Research On Image Registration And Image Stitching Based On Feature Points

With the continuous development of computer vision,people's demand for visual images is getting higher and higher,including the more abundant information in the picture and the pursuit of broader vision.From hardware perspective,people use larger screens with higher-resolution,and take photos with wide-angle lens cameras.While getting wide-angle images with efforts on hardware is expensive and inconvenient to change after equipment.From the other perspective,using software processing techniques,people can achieve the same goal from taking some normal images beforehand and stitching them,which is combining a clear wide-angle image with many images only with the help of a compute.With the advantages of platform compatibility and low cost image stitching algorithm is widely used in aviation,medical,virtual reality,military and other fields.However,the current techniques have strict input conditions,which need the imaged scene is planar.If the input images have different scene depth in their overlap regions,two problems related to image registration which including mismatch in overlapping regions and shape distortion in the other regions would not be escaped.Therefore,incorrect alignment and shape distortion problems under non-planar scenes are major issues for this paper.Currently,the researches on non-planar images registration have obtained some progress,including the local projection transformation method to improve accuracy and the shape correction based on lines to reduce shape distortion.But the state-of-art local projection transformation method lack of reasonable local regional partitioning mechanism,so the computing efficiency problem needs to be solved urgently.While the effect of existing shape correction method is limited to line characteristics which is difficult to expand the scope of use.In this paper,two new solutions are proposed:· Proposes a regional partitioning mechanism based on feature point clustering and quad-tree:The algorithm takes advantage of the inherent characteristics of feature point to do image partition.At first,we introduce k-means clustering analysis to label categorized feature points,then use quad-tree to divide images into local regions in different levels according to the label of key-points.After that estimate local projection transform matrixes for each grid and do image warping.Using the above intelligent region division,the number of local regions as well as projection matrixes are reduced,so it is more efficient to do local projective warping.From this algorithm,the efficiency is improved by reducing the computation redundancy and the registration accuracy of non-plane scenes is guaranteed at the same time.· Proposes a projection-similarity transformation model with regional control:Similarity transformation has been applied in some shape protection issues for its property of keeping the proportional relationship between point and point after transformation,but misalignment occurs in overlapping areas while dealing with the problem of non-plane scene image stitching.Thus,on the basis of the projection transformation,the proposed method introduces similarity transformation,and combines the projection-similarity transformation model with regional control method to achieve shape preserving.The projection-similarity matrix is construct by linear combination of the projection matrix and the similar matrix,and the relevant weights are controlled by local regions.The closer the overlap region is,the stronger the projection matrix is,and the further away from the overlap region,the stronger the similarity matrix will be.