Research On The Matching Algorithm For Infrared Images

Nowadays, more and more experts have noticed the special features of infrared image, and are going to use relevant technologies in their fields. Under this circumstances, this paper is mainly focus on improving matching algorithms for infrared images. The main contents of this paper include five parts as below:First, this paper does some research about some algorithms based on grayscale of the image and mainly focus on two algorithms phase correlation algorithm and log-polar coordinate mapping algorithm. To overcome the Shortcomings of those two algorithms, an improved algorithm is proposed to deal with the condition which involves translation, scaling and rotation at the same time. The improved method separate high-frequency signal from low-frequency one, and then use log-polar coordinate mapping algorithm to deal with those high-frequency signal to figure out the scaling and rotation parameters between those images. Once scaling and rotation parameters are known, the new image could transformed corresponding and then use normalized correlation algorithm to figure out the translation parameter.Second, considering the algorithms based on grayscale of the image would have to use all the pixels of each image, in order to reduce the amount of calculation, those algorithms that use features of each image for matching was studied in this paper. The Speeded Up Robust Feature algorithm is the key algorithm of this branch. The Surf algorithm was used to detect and describe the feature points of each paper, then the euclidean distance with the nearest neighbor distance ratio strategy was used to match those feature points. Finally, random sample consensus was used to estimate parameters of a mathematical model and figure out translation, scaling and rotation parameters.Third, in order to transplant the matching algorithm to hardware condition in a better way, feature signal combined with domain transformation was proposed. This paper studies the algorithms based on feature points or edge features together with domain transformation. It can be seen in the experiment results that these improved algorithms need less calculation and memory.