SAR Image Registration Based On Improved SIFT And A Novel Evaluation Method For Registration

To obtain abundant information in images,people often use the same sensor to take pictures of the same target in different time,or use different sensors to take pictures of the same target.But the obtained images are always having many differences in position,angle and some other indicators,and even some distortion is there often.These differences and distortion will seriously interfere with the image stitching,image fusion and change detection,and so on.In order to solve the problem,we should use the image registration technology.Besides,in order to know the result of image registration is good or not,we need an objective and comprehensive evaluating indicator to evaluate it.Because Synthetic Aperture Radar is an active microwave imaging method,it cannot be affected by the cover of the vegetation or the undulating planes of the topography.So the high resolution 2D SAR images can be provided day and night in any weather condition,and they have been used widely in various ground observation.Therefore SAR image registration has important values in theory and practice.Meanwhile,it also has important significance in theory and practices to explore the objective and comprehensive evaluating indicators.Based on the studies in SAR image registration algorithm and the evaluating indicators of the registration results,specific work of this paper is as follows:1.The scale feature of the SIFT algorithm is improved.The octave and the level of the image pyramid in SIFT is reduced in the improved algorithm according to the imaging characteristics of SAR images,and those octaves that have strong interference by speckle noise and those octaves that blurred by Gaussian with a high level are ignored,moreover,the factor of two nearby scales in every octave is modified to ensure the continuity of the scale of the image pyramid and the stability of the feature points.The amounts of match points of the improved algorithm is same with SIFT or even more than SIFT in some images,but the run time of the improved method is decreased significantly compared with SIFT,and the error rate of the match point pair of the improved method is reduced largely.2.A new method to remove the error matches is proposed.Center on the judging feature point pair,two neighborhoods are set which are called “near neighborhood” and “far neighborhood”.The number of match point pairs in these two neighborhoods is counted,if the ratio of the match point pair not less than a threshold,then the couple of judging feature point is matched correctly and they are remained or else is matched wrongly and they are removed.After using the new method,all the error match point pair is removed and the remained correct match point pair is more than 96% on our test SAR images.The result of experiment demonstrated that our new method has obvious advantages than those compared algorithms.3.A new comprehensive evaluation indicator of image registration is proposed.The matches point pair is classified by using of the matching error in the new evaluation indicator.Thus,the accumulated error which produced by statistical methods in evaluation indicator are eliminated.Consequently,the inaccurate evaluations on the match point pair are removed.Besides,the information on the distribution of match point pairs in images are added which is an important factor to measure the distribution of match point pairs.If the match point pairs are massed,the local distortions will produce in the process of image registration.Moreover,the different matching error is quantified to reflect the quality of matches with different matching error.At last,it has an obvious correspondence between the new evaluation indicator and the final results of the image registration.The evaluation tests of nature images and SAR images proved that the evaluation indicator fits the above description.