Research On Image Local Feature Matching Based On Nonlinear Pyramid

Image extraction and matching algorithm based on local feature points is a major research direction of computational vision,which is an important part of image Mosaic,video Mosaic,virtual reality and other fields.In practical applications,the images to be matched will have certain scale differences,so feature point extraction and descriptor establishment of the matching algorithm are carried out in the scale space.There are two ways to construct scale space: the first is to use SIFT algorithm with the help of gaussian convolution kernel function.The second method USES KAZE and AKAZE algorithm to construct a pyramid through nonlinear diffusion filtering,which can better protect the edge information while smoothing the image content.KAZE algorithm M-SURF floating-point descriptors,AKAZE algorithm M-LDB descriptor matching speed is slow,traditional binary descriptor are established according to the relationship between the two pixels,Causing they have poor robustness,and are vulnerable to illumination,scale,and noise interference,AKAZE/KAZE algorithm combined with the LATCH descriptor AKAZE/KAZE-LATCH algorithm is proposed.LATCH descriptor code with the aid of the relationship between the three pixel block,which has stronger robustness.When the AKAZE/ KAZE-LATCH algorithm is tested in the open data set,the matching accuracy and speed of the algorithm are significantly improved compared with the original KAZE and AKAZE algorithms,which is suitable for the scenarios with high requirements on matching speed and accuracy.ASV-KAZE algorithm based on KAZE algorithm is proposed for scenarios with high robustness requirements.Algorithm first uses KAZE algorithm in nonlinear scale space to detect feature points,then build multi-layer scale space for the feature points descriptor and the difference and absolute value of the scale space descriptors of two different layers are calculated,also set a threshold process descriptor,then the stability values of the descriptor obtained from all the scale pairs are added to obtain the ASV-KAZE(1M)descriptor of the first stage;Finally,and the ASV-KAZE(1M2M)descriptor of the second stage was obtained by binarization of the first stage descriptor.The ASV-KAZE algorithm was tested in the open data set,and the experiments showed that the ASV-KAZE(1M)descriptor and ASV-KAZE(1M2M)descriptor had strong robustness against fuzzy transform,light transform,Angle transform and JPEG transform,but the matching speed was not good enough.This paper has 31 figures,2 tables and 55 references.