Research For Image Matching Based On Improved SIFT Algorithm

Image matching is a main component of image processing area.The improvement of its speed,accuracy and robustness is always the focus for researchers.Image matching methods contains two categories: one is based on the gray,the other is based on the feature.The former algorithm is relatively simple but sensitive.It is mainly according to the gray information of image,so the matching efficiency is not ideal.The adaptation to invariability in deformation of the latter is relatively ideal.SIFT(Scale Invariant Feature Transform)is the stable algorithm based on feature.It can deal with the inference of rotation,illumination and others,but also exists the problems on efficiency and accuracy.This paper firstly analyzes SIFT algorithm,and then improves it.The improved algorithm is proposed for the images of prominent boundary.The image of SIFT algorithm is gray image.For the images with prominent boundary,this paper uses threshold images instead of gray ones.As the the pixel value of threshold image is 0 or 1,the boundary is clearer.Meanwhile it also can simplify the SIFT feature descriptor which reduces the matching time.In addition,the feature descriptor of 128-dimensional brings a huge amount of calculation.Descending dimension method in this paper makes a half of dimension of feature descriptor,but the 64-dimensional feature descriptor covers all information of128-dimensional feature descriptor.The descending dimension method which is used in this paper not only reduces operating time of algorithm,but also has no effect on veracity of algorithm.According to the study,it is not difficult to find that the matching with Euclidean distance reduces the veracity of algorithm.In order to improve the veracity of algorithm,this paper use the weighted Euclidean distance instead of Euclidean distance.Finally,this paper uses a series of experiments to examine the performance of the improved algorithm.The results of experiments show that,when the illumination,rotation and viewpoint among images vary,the efficiency of matching is still good.And the operating time of algorithm is far less than the original algorithm.So the improved algorithm in this paper advances the instantaneity and veracity.