SAR Vehicle-Target Classification And Recognition Based On Deep Learning

Synthetic Aperture Radar(SAR)has unique advantages that distinguish it from optical and infrared sensors,and plays a pivotal role in both military and civilian fields.After more than 60 years of research and development,SAR imaging technology has gradually matured,and high-resolution SAR image database under various background conditions has continuously enriched.Compared with the traditional low-resolution SAR images in the early days,high-resolution SAR images contain richer background and detail information,which provides a new opportunity and also poses a challenge to the automatic target recognition technology on SAR images.In recent years,deep learning has performed amazingly on natural image recognition.Its powerful feature extraction capabilities and integrated processing flow have opened up new directions for the SAR target recognition technology.However,due to the fundamental differences in imaging mechanisms,SAR and optical images have very different characteristics,it is necessary to modify the existing deep learning methods to fit the SAR images.Aiming at the problems of low efficiency and low accuracy of traditional SAR vehicle-target classification and recognition methods,this thesis studies the deep learning-based SAR vehicle-target classification and recognition methods,including classification of SAR image slices and multi-target detection and recognition in large scenes.The main accomplishments are summarized as follows: 1.Systematically studied and compared traditional SAR target classification and recognition methods with the deep learning-based methods.First,three key steps in traditional methods are presented,and the relevant algorithm principles are explained in detail.Subsequently,the advantages of the convolutional neural network in the deep learning algorithm are described and analyzed from its basic structure,key functions,and training optimization methods,respectively.2.Proposed an Inception-Residual-Dense Convolutional Neural Network to solve the problems in the shortage of SAR image samples and prone to over-fitting.This network can effectively improve the performance of the target classification model in multiple ways since it combines the advantages of residual networks,multi-scale networks,and densely connected networks.The proposed method is assessed in three metrics including prevention of over-fitting effects,model robustness,and detailed feature extraction capabilities.The experimental results with the measured MSTAR dataset have validated the effectiveness of the method.3.Proposed an improved feature fusion single shot multibox detector(FSSD)for the SAR target detection and recognition of small-sized targets such as vehicles with a complex background in a SAR image.First,the basic principles and characteristics of the single shot multibox detector(SSD)algorithm are discussed.Compared with the classic one-stage and two-stage algorithms,the SSD algorithm takes advantage of multi-scale features,but it performs poorly in the detection and recognition of small-sized targets.For this reason,a feature fusion single shot multibox detector(FSSD)algorithm is introduced.Through an analysis on the characteristics of the SAR image,we find that there are still many problems with the SAR image,such as small data volume and imbalance between positive and negative samples,and propose an improved FSSD algorithm,which modifies the original FSSD from five aspects: dataset augmentation,hard negative mining,transfer learning,anchor size change,and the soft non-maximum suppression(Soft-NMS).The proposed algorithm is evaluated in the model's evaluation index mean average precision(m AP),through multiple comparative experiments based on an artificially-synthesized SAR image dataset.The experimental results show that the proposed algorithm can effectively improve the model in many aspects.Finally,it is shown that,the proposed method can also outperform the classic deep learning algorithms for the multi-target detection and recognition in SAR images,which verifies the effectiveness of the method.