Image Mosaic Technology Based On Multi-scale Feature Point Clustering And Wavelet Fusion

This thesis studies image registration based on the feature point extraction and wavelet fusion algorithm.Focused on the issue that the effectiveness and the precision of traditional methods are poor,the paper combines multi-scale theory and Harris corner algorithm to extract the feature points,and the images are mosaicked by means of wavelet transform fusion method and the feature points.The specific work and innovation are as follows:Firstly,the Harris corner algorithm is combined with multi-scale theory and the feature points are described based on SIFT descriptor.The Harris corner algorithm with better performance is used to extract feature points in the process of image registration,but the feature points of Harris algorithm are only in single scale image,so they have no scale invariance,and it is difficult to describe the local invariant of the feature points.Therefore,the paper makes two improvements: Image and Gauss kernel do convolution operation to generate a series of scale space images and then the Harris corner points are extracted in each scale space image,so we can locate the feature points more accurately based on their coordinate position and scale.Then,SIFT descriptor is introduced in detail and then the feature points are described based on SIFT descriptor.Secondly,an image registration algorithm based on multi-scale feature point clustering is proposed.The local feature of the image is detected repeatedly within a certain range of scales when detecting multi-scale corner in image registration and it may generate lots of redundant points,so the algorithm is improved to eliminate redundancy and increase the matching accuracy.In this paper,the thought of clustering method is used in multi-scale corner extraction algorithm.The Harris corner points are extracted in each scale image,then the feature points of the largest scale image are in a category of their own,and then computing the Euclidean distance between the feature points of the small scale image and the feature points of the category in turn,the feature points of which are clustered with the idea of AGNES algorithm,selecting the corner point with the maximum response of each group as the feature point.The experimental results show that this method can eliminate lots of redundant points and improve the registration accuracy.Thirdly,a large number of experiments are carried out to verify the effectiveness and robustness of the proposed registration algorithm.The contrast experiment is adopted to analyze the accuracy of the proposed registration algorithm based on multiscale feature point clustering.In addition,the images may exist rotation changes,brightness changes,scale changes and noise in the application,so in the experiment,the images are respectively changed with the angle,brightness,scale and noise to analyze the robustness of this registration algorithm to image changes.The experimental results show that the robustness of the proposed registration algorithm to image rotation,brightness,size and noise is enhanced.Finally,the wavelet transform is adopted to improve the visual effect of the mosaic image.It does not play a good effect to eliminate the splice trace when using the traditional weighted average fusion,so the wavelet fusion with three layer is adopted to fuse the images that have been registered,which can effectively overcome the shortcoming of weighted average fusion method weakening the details.In the end,the paper makes the qualitative and quantitative analysis of different scenes image,the multi-dimensional contrast experiments verify the advantages of the proposed method in image mosaic.