Parallel Implementation Of Visible-light And Infrared Image Registration Algorithms

Image registration is the process of superimposing two or more images of the same scene obtained in different sensors or different shooting conditions(eg,illumination,angle,etc.).Image registration is an important part of obtaining accurate image information.Since the visible image and the infrared image have complementary features on the imaging scene information,it is a hot research topic to combine the infrared sensor and the visible light sensor to obtain accurate and abundant image information.But the imaging mechanism of the two sensors is different,so there is a big difference between the visible image and the infrared image,which increases the difficulty of image registration.At the same time,image registration is widely used in various important fields,such as medical image processing,transportation,remote sensing,military,etc.,so the need on speed of image registration is higher and higher.In recent years,the rapid development of GPU technology has become the first choice for high-performance parallel computing in many important fields.How to achieve fast and accurate visible and infrared image registration is a hot and difficult problem in current multi-source sensor image registration.This paper implements the registration of visible and infrared images based on the GPU platform.The main contents are as follows.For the difference in resolution between visible and infrared images,bilinear interpolation is used to amplify the infrared image.SURF feature points are detected for visible and infrared images respectively,and the reason for the inapplicability of traditional SURF feature descriptors based on pixel gray value gradients for infrared image characterization is analyzed.Local self-similarity descriptors are used.Since the characteristics of the target local shape information can be described by local self-similarity descriptors,the feature points of the visible and infrared image detected by the same scene are described.In the feature matching stage,the problem of poorly identifying the mismatching of feature matching by the nearest neighbor method is constrained by using bidirectional matching.In this paper,to solve the problem of low computational efficiency of visible and infrared image registration by other algorithms,parallelization design is performed on SURF feature points detection,local self-similarity descriptor and nearest neighbor method combined with bidirectional matching constraint feature matching based on CUDA programming and different levels of optimization methods such as texture memory,constant memory and shared memory.In this paper,the performance of parallel design of visible and infrared imageregistration algorithms is evaluated in several scenarios.Whether from subjective vision or from the evaluation of objective indicators,the proposed algorithm achieves better registration accuracy than other registration algorithms.Compared with the serial algorithm,the parallel algorithm designed in this paper can significantly improve the efficiency of the algorithm without losing the accuracy,and obtain a good speedup.