Research On Edge-based Image Registration Algorithm

Image registration means: aligning two or more images in space which is obtainedby different sensors at different times or under different imaging conditions but owe thesame scene.It is one of basic problems in image processing. The effect of registrationwill have a direct impact on the follow-up work.For example, Image mosaic, imagerecognition, tracking and other subsequent work, etc. In this paper, based onfeature-based analysis of image registration, feature points of the edge-based imageregistration and edge-based image registration has been studied.Based on edge feature point image registration Contains two important aspects:The first is the image edge feature point extraction, the second is matching criterion.The accuracy and stability of feature point extraction will impact on the work of thefollow-up registration. Feature points have better adaptability on the position change,gray scale, noise, partial occlusion in the extraction process. This paper matches thefeature points extracted as follow: first, based on the image edge characteristics, usingdynamic support domain of the improved CSS algorithm to extract edge corner, andthen using the singular value decomposition and the Mahalanobis distance method tomatch the feature points extracted. Singular value decomposition use data analysismethod and find out a large quantity of data implied by the model, and then matching.The experiments show that the singular value decomposition has invariance of rotationand translation. The size of Mahalanobis distance does not only relative to distributionof each set of points, but also to the distribution of its own. Using Mahalanobis distancematch for different sensor image. The experiments show that the Mahalanobis distanceis practical, and the excision for mismatch need to be further improved.Edge-based image registration is based on image edge feature unit. The Hausdorffdistance descript the similarity measure of two points, and has strong anti-interferenceability and fault tolerance.However, the simple Hausdorff distance is more sensitive to noise and outliers, leading to a higher rate of false matches.In the paper, using the formof improved Hausdorff distance which use the weighted average remove some specialpoints and achieve a portion of mean. As the influence of global search to registrationspeed in the registration process, introducing genetic algorithm as the search strategy.In genetic operator, using the improved MSE-Hausdorff distance as the similaritymeasure to construct the adaptive function of genetic algorithm, then determine theoptimal transform parameters, complete the registration of the benchmark image and theregistration image. Experiments show that the improved Hausdorff distance can wellovercome the impact of noise and partial occlusion on image registration accuracy andgreatly improved the operation rate.