Research On Multi-scale Target Detection In Synthetic Aperture Radar Images

Object detection via Synthetic Aperture Radar(SAR)can provide information sup-port for military reconnaissance and civil detection,it is an important part of SAR data interpretation.Object detection can also be used as a pre-step for other interpretation tasks,the results of object detection will directly affect the performance of Automatic Target Recognition(ATR)system.Therefore,object detection is an important part of the interpretation system.With the continuous updating of the SAR imaging systems,the difference between targets in images with different resolution taken by different radar is increasing.High-resolution imaging systems also enable small-scale targets to appear in images.In addi-tion,the actual size of targets also varies greatly.Therefore,the same targets appear in the image at different scales.The existence of multi-scale targets makes traditional tar-get detection methods unable to accurately estimate target and background area,causing false detections and missed detections.Specifically,stationary objects are sparse in large scene images,and the variety of target sizes makes it difficult to quickly and accurately detect these targets.For movable targets,the multi-scale characteristic makes it difficult to accurately estimate the target and background distribution parameters.It is necessary to proceed from the mission objective and study the common characteristics of multi-scale objects.In addition,Small-scale targets have limited information in the image.How to fully explore the context information in the image remains to be studied.There are many stationary targets and movable targets in the SAR images.Extensive research on multi-scale detection of these targets requires a lot of human and material resources.Considering the current research foundation and conditions,this dissertation will study the multi-scale detection of airports as typical stationary targets and ships as typical movable targets.The main contributions are as follows:1.In order to rapidly detect multi-scale airports,a algorithm based on line segments grouping and saliency analysis is proposed.A line segment grouping method is designed,the method explores airport support regions via breadth first search.Then potential air-port region can be acquired from the airport support regions that meet certain geometrical condition and distance via selective search.Finally,the pixel saliency and line segment saliency are proposed to remove false alarms.This algorithm effectively improves multi-scale airport detection efficiency.2.In order to accurately detect multi-scale airports,a algorithm based on multi-layer abstraction saliency is proposed.A airport support region refinement method based on superpixels is presented to obtain the airport components.Then clustering algorithm is used to get airport adobes.In this dissertation,three saliency cues are used: local contrast saliency,adobe deformation saliency and global uniqueness saliency.The bayesian infer-ence is adopted to acquired pixel-level saliency map.The proposed algorithm can output pixel-level results of multi-scale airport detection.3.Aiming at the problem that the unsupervised detection method cannot accurately divide the target and the background area due to the multi-scale ship target,and the dis-tribution parameter estimation is inaccurate,a multi-scale ship proposal generator is pro-posed.Selective grouping based on superpixel is used to explore multi-scale ship propos-als.All the ship proposals are scored via edges and contours.Specifically,we use edge density,contour completeness and contour tightness to rank the proposals.The proposed generator can detect multi-scale ships with high recall.4.Aiming at the problem that small-scale ship targets have less information available in the end-to-end detection system and even disappear in the deep feature map,feature pyramid network based on scale-transfer layer is proposed.Feature pyramid network can mearges feature maps of the same spatial size from the bottom-up pathway and the top-down pathway.Dense connection is adopted to fully merge feature maps.Scale-transfer layer is used to explore context information contained in channels.