SAR Image Denoising And Target Detection And Recognition Based On Convolutional Neural Network

Synthetic Aperture Radar(SAR)has many irreplaceable characteristics,such as all-time,allweather,active imaging and so on.SAR is widely used in military and homeland security fields because of its characteristics.It's very difficult to distinguish targets in SAR images with eyes,so it's necessary to achieve automatic recognition of SAR images which has a great significance to interpret efficiency and reduce manpower work.The imaging principle of SAR images results in speckle noise,which seriously affects the quality of the images.Thus,it is very important to study effective speckle noise suppression algorithm for improving the interpretation efficiency of SAR images.SAR ATR(Automatic Target Recognition)is also very necessary in SAR image processing and analysis.The specific works of this thesis are as follows:(1)A SAR image classification model is proposed based on the fusion of different characteristics which have different receptive fields.The feature vectors of different receptive fields in Alex Net are concatenated on channels,and then,the MSTAR dataset is used to evaluate this new method after data augmentation.Compared with some CNN methods which are used to the classification of MSTAR images in recent years,the proposed algorithm achieves higher accuracy in SAR image classification.Compared with VGG16,the proposed method has a faster training speed.(2)A SAR image despeckling algorithm based on deep convolution neural network is proposed,in order to solve the shortcomings of existing SAR image despeckling algorithms,such as low despeckling efficiency and poor flexibility.Residual learning mechanism is applied to the algorithm,because it can learn the mapping from noisy image to residual image instead of the mapping from noisy image to clean image,which reduce the difficulty of training.The proposed method is trained on artificial synthetic speckle noise images,and validated on both artificial synthesis noise images and real SAR images from Terra SAR-X satelite.Whether on synthesis noise images or on real SAR images,the proposed despeckling algorithm have better results and lower time loss.It can not only suppress speckle noise in homogeneous region,but also preserve edge information of images.(3)SAR image despeckling algorithms and Faster RCNN are combined,in order to achieve automatic and accurate detection and recognition of multiple targets in large scene SAR images.It's very difficult to acquire large-scale SAR images which contains targets.The large-scale scene images and three kinds of military targets from MSTAR dataset are synthesized for SAR target detection and recognition.Frost,Lee,curvelet,PPB algorithm and the algorithm proposed in this article are used to despeckle the images.The Ro I Align is used to replace the Ro I Pooling in Faster RCNN to improving the recognition performance for small targets,and then,the improved Faster RCNN is trained by using different noise levels datasets,so as to achieve high precision detection and recognition of military targets in SAR images.