Research Of Image Registration And Filtering Method Using CUDA

As one of the hot topics in the field of image processing,image registration aims to correct and align two or more pictures,which have different original location,by performing an optimal geometric transformation.With image registration technology,people can share and complement the relative information.As an indispensible process in image preprocess,image filtering aims to enhance the vision quality of images by restraining noise while reserving the images detail.As the key technologies in image processing,both image registration and filtering are widely used in many fields,such as computer vision,3D reconstruction,pattern recognition,artificial intelligence and image analysis.The main problem of image registration is its complexity and huge amount of computation.As a result,it is hard to process in real time,and then greatly constrains its application in projects.With the rapid development of NVIDIA Compute Unified Device Architecture(CUDA),it is possible to combine software and hardware to realize parallel acceleration.Due to the advantage of parallel programming and bi g data processing of CUDA,this paper researches the image registration algorithms based on CUDA and discusses its parallel programming,theory,architecture and parallel processing mechanism in detail.This paper first introduces the application prospect and state-of-art development of image registration algorithms and then introduces the basic theory of SURF and its parallel process ways.This paper mainly studies some parts of SURF algorithm,such as establishing the integral images,structure of scale space,location of the key points,determining main direction of feature points,construction of feature point vectors,and feature points matching,on GPU kernel function.As a new image processing algorithm,guided filtering not only has the advantages of bilateral filtering,but also overcomes the influence of gradient reversal artifacts.Moreover,the guided filtering is widely used due to its close relation with the Laplace matrix.For the shortcoming of Guided filtering,such as slow operational speed and non-real time processing,the algorithm is speeded up using CUDA.In the proposed method,the sum of neighbor pixels' value is calculated based on parallel programming,and then the mean value is calculated.The key parameters of guided filtering are obtained using texture memory and registers,as well as algorithm optimizing.Finally,the whole optimum of approach is achieved.Experimental show that both the proposed image registration and guided filter algorithm based on CUDA have a good acceleration results.Experimental results show that the speed of algorithm has a big improvement after parallel processing and can meet the requirement of real-time processing.