Harbor Extraction From SAR Imagery

With the demand on harbor interpretation with SAR image, the techniques ofextracting harbor from SAR images are studied in this thesis. In order to deal withlarge-scene images in practical applications, this paper proposes a hierarchicalprocedure for harbor extraction. Firstly, the sea area is separated from the large-sceneimage (sea-land binary image extraction); secondly, harbor detection is implementedfrom the sea-land binary image (ROI extraction); then, the ship detection is realizedinsidethedetectedharbor;finally,theshipdiscriminationisimplementedforshipROIs.According to this procedure, this paper has a detailed research on the techniques ofsea-land segmentation, harbor extraction, ship detection and discrimination inside theharborarea.Themainworkofthisthesis includesthefollowingaspects.(1) The disposal of the harbor and the characteristics of SAR images are firstlyanalyzed. The general disposal ofthe harbor is concluded and the general harbor modelis established. Then the main scattering mechanisms encountered in harbor areas andtheir characteristics in SAR images are analyzed. The above analysis on the harbor isthefoundationofthe subsequentresearch.(2)Accordingtotherequirementofprovidingthe sea-landbinaryimageforharbordetection, an accurate and efficient method to segment the sea areas from SAR imagesis proposed. Firstly, the objects and the background are easier to distinguish in a 2Dhistogram than that in a 1D histogram. Secondly, the assumption about themain-diagonal probabilities is unreasonable, which is used with the 2D histogram in atraditional 2D OTSU method. We corrected the calculation of the probabilities at themain-diagonal region,andthesegmentationprecisionis greatlyimproved. Accordingtothe theoretical analysis, a fast recursive method for realizing modified 2D OTSU isobtained, which makes the sea-land segmentation algorithm more practical. Moreover,proposed method can meet the practical application demands of providing sea-landbinaryimageryforharbordetection.(3) Since the existing methods of harbor detection from SAR images are notapplicable for images with different types of harbors, this paper proposes a method ofhabor detection based on features, in order to detect harbor and acquire correspondingboundaries accurately. This algorithm not only makes use of the characteristics ofharborjetty, which has longstrip in shape andconcentrative space distribution, but alsoutilizes the characteristics of the closed harbor coastline, which is surrounded by theland. The combination of the characteristics of harbor jetty and harbor coastline canovercome the problems, that the performance of harbor detection is worse when harborjetties has comparatively incompact distribution and the shape of coastline iscomplicated.Theexperimentalresultsshowthatthenewmethodiseffectivewithahigh detection rate, a low false-alarm rate and good localization performance. The detectionresults can meet the demands of providing harbor ROI for ship detection inside harborregion.(4) According to the requirement of providing ship ROIs for ship discrimination,an effective and efficient method of detecting the ships inside the harbor region fromSAR images using a CFAR detector based on the G0distribution is proposed. Firstly,theSAR imageoftheharborcoastwise regionisextractedbasedontheharborcoastline.Then, a detailed analysis is presented on the clutter statistical properties of harborcoastwise region in SAR image. Further, the ship detection is completed based on theCFAR detector with the G0 distribution. The proposed method can precisely modelclutter data under different clutter environment statistically. By introducing theautomatic censoring of effective clutter pixels, our method has a constant false alarmrate and good performance of detection. The detection results can meet the applicationdemandsofprovidingshipROIsforshipdiscrimination.(5) Aiming at higher precision for discrimination, a method of ship discriminationbased on feature extraction and selection is proposed. Firstly, a new shape feature isconstructed for ship discrimination. Then, the common features for discrmination arequantitatively analyzed by the redundancy, robustness and separability of features. Amethod of selecting the optimal features for target discrimination is given. Finally, aweighted minimum distance classifier is designed to improve the performance of theexistingclassifiers.Theexperimental results show that thenewmethodiseffectivewithhighclassificationaccuracy andexcellentdicriminationperformance.