Research On Multi Visual Sensor Image Matching Technology Based On Gross Contour

Multi visual sensor image matching technology is applied widely. Owing to the great differencesbetween multi-sensor images, there are still some technical problems unsolved. This paper studies themulti-sensor image matching technology based on gross contour on the basis of human’s cognitionprocess of image.First, the idea of image segmentation first and edge detection second was used to get grosscontours, according to the problems existing in relative research, the imaging characteristics ofmulti-sensor images and the fact that gross contour reflects common features in multi-sensor images.A feature space was established consisting of gray average, gray variance and entropy to represent thefeatures of different objects in images, then the initial clustering number and centers optimized byACO was taken as initial conditions of FCM clustering algorithm which can describe uncertainties ofimages and the improved clustering algorithm was used in the feature space. Canny operator was usedto get gross contours. The experimental results show that the proposed image segmentation methodcan be applied in multi-sensor images, compared with traditional methods the segmentationinaccuracy average was reduced by1.7%~6.6%, and the gross contours are more complete.Second, the gross contours were tracked to eliminate the wrong edges in edge detection results.C-scale segment was used to obtain curvature information of points in contours, then the locallycurvature maximum points was taken as feature points to simplify contours. The experimental resultsshow that compared with Douglas-Peuker method, information of contours after contoursimplification using NMS is more complete in favor of following image matching process.Finally, a new local contour matching method based on sub-matrix was researched to solve theproblem that there is rarely local contour matching method available for the occlusion problem.Firstly, distance matrix and tangent vector angle matrix were computed; secondly, k rank sub-matricesextracted from the matrices above to estimate whether there are matching parts or not; thirdly, RSTinvariant features were used in find image matching and the final matching curve segments and points.The experimental results show that the proposed image matching algorithm can be used when aretranslation, rotation and scaling transformation or even occlusion between two images. Comparedwith two methods based on global contour information, matching accuracy is improved by40%.