The Research On Image Mosaicing And Parallelization Based On Improved SIFT

Image mosaicing is an important direction of image processing research. It could be a good solution to certain applications (such as medical image analysis, remote sensing measurements, and so on) due to the hardware constraints of the image acquisition device, could not be collected large visual field and high resolution images. Therefore, the research on high precision image splicing technology and its rapid implementation has important theoretical and application value.Image matching is the core algorithm of image mosaic. The main disadvantages of matching algorithm commonly used based on SIFT Feature have:First, the large amount of calculation and slow speed operation; Second, the estimation accuracy of the main direction of the key points directly affects the stability of its descriptors. These made mosaicing results difficult to meet the actual requirements. Aiming at these shortcomings, image mosaic and parallelization based on an improved SIFT was proposed in this paper. The main contents are as follows:(1) In order to avoid the errors of the key points’ main direction estimated, an improved SIFT algorithm was proposed at the moment. The local coordinate system was established on the circular area at the center of key points.(2) The nearest neighbor algorithm was put forward to match these key points. In order to delete these points matched wrong, the RANSAC algorithm was put forward.(3) Genetic algorithm was proposed to resolve the projection model parameters. As the reference frame of the images acquired might be different, the image sequence was projected to the same coordinate system before image stitching. To resolve the projection parameters, genetic algorithm was used. The image sequence projected was spliced with the method of weighted average fusion algorithm.(4) Due to SIFT, RANSAC algorithm has large operation that can be converted into parallel operation. These operations were made in GPU (Graphic Processing Unit) platforms.The algorithm proposed in this article was respectively tested on the CPU and GPU platform. The experiment indicated that the algorithm researched in the article conformed to the human visual effect, and the parallel algorithm could achieve several times speedup (the speedup varies according to the input image resolution, the higher the resolution, the larger the speedup). The algorithm can meet the real-time requirements.