| Synthetic aperture radar as an important tool for earth observation,has experienced more than 60 years of the development and testing.Its all-time and all-weather characteristics make it a very wide range of applications,and the detection of surface aircraft target in SAR image is one of the important applications,the detection of aircraft targets are important in both the military field and civilian areas.With the development of synthetic aperture radar(SAR)technology,the acquisition of SAR image is becoming more and more abundant,and the image resolution is getting higher which makes it possible to use deep learning in the field of SAR image processing.This dissertation analyzes the problem of target detection of SAR image surface aircraft,introduces the related technology of optical image target detection,and puts forward a more effective method of aircraft target detection based on the characteristics of SAR image.The main contents of this dissertation include the following four aspects:Firstly,this dissertation summarizes the optical image and SAR image target detection technology,combing out the basic framework of the typical target detection algorithm.The basic framework includes:1)candidate region extraction;2)feature extraction;3)classifier classification.In the aspects of candidate region extraction and feature extraction,the methods used in the detection of optical image and SAR image are quite different.In this dissertation,these methods are briefly introduced.Besides,the problems existing in the SAR aircraft target detection task are also summarized:1)SAR scattering mechanism caused by the SAR scattering mechanism is separated,which is called sparseness;2)SAR image scattering mechanism and the aircraft target itself particularity caused a variety of scattering,which is called diversity.Aiming at sparsity,an improved DPM algorithm for adaptive component selection is proposed.Considering the diversity,the part based multilayer parallel network target detection algorithm is proposed.Secondly,the part-based target detection algorithm and the network-based target detection algorithm are introduced respectively in the optical image target detection technology.This dissertation mainly introduces the DPM algorithm and YOLO algorithm,explains the key ideas,the procession of training and detection.And the advantages and disadvantages of the two algorithms are analyzed in combination with the SAR image itself in this dissertation.Thirdly,this dissertation introduces the specific ideas of the two algorithms proposed in this dissertation,and introduces their design principles,training and testing process in detail.The improved DPM algorithm for adaptive component selection is to solve the sparseness problem of aircraft target scattering point in SAR image by using the component information.At the same time,this algorithm is improved according to the existing problems of the original DPM algorithm and the characteristics of the SAR image.The part based multilayer parallel network target detection algorithm introduced the deep feature and the multi-layer network structure in the deep learning to solve the problem of the diversity of the aircraft target scattering and the sparseness of the scattering point in the SAR image,and perfected the component network and the whole Network,proposed a hierarchical component-overall detection and identification network framework.In particular,the decision constraint layer further improved the SAR image aircraft target detection accuracy,and reduced the false alarm rate.Finally,by setting the contrast test,the validity of the proposed method in this dissertation is explained.The improved DPM algorithm improves the detection accuracy to a certain extent,but the part based multilayer parallel network target detection algorithm has higher precision and lower false alarm rate. |
