Affine Invariant Image Feature Extraction With Registration Algorithm And Application

Evolution Computer Vision(ECV)is a new research method combining evolutionary computation and computer vision.It represents a new interdisciplinary research topic that combines analytical methods with stochastic optimization and heuristic methods.ECV is committed to design hardware and software solutions that help solving computer vision challenges.In the early days of artificial intelligence,people tend to believe in the cognitive aspects of intelligence,and they underestimated the study of visual perception.Nowadays,computer vision is a successful research and application technology field.There are many ways to solve many practical problems,such as 3D modeling,tracking objects in image sequences,fingerprint recognition,finding a person in the scene,optical character recognition,etc.A common feature of these applications is the attempt to describe the world through one or more images and to recreate the properties of the scene,such as traits,colors,and lighting,as well as other attributes that help present the entire visual experience.This paper first introduces the state of the art of computer vision,and then combines the evolutionary algorithm to summarize the research results in the field of computer vision.Then it mainly discusses the genetic algorithm in the evolutionary algorithm.Then three aspects of research work are mainly carried out: 1.Studying the affine invariant image feature extraction algorithm,in which the ASIFT algorithm is discussed and improved.A more accurate and efficient RA-SIFT algorithm is proposed.2.Combining affine invariance and multi-objective genetic algorithm in evolutionary algorithm,image registration is studied,and a MOGA algorithm suitable for real-time registration of complex images is proposed.3.Attempting to apply the proposed algorithm to the real-time matching of vision images of "Qi-Zhi" ROS robots.In order to obtain the affine invariance of the image,the affine invariance algorithm is carefully studied.Routine algorithms for image feature extraction,such as SIFT,SURF,ORB,etc.,all have partial affine invariance to the image,so this paper focuses on the improved SIFT algorithm of SIFT,which is mathematically proved to be complete.The affine invariance has a good matching effect on the changes in the image angle of view.The 1st innovation of this paper is mainly embodied in:by ombining the RANSAC algorithm and the distance deletion to improve the ASIFT,contructing RA-SIFT algorithm--a higher precision and higher efficiency affine invariant feature matching algorithm,the experiment results show that compared with ASIFT,RA-SIFT improves the accuracy and efficiency of feature extraction significantly.By combining with affine invariance algorithm and multi-objective genetic algorithm,this thesis also proposes a complex image registration algorithm--multi-objective genetic algorithm with affine invariance(A-MOGA).This multi-objective genetic algorithm is necessary to set the objective function from 3 aspects: angle,distance and point matching to match the image.In the matching process,it is necessary to calculate the matching degree(fitness)of each individual(solution)in the population.According to the matching degree of each individual in the population of this generation,new individuals(new solutions)are formed by crossover and mutation operators.Experiments show that another innovation of this paper,A-MOGA algorithm,has obvious validity and feasibility: compared with the nearest neighbor similar matching method and the violent matching method,A-MOGA algorithm has obvious advantages for complex images with more feature points in terms of matching time length,and can meet the requirements of real-time image registration.This paper also attempts to apply the proposed A-MOGA algorithm to vision image matching of "Qi-Zhi" ROS robot: adjusting algorithm parameters according to hardware parameters and software system of ROS robot,pre-acquisition of reference images,image matching experiment between real-time image of ROS robot and these reference image.The experimental results show that the proposed A-MOGA algorithm can be applied to ROS robot effectively.