Study On Ship Detection Methods For SAR Images

Ship detection is a necessary means to safeguard the maritime rights and interests of coastal countries,which is of important significance in both civil and military fields.Due to the all-day and all-weather observation capability of synthetic aperture radar（SAR）,it has be-come an important data source for ship detection.With the improvement of SAR system and the progress of signal processing techniques,the scattering characteristics of targets and clutter in different wavebands and polarization modes can be acquired.With enriched di-mension of information,target detection performance can be further improved.In this paper,ship detection methods for single-polarization SAR images and polarimetric SAR（PolSAR）images are studied,aiming to improve the detection rate,reduce the false alarm rate,and improve the efficiency of target detection algorithms.The main content of this dissertation is summarized as follows.1.For ship detection in large scenes,a two-stage superpixel-level constant false alarm rate（CFAR）detection method is proposed to improve detection efficiency,which consists of a global CFAR detector and a local CFAR detector.The global CFAR detection is conducted to prescreen the candidate target superpixels with the weighted information entropy（WIE）feature.The WIE is extracted to describe the statistical characteristic of each superpixel,so as to improve the separability between targets and clutter.In the local CFAR detection step,only the candidate target superpixels are judged.Besides,taking the global CFAR detection results as outlier indicator,the impact of adjacent targets on the clutter parameter estimation accuracy can be reduced.Thus more accurate detection results can be obtained in the multiple-target situation.The improvement of detection efficiency can be attributed to two aspects.First,the superpixel is regarded as the basic unit of target detection;second,a coarse-to-fine target detection scheme is adopted.Therefore,the number of detection units to be processed is significantly reduced,thus improving the detection efficiency.The experimental results based on TerraSAR-X images show that the proposed method has robust target detection performance and high detection efficiency under different scenarios.2.For ship target detection in single-polarization SAR images with island and near-shore sea background,a target enhancement and clutter suppression method based on statistical dissimilarity between superpixels is proposed.With improved target-clutter contrast,target detection performance can be improved.Assuming the intensity of pixels in the superpixel obeys the Gamma distribution,the parametric expression of superpixel dissimilarity is de-duced with the Bhattacharyya distance measurement.With this basis,the global and local superpixel-level contrast can be obtained,respectively.Then the global and local contrast maps are combined to construct target-clutter contrast enhancement image,based on which the Gamma-CFAR detector is conducted to obtain the target detection result.In the ex-periment,three SAR images acquired from different harbors are used.The experimental results show that the proposed method can achieve both low false alarm rate and high target detection rate in both port region and coastal area.3.Aiming at the problem of target detection in high-resolution SAR images,an improved active contour model（ACM）based on regional statistical information is proposed to ob-tain accurate target contour segmentation results.Based on the assumption that the pixel intensity in local region follows the multilook Gamma distribution,the energy functional based on regional statistical information is derived.The global minimization active contour（GMAC）framework is utilized to realize the global energy minimization of the energy func-tional,so as to obtain the accurately segmented target contour.Compared with detection methods based on superpixel segmentation,the proposed method does not need to adopt post-processing,such as boundary connection or region merging,to obtain the final target detection results.Although the proposed algorithm is proposed for ship detection,it is also applicable to ground military target detection.The effectiveness of the proposed algorithm is verified by experimental results based on simulated SAR images and measured ground and sea SAR images.4.For ship detection in quad-polarization SAR images,a new detection method is proposed based on polarimetric G_p⁰mixture model and unsupervised clustering.The core idea of the proposed algorithm is to assume that all target pixels belong to the same class.With unsupervised clustering,SAR data can be classified into several clusters,such as target,clutter,interference,etc.Then the feature of each cluster can be extracted to identify the target cluster.The weighted linear combination of G_p⁰distributions with different parameters is utilized to describe the whole PolSAR data,providing effective discrimination between targets and azimuth ambiguities or clutter.Therefore the false alarms can be significantly reduced in the detection result.Besides,considering the sparse distribution of targets in SAR images,only potential target pixels are selected through prescreening for clustering,which greatly reduces the time consumed by class parameter estimation.In addition,the removal of complex massive background clutter pixels yields a faster convergence speed,thus improving the efficiency of the algorithm.In the experiment,the AIRSAR airborne and RADARSAT-2 satellite-borne quad-polarization data with different wavebands are used to verify the effectiveness and efficiency of the proposed algorithm.5.To reduce the influence of speckle noise on the pixel-level target detection method,a novel ship detection method is proposed for quad-polarization SAR images based on superpixel-level features.Due to the effectiveness of extracted superpixel feature,the false alarms caused by azimuth ambiguities and side lobe ambiguities can be remarkably decreased.Firstly,a fast superpixel segmentation method is proposed for PolSAR images.In order to improve the separability between the target and clutter,the Bartlett distance and Riemann distance are exploited to measure the dissimilarity between two superpixels.Then the two local dissimilarity features are extracted at superpixel level.Considering the difference of scattering characteristics between targets and azimuth ambiguities,the polarimetric coher-ence matrix of each superpixel is decomposed to extract the eigenvalueλ₃to better distin-guish targets from azimuth ambiguities.Finally,based on the constructed 3D feature vector for each superpixel,the detection result is obtained using a classifier.Experimental results based on AIRSAR airborne quad-polarization data show that the proposed method can ef-fectively improve target detection rate and reduce false alarms caused by azimuth ambiguity and background clutter.