Research On Local Feature Matching Algorithms Based On SAR Images

How to perform robust high resolution SAR image feature matching has always been a difficult and bottleneck of features matching algorithms.The SAR image matching technique of feature pattern has been no robustness and wide universality because of its main reason lies in the inherent speckle noise of high resolution SAR images.The development of Optical image matching technology algorithms is gradual mature.Because of its conplete robustness and university,the Scale Invariant Feature Transform(SIFT)algorithm has been widely used in image matching and target recognition.In the near future,matching technique of the feature pattern of optical image is divided into two directions:1.The construction of the local feature points neighborhood description information(patch based image including gradient,phase,gray,transform domain information,etc);2.Local feature points graph generation and similarity constraints(features-map mainly methods Delaunay triangular network,etc).High resolution SAR image matching technology is divided into three major categories:Area-based pattern,Frequency domain pattern and Feature-based matching algorithm.The Area-based matching algorithm is mature and its effect is high stability and good robustness.However,the Area algorithm is time consuming and SAR images' size are large,which can not be applied to the military field in real time.Feature based on matching algorithm is high efficiency,anti-blocking and so on In this paper,the high resolution SAR image matching based on feature pattern is studied along the SIFT algorithm framework.The main contents are as fellows:A scale space module based on nonlinear filter is proposed to construct the scale space under the framework of SIFT algorithm.The nonlinear diffusion filter uses the fast explicit diffusion algorithm for speed up numerical decomposition Meanwhile,its diffusion coefficient uses the exponentially weighted mean proportional operator as the weight coefficient(Nonlinear Diffusion Filter ROEWA,NDFR).Scale space noise reduction adopts the exponentially weighted proportional operator that can effectively suppress the speckle in high resolution SAR images.The feature detection algorithm uses the SAR-Harris combined with ROEWA algorithm to feature detection,which can effectively extract the robust points.Finally,the MSURF descriptors can resist the SAR matching problem that the probability of low score and high mismatch.Experimental results show that our proposed algorithm is robust and stable.Because of the poor anti-affine performance of the SIFT algorithm,this paper study the fully affine invariant algorithm which is ASIFT.The affine invariant high resolution SAR image matching is resanpled along the ASIFT framework.But the resaimpling mode which dates from camera imaging simulation,is single source on realrtime SAR image and the reference image is not resampled.After resampling the real-time map,we adopts the proposed algorithm to match.The experiments are compared with the Harris-Affine,Hessian-Affine,MSER algorithm,which are currently anti-affine invariant operators.The results show that the total affine invariant algorithm under the framework of ASIFT can effectively solve the large views high resolution SAR image matching and registratio.After the global match is completed,Firstly the real-time map is subjected to the reference map to global coarse matching transformation according to the estimated transformation matrix.Secondly,the sub-block region is equally divided into the same region,and the control points of each sub-blbck regions are constructed with Delaunay triangles.For each sub-btock area using gradient temple matching,and then sub-block center point of the location update.Finally,the control point from the real-time map and reference map used to generate the triangular mesh for nonlinear parameter estimation and then interpolation of multi-triangular meshes.So the local image matching process is completed.