Target Detection And Discrimination In SAR Images Based On Deep Networks

Synthetic Aperture Radar(SAR) is a sensor device that is not affected by light or climate.As it can work in all days and under all weather conditions,SAR has a wide range of applications in military,agriculture,scientific research and other fields.The automatic target extraction technology for SAR images can accurately locate the targets,so it has important research significance in the SAR image interpretation field.This paper focuses on the problem of automatic extraction of vehicle targets in complex SAR image scenes.It thoroughly analyzes the deficiencies of existing algorithms from the perspectives of target detection and target discrimination,and then makes modifications with deep learning methods.The main research contents are summarized as follows:1.A 10-layer Convolutional Neural Network(CNN)is designed to solve the problem that traditional discrimination algorithms need to artificially select feature combinations.By using deep learning methods,the feature extraction,combination and classification are realized automatically,which effectively helps to avoid the complex feature engineering.This method directly takes original image slices as input,which greatly reduces the need of data preprocessing.The experiments on the measured data of Mini SAR demonstrate that,compared with traditional discrimination algorithms,the CNN-based discrimination algorithm has stronger classification ability and higher classification accuracy for the potential target slices.2.The optical image target detection algorithm—Single Shot Multi Box Detector(SSD)is studied and applied to SAR image processing to achieve the end-to-end target detection.In view of the difficulty in obtaining SAR images and insufficiency of training data,a more reasonable data augmentation method is found through the comparison experiment on Mini SAR data.Moreover,the network structure of SSD is modified based on the analyses of SAR image detection tasks' characteristics.The experiments demonstrate that the modified network can detect more targets without increasing false alarms and has more excellent detection performance.3.Due to the fact that the target detection and discrimination needs to be addressed in two successive stages,we designed a SAR image target extraction software.By integrating all studied algorithms into the human-computer interaction interface,the entire process is achieved automatically.The software is designed by principles of high flexibility and scalability.It includes four modules: data reading,preprocessing,target extraction and auxiliary functions.Every module can be executed sequentially or independently.All parameters in the software have default values or can be modified through external interfaces,which greatly facilitates the understanding and usage of the software.