Image Registration Techniques And The Optimization Algorithm

Image alignment is a fundamental and widely researched subject in area of image processing and recognition. It's generally defined as a procedure of aligning images which are obtained by different imaging equipment or under different environments, and calculating the model which cause the distortion of images. Applications of image alignment exist in many areas including military imaging, remote sensing, medical imaging, and object recognition.The Innovation points of this paper include:1. This paper puts forward a simulated annealing-based image alignment method to solve the optimization of object functions which can jump out local optimums for its feature of random walking and finally find the global optimum. Classic optimization methods such as Lucas-Kanade and its variations may lead the solution falls into local optimums for its feature of gradient descending and these kinds of method rely on initial values strongly. And results of these two kinds of methods are compared.2. Normal methods of image alignment assume that all pixels have equal weights, but in some applications, regions of interests tend to be aligned rather than others such as face alignment and object tracking. This paper imports saliency map to image alignment and fix this problem well.3. To correct the gray-scale distortion, many methods pre-process the images simply, but larger distortion or more strict requirements lead this method to ineffectiveness. This paper brings in the gray-scale transformation model to the object function and solve it synchronously, and verifies its effectiveness by experiments.4. Apply image alignment to image mosaicking, and put forward bi-directional matching for feature points, then remove outliers by RANSAC method for accurate results, and finally use simulated annealing to solve perspective transform matrix for avoiding entrapping in local optimum.