Research On Target Interpretation In SAR Images Based On Shape Prior

Synthetic aperture radar(SAR)is able to explore Earth Observation independently of the lighting and weather conditions.Due to this unique ability,SAR becomes a kind of continuous and reliable sensors.The development of SAR leads to the explosive growth of SAR data.By contrast,however,the progress of SAR image interpretation is relatively slow.Target interpretation is an important application of SAR,where target extraction and recognition is the key means to obtain valuable information from SAR images.The current techniques of automatic target interpretation still have a gap from practical application.Although high resolution SAR images present detailed information of targets,there are some problems frustrating target interpretation,such as speckle noise,intensity variations in target and background regions,the indistinct and incomplete target contour and so on.Therefore,this dissertation implements the research on target interpretation in SAR images based on shape prior.The aim is to improve the performance of target extraction and recognition,by introducing the shape information of targets.The main contributions of the dissertation are shown as follow.(1)Improved nonlocal active contour model is proposed.Focusing on the multiplicative speckle noise in SAR images and the irregular intensity variations,the dissertation introduces a kind of ratio distance based on the probability density function to enhance the robustness of the model.Besides,the nonlocal image processing is used to deal with texture variations and improve the target extraction.Finally,the integral histogram is introduced into the pairwise interactions to fasten the speed of convergence.(2)Refined segmentation of ship target in SAR images based on GVF snake with elliptical constraint.Unexpected distortions stemming from speckle noise and side lobe effect are usually appear on ships in SAR images.They make a great influence on the ship appearance and are hardly recovered using image information only.According to elliptical contour of ships,the elliptical constraint is introduced to facilitate the ship target extraction.On one hand,the ROEWA operator is used to extract edges from images,which produce the gradient vector flows.On the other hand,the ellipse fitting is implemented by the least square method based on the contour,and the elliptical constraint term is added to the energy functional.With the guide of the gradient vector flows and the elliptical constraint,the contour evolution can overcome shape distortions and reach a refined segmentation.(3)A refined segmentation strategy for aircraft recognition in SAR images by statistical shape model.It is difficult to generate a proper representation for aircrafts recognition in SAR images,mainly due to the complicated backscattering phenomenology.To solve this problem,the dissertation proposes a segmentation strategy according to the elastic model.The shape training set is established to model the aircraft shapes.Then the shape analysis is implemented by the statistical shape model,where the sparse representation of the input shape is calculated with the sparse constraint.By this means,a limited number of shape instances are selected to reconstruct the input shape.The contour evolution is implemented by GVF snake model.On one hand,gradient vector flows guide the contour toward edges.On the other hand,with the control of statistical shape model,the contour keeps a regular shape.The real contour of the target can be obtained through iterations.Among the selected candidates,the shape with the highest reconstruction weight is the closest to the target.Hence,the type can be verified.(4)Statistical shape model based on manifold learning.According to the work of(3),a manifold learning technique is introduced into shape analysis.The shape instances and the input shape are supposed as the data points on the a manifold in high-dimensional space.The intrinsic structure of the data are reflected by dimension reduction using ISOMAP.On the low-dimensional embedding,the shape instances close to the input shape are selected by KNN.After calculating the reconstruction weights by Lagrange multiplier,the reconstruction of the input shape can be obtained.The statistical shape model can be incorporated into the iterative framework in(3)to achieve target extraction and recognition with higher precision.