Research And System Implementation Of Multi-object Behavior Recognition Method For SAR Image

Synthetic aperture radar image target recognition and behavior recognition are key technologies for SAR image processing and interpretation.How to improve target recognition and behavior recognition is a research hotspot in SAR image processing.Due to the special imaging mechanism of the SAR image,the target information is heavily polluted by the speckle noise,which makes the fine structure of some contour details in the image be weakened to some extent.In order to overcome the effect of speckle noise on image details,reduce the reliance on artificial design features and classifiers,and further improve the accuracy of target recognition in SAR images,this paper researches on a SAR image target recognition method based on wavelet threshold denoising and convolutional neural network.In interactive behavior recognition,the traditional feature selection method is usually designed manually.In particular,a lot of prior knowledge and experience are needed to hierarchically design the interactive behavior matrix.It is very difficult and time-consuming to extract features from the interactive behavior matrix.At the same time,as the number of interactive objects increases,the complexity of parameter estimation in some traditional method increases.In order to overcome these problems and improve the accuracy of multi-object interactive behavior recognition,this paper researches on the multi-object interactive behavior recognition method of SAR images based on convolutional neural network.The main research contents are as follows:1.SAR image target recognition methods are studied,and a SAR image target recognition method based on wavelet threshold denoising and convolutional neural network is proposed.The input SAR image is decomposed by two-dimensional wavelet and denoised by the Bayesian estimation using threshold for wavelet coefficient.The convolutional neural network is used to automatically learn the multi-layer feature representation from the synthetic aperture radar image and identify the target type with the learned features.This method can reduce the influence of speckle noise on recognition and improve the SAR image target recognition effect.2.Multi-object behavior recognition methods of SAR image are researched,and a multi-object interactive behavior recognition method of SAR image based on convolutional neural network is proposed.The Surendra background updating background difference method is used to detect the location of moving targets from the input SAR image sequence,and the Bayesian estimation with wavelet threshold denoising and convolutional neural network is used to identify the moving target type.Then,the interactive behavior characteristics of the moving target are extracted and the feature matrix is built.At last,the convolutional neural network is used to identify the target interactive behavior type.This method can overcome the problem that manual selection of features in the interactive environment requires a large amount of prior experience.The method can automatically learn the multi-layer feature representation vector and obtain better behavior recognition performance.3.SAR image target recognition method based on wavelet threshold denoising and convolutional neural network and multi-object interactive behavior recognition method based on convolutional neural network are implemented in SAR multi-object behavior recognition system.The registration and entry module,data reading module,target recognition module and behavior identification module of SAR target recognition system are designed and implemented.The feasibility and effectiveness of the method are verified through the system.