Research On Image Feature Extraction And Matching Algorithm Based On Nonlinear Scale Space

The image feature extraction and matching technology aims to match two images containing the same scene that collected under different conditions through some specific algorithms,and seek the similarity and transformation relationship between images.There are two categories of image matching algorithms: one is gray-based matching algorithm,and the other is feature-based matching algorithm.In the feature-based matching algorithm,the images are described and matched by using more stable local features.Currently,the popular local feature matching algorithms(such as SIFT algorithm)extract feature information based on Gaussian linear scale space,in the process of constructing scale space,Gaussian blur will cause loss of edge and detail of the image,resulting in reduced accuracy of feature detection and matching.However,the KAZE algorithm can effectively solve the above problems by using nonlinear diffusion filtering instead of Gaussian blur,but the KAZE algorithm also has problems such as lack of color information and insensitivity to changes of image viewing angle.Therefore,this thesis mainly proposes improvements to the problems existing in the KAZE algorithm.The specific work has the following two aspects:(1)Aiming at the problem that the KAZE algorithm lacks the color information generated during the graying process,which makes the extraction of some feature points difficult and the correct matching rate is low,a feature extraction and matching algorithm based on improved image graying is proposed.The algorithm first calculates the contrast between pixels in the color image by weighting the Euclidean distance,then optimizes the multivariate polynomial model representing the linear combination of colors to obtain the gray value of the pixel.Finally,the bimodal distribution function is used to minimize the grayscale difference between pixels and the inter-pixel contrast to obtain the grayscale image,so that the grayscale image retains more color contrast information in the original image;the feature extraction is performed on the basis,and the second-order gradient values mean of the neighborhood region of the feature points are calculated to replace first-order gradient values in the original descriptor to form a new descriptor,then performs image matching.The feasibility and effectiveness of the algorithm are verified by the final matching effect.(2)Aiming at the problem that the KAZE algorithm is not sensitive to changes of image viewing angle and the matching result does not have affine invariance,a feature extraction and matching algorithm based on affine invariance is proposed.Firstly,according to the principle of camera imaging,the local affine transformation of the image can be used to replace the image perspective transformation.Secondly,this method is used to construct the affine transformation image set.By simulating the transformation parameters such as scale,longitude and latitude,and normalizing the translation and rotation generated during the process,the simulation of the whole affine transformation space is realized.Then feature detection is performed on a series of simulated images,and the similarity of features is measured by combining the affine invariant property of Mahalanobis distance to complete the final matching process.Finally,the algorithm is verified by experiments to have affine invariance.