Research On Mapping Relationship Between RCS Data And Target Attitude

In recent years,with the continuous development of radar monitoring and deep learning technology,RCS data analysis based on deep learning has become a new research direction,and has been widely used in many fields.Because RCS data acquisition is difficult,samples are rare,and the data dimension is small,it is not enough to provide sufficient samples,so the gradient of neural network in the process of classification training according to RCS attitude is unstable,resulting in low accuracy of training results.In view of the above problems,this paper improves the existing three network structures,and finally finds that the meta learning method can make the model realize the pose classification of the target object in the case of a small number of RCS data samples,so as to obtain the one-way mapping relationship between the RCS data and the target pose.The main research work of this paper is as follows(1)This paper studies and improves three kinds of network model structures:RESNET 18 layers,LSTM and MAML,respectively from the perspective of CNN,RNN and meta learning to solve the problem that the neural network can not classify the pose of the target object under the condition of a small number of samples.On the basis of the original network structure,a new convolution pooling layer is added,and a more appropriate loss function is selected to realize the recognition of the target attitude.(2)In the case of a small number of sample data sets,the original model and the improved model are tested respectively.Since the training of convolutional neural network needs a lot of data as support,this paper makes the following improvements to the basic neural network model:a)adding the attention mechanism of computer vision into the original network;b)adding hourglass convolution layer at the input end to enhance the extraction of data features;c)adding the attention mechanism of computer vision into the original network;c)Convolution pooling layer is added at the output end to calculate the eigenvector of the output feature;d)cross entropy loss function is used to calculate the loss,which improves the accuracy;E)experiments are conducted to evaluate the effectiveness of the three network improvements,and horizontal comparison is conducted.To sum up,this paper uses different mathematical modeling methods to improve the following three network structures:RESNET,LSTM and MAML,and studies the relationship between RCS data and target attitude.Among them,the recognition accuracy of meta learning MAML based on task learning is better than that of recurrent neural network LSTM based on time series data,and the performance of convolutional neural network RESNET based on finding data features is worse than that of LSTM and MAML.The improved network model proposed in this paper can realize the object pose recognition with less samples and solve the problem of target pose recognition At the same time,the average recognition accuracy of the improved network structure is improved from 75%to 80%.