The Algorithm Research And Implementation Of Image Registration

How to achieve the image registration algorithm, which is fast, robust, and accurate isa fundamental problem in computer vision. This paper studied and implementedhomologous visible images registration, and heterologous visible and infrared imagesregistration. For the problem of visible images registration, we design and implement aregistration algorithm that basing on feature points; for the problem of visible and infraredimages registration, we design and implement a registration algorithm that basing onParticle Swarm Optimization and Mutual Information; for the feature of highcomputational complexity and parallelism, we use heterogeneous program with CUDA toaccelerate these algorithms.The main contents are as follows:(1) For the visible images registration, in order to make feature points have scaleinvariance, we construct a Gaussian scale space, and extracting the local maxima points inthe Gaussian pyramid by covariance matrix. In the local area around the feature points, tomake feature points have rotation invariance, we determine the main direction of the localarea by statistical moment invariants. In order to make the feature points have highrecognition and have low sensitive to blur or changes in brightness, we get a256bitsbinary string in Gaussian distribution sampling pattern from the area around the featurepoint. Through a lot of experiments，we prove that the algorithm we designed has a highcorrect matching rate, high robustness, and high computational efficiency.(2) For the parallelism in image preprocessing algorithms of feature point’sextraction (filtering and up sampling), constructing Gaussian pyramid, and computingcovariance matrix, we achieve a heterogeneous GPU-CPU programming by CUDA C.Experiments show that we get a very good speedup ratio. Through experiments we analysisthe factors affecting the acceleration ratio, and obtain the condition that affects the extremevalues of speedup ratio through theoretical derivation.(3) For visible and infrared images registration, in order to get result gradually bychanging resolution from low to high, we use Mutual information as similarity measurecondition, and use the Particle Swarm Optimization algorithm to search for the maxima insearching space. Experiments show that the algorithm improves the computationalefficiency without reducing the accuracy.(4) For the parallelism in Particle Swarm Optimization algorithm, Mutual Information algorithm, and affine transformation, we achieve a heterogeneous GPU-CPU programmingby CUDA C. Experiments show that we get a very good speedup ratio. Throughexperiments we analysis the factors affecting the acceleration ratio, and obtain thecondition that affects the extreme values of speedup ratio through theoretical derivation.