Research On Image Registration And Mosaic Based On CUDA

Image is the visual form of the world’s energy and state, which provides a wealth of information for people to understand the world. The image registration and mosaic, one of the basic problems of image processing, has a wide range of applications in the field of virtual reality, geological surveying, medical imaging, weather forecasting, emergency response, etc. Unlike traditional CPU, GPU use more crystal unit cell as compute unit, greatly improving the computing power. GPU technology has been widely used in data mining, image recognition, genetic engineering, global climate forecasting accuracy, also provides a new solution for processing of remote sensing images.In view of the current image registration and mosaic computational bottleneck problem, based on the analysis of the current image registration and mosaic technology, combined with CUDA parallel computing technology, the research is mainly focused on large-size remote-sensing registration and real-time mosaic method for aerial images. The main research contents include:(1) Large remote sensing image registration. For traditional image registration methods are difficult to adapt to the large-size high-resolution remote sensing image registration problem, we propose a coarse-to-fine control point matching method. To fix local distortion, image rectification is based on a linear mapping function computed from triangulations, and a self-adaptive scan filling algorithm is proposed to determine which triangle each pixel belongs to. Through the analysis of computing bottleneck problem in registration, the control point matching and image rectification processing are accelerated with CUDA parallel computing techniques. The experiment results with a Pair of IKONOS Panchromatic Images, Geoeye panchromatic and Geoeye multispectral images, ZY-3satellite image, show that this method can achieve high accuracy registration results and a higher speedup.(2) Real-time online aerial image mosaic. For the bottleneck of real-time online aerial image mosaic, this paper presents a method of image mosaic based on the CPU and GPU co-processing. Based on POS data, the CPU is used to calculate transform relations between the original image and the corrected image, and then using the GPU parallel computing to realize image correction. Because the aerial image usually has a larger degree of overlap between images, so an adaptive splicing method was employed to reduce redundant computation. Simulation experiments are carried out with two computers, one of them simulates a camera to shot and transform images to the processing machine. The experimental results show that the proposed method can achieve online real-time mosaic.The image registration and mosaic method proposed in this paper has been implemented by using Visual C++, and the corresponding prototype system is developed. The experimental results show that the method in paper can achieve efficient registration of large remote sensing images and the aerial image online stitching.