be obtained from a single frame of data. If the features of multiple time frames are integrated and the long-term dependency features of the target are extracted, a better recognition effect will be obtained. Therefore, this embodiment proposes a network structure based on DCNN_LSTM to extract features. After the short-term features are extracted by the deep convolutional neural network DCNN, the azimuth difference ratio and pitch difference ratio information corresponding to each point trace vector are globally convolved to obtain the abstract features of the sum difference ratio information part, and then the recurrent neural network RNN ​​structure of LSTM is used to capture the long-term dependency of the features extracted by the convolution operation of the deep convolutional neural network DCNN, and the extraction of the target time dependency features in the sequence data is completed.

[0028] Recurrent Neural Networks (RNN) is a structure that repeats over time. This structure can achieve a certain "memory function", so it can extract the target time-dependent features in the sequence data. Unlike ordinary neural networks, the neural network unit A of the RNN network is not only connected to the input and output, but also has a loop with itself. That is, the network state information at the previous moment will act on the network state at the next moment.

[0029] The Long Short-Term Memory network (LSTM) is a variant of the RNN. It inherits most of the characteristics of the RNN model, and effectively solves the gradient vanishing problem caused by gradual reduction during the gradient backpropagation process, and can learn long-term time dependencies.

[0030] First, the network introduces a new internal state, which is dedicated to linear cyclic information transmission. At the same time, the input layer outputs information to the hidden layer, and flows through the hidden layer into the external state of the output layer, flowing forward. Secondly, the long short-term memory network LSTM network introduces three gate mechanisms: input gate, forget gate, and output gate to control the path of information transmission. The input gate, forget gate, and output gate all contain a sigmoid neural network layer and a pointwise multiplication operation. The sigmoid layer outputs a value between 0 and 1, describing how much of each part can pass. Among them, the forget gate controls how much information the internal state of the previous moment needs to forget, the input gate controls how much information the candidate state of the current moment needs to save, and the output gate controls how much information the internal state of the current moment needs to output to the external state.

[0031] The LSTM network first uses the external state at the previous moment and the input at this moment to calculate three gates and candidate states. Secondly, it combines the forget gate and the input gate to update the memory unit. Finally, it combines the output gate to pass the information of the internal state to the external state. The LSTM recurrent unit can be used to establish the long-distance temporal dependency relationship of the entire network. After the short-term and long-term dependency features of the target are extracted by the above-mentioned DCNN and LSTM networks respectively, a splicing layer is used to merge the two features, and then the recognition output is completed through the subsequent network layer.

[0032] Network model optimization

[0033] See also Figure 4 The network model of the above complexity takes about 50ms to complete a test operation. To meet the real-time requirements, the network structure needs to be optimized to achieve compression and acceleration of the network model.

[0034] Model pruning is an important means to achieve model compression and acceleration. Model pruning is to select unimportant parameters in the model to reduce network parameters and reduce the amount of calculation. Removing unimportant parameters or channels will not have a significant impact on the effect of the model. The compressed model can also restore some network performance through the retraining process.

[0035] The key to model pruning is to find an effective means of evaluating the importance of parameters. In this embodiment, channel entropy is used to evaluate the importance of channels and prune the channels in the above-mentioned deep recognition network.

[0036] The basic idea of ​​the entropy pruning method is to represent the output of each layer as a tensor of c×h×w, where c represents the number of channels, and h and w represent the length and width of the feature map respectively. The weighted calculation module is based on the model pruning method of channel entropy. The output tensor is converted into a 1×c vector through global average pooling. Each channel generates a corresponding score for a frame of input sequence data. When n frames of data are input, a matrix M∈R is finally obtained. n×c ; For channel j, the obtained M :,j The matrix is ​​divided into m bins, and the probability p of each bin is counted i , through the formula Calculate its entropy value and use it to determine the importance of the channel. By setting a threshold, all channels with entropy lower than the threshold are eliminated, and their corresponding feature maps are also eliminated. Then, fine-tune to restore their generalization ability.

[0037] 4) Network incremental learning

[0038] In practical applications, it is impossible to obtain all the flight information of different birds and different types of drones in different environments in advance, and it needs to be obtained gradually over time. Therefore, the training samples initially used to train the deep network are incomplete. When new target samples with large differences appear, the generalization ability of the model decreases. In order to realize the recognition of new targets by the model, it is necessary to build a network self-learning mechanism to realize the rapid incremental learning of the network model on new data and improve the generalization ability of the model.

[0039] The purpose of incremental learning is to simulate the learning ability of the human brain, so that the network can continuously learn new knowledge while retaining the previously learned knowledge. In the deep recognition network of this method, the DCNN and LSTM layers respectively complete the task of extracting the short-term dependent and long-term dependent motion features of the target. We believe that when processing new samples, the acquisition method of these two types of features can remain unchanged, but the distribution of features extracted from new samples will change. Therefore, only the parameters of the last few layers of the classification network layer need to be updated, and the recognition result output can be obtained using a new feature weighting method. To this end, this method proposes a transfer learning method combined with data review to achieve model incremental learning.

[0040] The conventional transfer learning method is to keep the network parameters of the DCNN+LSTM layer unchanged when new samples appear, and use the new data to fine-tune the network parameters of the subsequent classification layer so that the fine-tuned network can recognize the new data. However, fine-tuning only the last few layers with new data may result in the new network being able to recognize new samples well, but unable to recognize the original target samples.

[0041] In order to enable the fine-tuned network to recognize new data while maintaining the recognition of previously available data, and to achieve incremental learning of the network, the data review method is used when fine-tuning the last few classification layers, namely:

[0042] First, the joint feature vectors and category labels of all samples in the previous network training are saved as historical data in advance; then, when parameter fine-tuning is required, the historical feature data is exported and used together with the feature data of the new samples to fine-tune the parameters of the next few layers of the network.

[0043] Data Validation

[0044] First, the target radar coordinates in the track data and each track data, which contain the radial velocity value of the target, are converted into rectangular coordinates. The 4-dimensional time series is framed by sliding windows, and the overlap of adjacent frame sampling points is 50%. Each frame of data is used as a sample for training and testing the deep learning network, and a 10-layer deep recurrent neural network CNN is designed to achieve target classification and recognition. In the recurrent neural network CNN structure, each of the first four parts contains a convolution layer and a pooling layer. The convolution layer also contains a normalization layer. The convolution layer convolves the input or the output of the previous layer with a group of convolution kernels. The convolution output must pass through the rectified linear unit (ReLU) layer, and the activation function relu(v) = max(v; 0) maps the convolution output to a new output. The maximum pooling layer can find the maximum value of the feature map in the local time domain. The normalization layer is used for the values ​​of different feature maps in the previous layer, using the normalization formula: Normalize it,

[0045] Among them, κ, α, β are hyperparameters, and G(j) is the set of element indices participating in normalization in the feature graph.

[0046] In order to unify the feature maps output by part 4 of all motion features, the above deep learning network is trained and recognized using the framed 4D sequence data. A special fully connected layer is used to process the output of part 4. The calculation method of the jth value in the feature map of this layer uses the normalized exponential function softmax in probability theory and related fields: "Compress" a K-dimensional vector containing any real number into another K-dimensional real vector. In the sixth part of the recurrent neural network CNN, the basic fully connected layer is used to get the final output. The output is calculated based on the output category C, using the normalized exponential function softmax as the last layer of the neural network:

[0047]

[0048] Accept the input value from the previous layer of the network, then convert it into probability, map the output of multiple neurons to the interval (0,1), and the output value is the posterior probability of the classification result. Then construct an objective function, find the cross entropy cost function that can make the objective function reach the maximum / minimum value, and seek a smooth maximum function to effectively limit the total amplitude of the gradient. The gradients obtained from each non-target score are distributed through the softmax function, and gradually transformed into the softmax cross entropy loss function to control the amplitude of the entire neural network.

[0049] The learning rate is optimized using the Adam algorithm during model training, batch size = 32, and the number of iterations epoch = 200. The number of samples used for model training and recognition testing is shown in Table 1.

[0050] Table 1 Number of sample sets

[0051] Training set Validation set Test Set Drones 529 133 106 birds 423 105 106

[0052] Through verification, the correct recognition rate of the recognition method of the present invention tends to be stable after 25 iterations, and the correct classification recognition rate of low, slow and small target test samples such as drones and birds is about 75%.

[0053] The embodiments of the present invention have been described in detail above. After understanding the basic creative concept, those skilled in the art can make various changes and modifications to the embodiments without departing from the design spirit and scope of the present invention. The claims of the present invention include these modifications and variations.

Claims

1. A classification and recognition method for low, slow and small targets, characterized in that: The steps include: Obtain surrounding geographical environment information, historical track data, related point track data and pixel density PPI image formed by original point tracks. The data preprocessing module first preprocesses the original data and generates a training set after the preprocessing is completed; Construct a deep learning network model; the deep learning network model uses deep learning algorithms and image recognition related algorithms to perform classification and recognition calculations on two types of PPI images formed by low, slow and small targets; the classification and recognition calculations are specifically as follows: the deep learning network model first uses a deep network to perform sliding window data frame division for the sequence data input by the data preprocessing module, and then sends the fused data and the corresponding labels to the supervised deep convolutional neural network DCNN to extract short-term dependent features, and then performs global convolution on the azimuth difference ratio and pitch difference ratio information corresponding to each point trace vector, respectively, to obtain the abstract features of the sum difference ratio information part, and then uses the LSTM recurrent neural network RNN ​​structure to capture the long-term dependency on the features extracted by the deep convolutional neural network DCNN convolution operation; Then, the fusion features extracted by DCNN are concatenated with the short-term and long-term dependency features of LSTM. Finally, two fully connected layers and a softmax layer are used to complete the target recognition and classification tasks, and the classification task results are output. Network model optimization: Based on the DCNN_LSTM architecture recognition network, the network self-learning mechanism is used to build a network optimization module. The network optimization module uses the training set to perform incremental learning training on the joint features of the deep learning network model DCNN_LSTM recognition network to recognize new targets that have not been learned. The target track data of deep learning image recognition is input into the weighted calculation module in real time, the trained deep learning network model structure is compressed and accelerated, and the channel entropy value of the DCNN part of the DCNN_LSTM recognition network is calculated based on the model pruning method of channel entropy value. The deep features required for accurate classification are automatically learned from the target flight raw data and flight trajectory graphics to perceive the subtle differences in target movement. The two recognitions are weighted to obtain the final recognition result, completing the accurate classification and recognition of small birds and small drones.

2. The classification and recognition method for low, slow and small targets as claimed in claim 1, characterized in that: The data preprocessing module pre-trains the original data set in the data preprocessing stage, fuses the obtained data with the original data, pre-processes the obtained data, and performs the following steps on the target track sequence data input in real time: coordinate conversion, abnormal data cleaning, and sample equalization processing. In the coordinate conversion, the radar coordinates of the target track are converted into three-dimensional coordinates (x i ,y i , z i ), compare the echo amplitude value with the calibration signal amplitude for the input target track sequence and echo amplitude value change sequence to obtain the equivalent target RCS value.

3. The classification and recognition method for low, slow and small targets as claimed in claim 2, characterized in that: In the abnormal data cleaning, the data preprocessing module performs anomaly detection on the sequence data according to the sampling time, deletes irrelevant data and duplicate data in the original data, and processes missing values ​​and abnormal values; adds a label containing at least one complete target track to the data of the corresponding target, and arranges the historical point vector of the target in chronological order for each label and each track; if the data change of adjacent sampling points exceeds the predetermined threshold, it is confirmed that the current sampling value is abnormal, and the average of the two sampling values ​​before and after the rectangular window with a width of L is taken to replace the current abnormal sampling value, and slides on each training data. The number of target historical points of each track does not exceed the width of the rectangular window, and the track data with a length less than L is padded to make each label reach the same length.

4. The classification and recognition method for low, slow and small targets as claimed in claim 1, characterized in that: During the model training phase, the data preprocessing module balances the samples in the training set, uses the SMOTE oversampling algorithm to analyze and simulate minority class samples, and adds new artificially simulated samples to the data set. It achieves sample balance by increasing the number of minority class samples in the classification, unifies them in multiple channels, and saves the joint feature vectors and category labels of all samples in the previous network training as historical data in advance, and stores the data cleaned from abnormal data. Then, when parameter fine-tuning is required, the historical feature data is exported and used together with the feature data of new samples for parameter fine-tuning of the subsequent layers of the network.

5. The classification and recognition method for low, slow and small targets as claimed in claim 2, characterized in that: The data preprocessed by the data preprocessing module is the spatial coordinates of the target (x i ,y i , z i ), radial velocity V i , the time series of RCS information, that is, the multi-channel time series, performs sliding window processing on the input sequence data, and constructs a deep convolutional neural network DCNN for deep feature extraction for a frame of data intercepted by the sliding window. For each frame of data, which is a multi-channel time series, convolution and pooling kernels are constructed according to the time dimension, and the maximum value of the pooling window is taken as the output of the current window. The output data of the convolution layer is subjected to maximum pooling processing, and the training set data is input into the convolutional neural network. After the input layer, convolution layer, pooling layer, fully connected layer and regression layer, the target category representing the track dimension is obtained.

6. The classification and recognition method for low, slow and small targets as claimed in claim 1, characterized in that: The Long Short-Term Memory Network (LSTM) network introduces three gate mechanisms: input gate, forget gate, and output gate to control the path of information transmission. First, the external state at the previous moment and the input at this moment are used to calculate the three gates and candidate states. Secondly, the forget gate and input gate are combined to update the memory unit. The output gate is combined to pass the information of the internal state to the external state. The LSTM recurrent unit is used to establish the long-distance temporal dependency relationship of the entire network. After the short-term and long-term dependency features of the target are extracted by the above-mentioned DCNN and LSTM networks respectively, a splicing layer is used to merge the two features, and then the recognition output is completed through the subsequent network layer.

7. The classification and recognition method for low, slow and small targets as claimed in claim 1, characterized in that: The weighted calculation module uses a model pruning method based on channel entropy values. It converts the output tensor into a vector of 1× the number of channels c through global average pooling. Each channel generates a corresponding score for a frame of input sequence data. When n frames of data are input, a matrix M∈R is obtained. n×c ; For channel j, the obtained The matrix is ​​divided into m bins, and the probability pi of each bin is counted, using the formula Calculate its entropy value and use it to determine the importance of the channel. By setting a threshold, all channels with entropy lower than the threshold are eliminated, and their corresponding feature maps are also eliminated. Then, fine-tune to restore their generalization ability.

8. The classification and recognition method for low, slow and small targets as claimed in claim 1, characterized in that: First, the target radar coordinates in the track data and each track data, which contain the radial velocity value of the target, are converted into rectangular coordinates. The 4-dimensional time series is framed by sliding windows, and each frame of data is used as a sample for training and testing the deep learning network. A 10-layer deep recurrent neural network CNN is designed to achieve target classification and recognition. In the recurrent neural network CNN structure, each of the first four parts contains a convolution layer and a pooling layer. The convolution layer also contains a normalization layer, in which the convolution layer convolves the input or the output of the previous layer with a group of convolution kernels. The convolution output passes through the rectified linear unit (ReLU) layer, and the activation function relu(v)=max(v;0) maps the convolution output to a new output. The maximum pooling layer finds the maximum value of the feature map in the local time domain. The normalization layer is used for the values ​​of different feature maps in the previous layer, using the normalization formula: Normalization is performed, where κ, α, β are hyperparameters, and G(j) is the set of element indices participating in normalization in the feature map.

Images