With the rapid development of Synthetic Aperture Radar(SAR)imagery technology,the resolution of SAR images is higher and higher,and the amount of data becomes larger and larger.Facing the massive SAR image data,how to detect the targets quickly and accurately from these SAR images has been a hot topic in SAR image interpretation,and has very significant research value.In recent years,due to its powerful ability of feature learning,Convolution Neural Networks(CNNs)have made a series of breakthroughs in the various fields,such as image recognition,target detection and so on.Recently,the Single Shot multibox Detector(SSD)algorithm based on CNN has shown great performance in target detection of the natural images,and has received extensive attention from the researchers.This thesis focuses on the application of SSD in SAR image target detection.The main works of this thesis are as follows:1.Aiming at the vehicle target detection task in SAR image with complex scenes,in chapter 3,the SAR target detection algorithm based on SSD is proposed.Compared with the optical images,the amount of SAR images is more lacking.In order to solve the problem of insufficient training samples when applying SSD to the SAR target detection,we adopt the strategies of data augmentation and transfer learning.For data augmentation,the first method is to deal with the original training images from the Mini SAR target detection dataset to generate new training images via adding noise,filtering and flipping.The second method is to perform the filling and interpolation operations on the target images from the MSTAR dataset,in order to obtain the augmented samples to assist the target detection task.For transfer learning,first,we apply the sub-aperture decomposition method to convert the 1-channel SAR images to the 3-channel SAR images.Then,the parameters of the convolutional layers from the basic network in SSD are initialized via the 3-channel network model well-trained on the large-scale optical images to realize the model parameter transfer.The experiments on the measured Mini SAR data have verified the effectiveness of the proposed method.2.Aiming at the vehicle target detection and recognition task in SAR image with complex background,in chapter 4,the integrated algorithm for SAR target detection and recognition based on two-stream SSD is proposed.On the basis of SSD,we add a branch network to introduce the prior information from the saliency map,and construct a two-stream SSD network model.Then the dual convolution neural networks are used to extract the multi-scale features maps from the input SAR image and its saliency map respectively,and the multi-scale feature fusion is operated on these feature maps.Finally,the target detection and recognition tasks are completed at one time based on these fusion features.At the same time,we also employ the data augmentation and transfer learning as the auxiliary ways to improve the performance of the two-stream SSD model for SAR target detection and recognition.To satisfy the requirements of simultaneously accomplishing target detection and recognition,we utilize the clutter and target images from the MSTAR dataset to synthesize the SAR image data of multi-class vehicle targets with complex background.The effectiveness of the proposed method is demonstrated using the synthetic SAR image data. |