Research On High Resolution Range Profile Recognition Methods For Ships

In the field of radar automatic target recognition(RATR),target recognition based on high resolution range profile(HRRP)is an arduous and meaningful task.Estimating the physical length of ship target based on HRRP can provide important information for the recognition of ship target;Segmenting the HRRP data set according to the attitude angle can improve the generalization performance of classifier;Extracting HRRP features can not only expand the amount of data,but also overcome the various sensitivity problems of HRRP;For the recognition of HRRP,construct a suitable classifier can improve the final recognition rate.Based on the estimation of ship target length,the segmentation of HRRP data set,the feature extraction and the construction of classifier,the main works of this thesis are as follows:(1)Aiming at the problem of ship target length estimation,a new process based on sequential HRRP and attitude angle is proposed.According to the HRRP of the ship target,an algorithm of extracting the projection length of the ship target is proposed in this process.According to this algorithm,the projection length of the ship target at different attitude angles is extracted,and the physical length of the ship target is estimated based on the attitude angle,which provides some prior information for further recognition.The calculation results show that the algorithm has good stability and accuracy,and the physical length error estimated by the process is also within a reasonable range.(2)To solve the problem of HRRP segmenting,an adaptive segmenting process based on unified manifold approximation and projection(UMAP)is proposed.This process uses UMAP and a clustering effect evaluation mechanism to segment the HRRP data set,which overcomes the difficulty of parameter searching and provides a new idea for HRRP segmenting.The calculation results show the superiority of the segmenting process and improve the recognition rate.(3)To solve the problem of HRRP feature extraction and classification,a convolutional neural network(CNN)recognition process based on reconstructed bispectra is proposed.In this process,a special CNN structure is designed,and the key area of HRRP bispectral image is reconstructed as the input data of CNN.The results show that the recognition process has a good recognition effect on largescale HRRP data sets with a large number of types,and overcomes the problem of sample data imbalance and CNN over fitting.