Research On Image Mosaic Based On Feature Points

Image mosaic technology is a significant research area which involves digital image processing, computer graphics and computer vision. With the continuous development of image splicing technology, it has been widespread applied in virtual reality, remote sensing, medical image analysis and military affairs. Image splicing quality is decided by image registration precision. Feature-based image registration technology has good robustness when the images involve translation, rotation, scaling, illumination and view-point transformation. It has many advantages, such as, high registration accuracy and good splicing effect. So the study on feature-based image registration and mosaic technology has significant research value and application value.In this paper, the main work was focused on feature-based image registration and mosaic technology. It contains image preprocessing, SIFT features extraction, image registration and image fusion. Algorithm principle is analyzed in detail and specific realization algorithm is proposed. At last, image mosaic is realized by MATLAB 2012b. The followings are the main research work:(1) The basic geometrical principles of the camera was studied. The relationship of the world coordinate system, the plane coordinate system and the camera coordinate system was clarified. Some basic movement forms of camera were introduced and several image transformation models were analyzed.(2) The extraction of feature points. Establishing scale-space, detecting local extreme points, determining accurate positioning of local extreme points, and generating characteristic vector descriptor were processed to maintain rotation invariance and scale invariance of the extracted feature points. It determined the position, direction and the scale of each feature point, and ensured the higher stability of detected feature points.(3) Image registration mainly includes the initial matching and accurate matching. The initial matching points were determined by the Euclidean distance of similarity criterion that was compared the ratio of Euclidean distance between nearest neighbor feature points and next nearest neighbor with set the distance threshold. If the ratio was less than the distance threshold, it would be the initial matching point. The accurate matching was determined by using improved RANSAC algorithm that mismatching points of initial matching are removed. The improved algorithm was that error provisional model was excluded by method of pre-detection feature points. It reduced iterations of the algorithm, saved stitching time and improved stitching speed. And all feature points are processed by the average block to ensure that pre-detected feature points could be distributed evenly. It improved candidate model accuracy to improve the candidate model accuracy, and improved the matching precision. This results show that improved RANSAC algorithm improves the stitching speed and accuracy on the basis of the stability of the traditional RANSAC algorithm.(4) Image fusion. The sub-regional fusion image method was used to eliminate splicing gap of image overlapping area and to make image smooth. The pixel values of each point on not-overlapping area of two images are determined by bilinear interpolation method. And fade weighted average method was used for the fusion of overlapping area of two images. The results show that the trace of image splicing is eliminated by the fusion method, and stitching quality is improved.