Research Of Image Registration And Mosaic Technology Based On GPU

Image registration technology is widely used in many fields, such as imageprocessing, computer vision, artificial intelligence, machine learning, medical image,national defense science and technology, and geological exploration. Imageregistration technologies based on feature points are generally used due to its preciseregistration, stable performance, strong robustness and high adaptability.General parallel computing technology has made great improvements in recentyears, the NVIDIA GPU technology based on CUDA architecture has been rapidlydeveloped and widely used in intensive data processing fields. This paper firstlyintroduces the development of NVIDIA GPU, mainly on CUDA hardware andsoftware architecture, such as CUDA storage model, CUDA software system, CUDAprogramming model and CUDA hardware execute model. Moreover, this paperdescribs some heterogeneous system models, such as Host code (run on CPU) andDevice code (execute on GPU).The thesis discusses the Speeded up Robust Features (SURF) algorithm in detail,especially on its feature point extraction and description process, which has higherparallel feasibility. Moreover, the SURF algorithm is implemented in CPU+GPUheterogeneous system, in which CPU is responsible for memory allocation andprocess control, while GPU is responsible for parallel implementation of kernelfunctions. In addition, this thesis mainly studies parts of the SURF algorithm, such asthe gray transformation process of color images, the establishment of integral images,calculation of Hessian determinant, the maximum inhibition, main direction of featurepoints determinating, and generation of feature point vectors’ realization, on GPUkernel function.In the final part, the simulation experiment is tested on synthetic images and realimages respectively, which have translation, scaling, rotating and complex scene.SURF algorithm is implemented on GPU system to extract feature points and generatedescribed vectors, and then feature points matchment are performed on CPU. Afterthat, the affine transformation is calculated according to the matching points, the finalimage registration results are obtained by least square method, and finally the imagemosaic is implemented. The experimental results shown that sub-pixel accrancy registration can be achieved by SURF algorithm based on GPU system, and perfactvisiual performance of final mosaic images can be gotten as well. Moreover, therunning time of GPU system can be decreased greatly compared with CPU system.The execution speed acceleration ratio can be about11~14on average. This paperlays the foundation for real-time image mosaicing in the future.