Synthetic Aperture Radar(SAR)has advantages of working all-time and all-weather and have a certain ability to penetrate the observation target.SAR has become an irreplaceable remote sensing information acquisitio n.At present,SAR has been widely used in national defense and national economy.In recent years,SAR imaging technology has matured,the data obtained by the SAR system is increasing,and the image quality is getting better and better.How to extract the target of interest from the massive data quickly is the urgent problem that needs to be solved.SAR image interpretation is a research hotspot in recent years.SAR target recognition technology is an important pa rt of SAR image interpretation.SAR target recognition technology mainly includes three parts: data preprocessing,feature extraction and classifier design.How to design and extract the discriminant characteristics of the target is one of the main difficulties of SAR target recognition technology.The deep learning techniques that have emerged in recent years have the advantage of automatically extracting hierarchical features from massive data,which has achieved a number of outstanding results in the field of image recognition.In this paper,a new SAR target recognition method based on improved convolutional neural network model is proposed in combination with modern deep learning techniques.The main work of this paper includes the following aspects:Firstly,the working principle of convolution al neural network is analyzed.On this basis,the core structure of convolution neural network is deeply studied.Feature extraction is an important step in SAR target recognition.The traditional feature extraction method is often designed according to th e professional knowledge.This process is very difficult an d time consuming.In this paper,a method of target recognition for SAR images is presented.Secondly,since the number of SAR images is limited,it will yield severe over-fitting when using traditional CNN algorithm.A few strategies such as Re LU activation function,L2 regularization,batch normalization and Dropout are investigated to retrain over-fitting.Four kinds of data augmentation method are proposed to suppress the over-fitting problem.Finally,the validity of the algorithm i s verified by the MSTAR dataset and the OKTAL simulation dataset,and the sensitivity of the algorithm to the target mirror,displacement,rotation and noise is verified by the extended test set.The experimental results show that the algorithm proposed in this paper has achieved high recognition accuracy on both data sets,and the model is not sensitive to target mirror,displacement,rotation and noise. |