Research On Image Registrations And Cylindrical Panoramic Image Constructions

Constructing panoramic images is an important research content of imaged-based rendering (IBR), concerned with many subjects such as Computer Graphics, Image Processing and Calculating Geometry. The panoramic images may express complete surroundings information of a scene from one fixed viewpoint. The constructing processes include four steps normally: choice of mapping patterns, collection of data, images matching as well as images blending.The images stitching is the one of key steps in the constructions of panoramic images, which generates a scene mosaic or full scene by the overlapping images taken by (videos) cameras circumrotation or parallel move. Some kinds of algorithms of the images stitching have been studied in this paper:In the equidistant matching algorithm for constructing cylindrical panoramic images, the color differences for all correspondence points in strips need to be calculated so that matching speed is slow. A fast cylindrical panoramic image mosaic algorithm is proposed to speed up it. In two overlapping images, the two regions selected respectively are moved with the same step. The color differences of corresponding sampling pixels which are got by random function are calculated and the differences are accumulated. For any step, if the accumulated value is over a threshold, the calculation of this step is stopped. For all steps, the position with the most accumulated numbers is the best mosaic location. For an image sequence with arbitrary order, a new order with real scene content, in which neighborhood images are overlapping, can be sorted using the ideal of the algorithm.When stitching images by phase correlation algorithm, the color values of original images are carried on Discrete Fourier Transform directly. So the algorithm is sensitive for illumination. A contour phase correlation algorithm has been studied. Firstly, images contours are extracted. Then the contour images are calculated by phase correlation algorithm. There are several peak values close to the maximal peak value in the array of 8 function by DFT. It is difficult to select the right peak value. Every of the peak values correspond to one pair of parallel move values (x, y). The all pairs of parallel move values are regarded as candidates to get the best stitching location. For each pair of parallel move values, the corresponding overlapping part of two images are carried on phase correlation algorithm again and the rightpeak value is obtained. The parallel move value corresponding to the right peak value is the final value.In the traditional template matching algorithm, complicated calculations cause a slow matching speed. Though Sequential Similarity Detection Algorithms (SSDA) may reduce the calculation quantity, it is difficult to select right threshold. A SSDA of automatically changing thresholds has been studied. After selecting the initial threshold, thresholds are updated unceasingly in the latter matching calculations. Following to the principle of stopping the calculations while exceeding the current threshold value, the best matching position is sought. The algorithm studied in the paper may not only hold the real matching point but also reduce the calculations.A system of image registrations and panoramic image constructions has been exploited based on the images stitching algorithms studied in the paper. The system includes several modules, such as the improved equidistant matching algorithm, the contour phase correlation algorithm, the SSDA of automatically changing thresholds algorithm and stitching images after be sorted and so on. The better effects of image registrations and panoramic image constructions can be got.