Research On Technologies Of UAV Remote Sensing Image Mosaic

Remote sensing image of Unmanned Aerial Vehicle (UAV) have several advantages: goodflexibility, detecting of high risk areas, high resolution, low cost and high efficiency. UAV imageshave been applied widely in military, agriculture, city planning, geographic mapping andenvironmental monitoring, etc. In order to expand field of vision and obtain better results of UAVimage analysis, research and processing, two or many remote sensing images with overlap portionneed be spliced into one image, and it has a very important significance to study how to complete theimage mosaic accurately and quickly for the unique characteristics of UAV remote sensing image. Thepaper is focused on the research and improvement of image registration and image fusion algorithm,which are two key steps in image mosaic. The main work is as follows:Firstly, make a thorough study of the key technology of image preprocessing, image registrationand image fusion, which three are composed into image mosaic. In this part, several common imageregistration and image fusion technology are highlighted.Then, Scale Invariant Feature Transform (SIFT) algorithm cannot extract enough feature points forlow-contrast image and SIFT algorithm is not efficient most of time, an improved SIFT algorithm,which has an adaptive contrast threshold with a variable coefficient, is designed in the paper. Thealgorithm Initially identified contrast threshold according to the gray level information of the localneighborhood, and also determine the size of the coefficient according to the current number offeature points; more feature points, the greater the coefficient increases the contrast threshold so thatthe number of feature points avoids too large. The experimental results show that the improved SIFTalgorithm significantly enhances the robustness of the algorithm for low-contrast image matching,while the final number of the feature points is not too large and stabilized at a predetermined interval,so the algorithm efficiency significantly improves.And then, Random Sample Consensus (RANSAC) algorithm will waste a lot of time while lookingfor support set in the process of selection of the sample set; besides, if two candidate points are tooclose, they may be taken with one point so that mismatching will be produced. In order to solve thetwo problems, an improved RANSAC algorithm is used in the paper. The algorithm first divides thearea containing match points in the image into several blocks, and selects the match points from eachblock to make candidate points avoid too close; then dynamically selects a random sample set andmake sure that its number is always1more than number of match points required in calculating thetransformation matrix, thus time can be saved during looking for support set. Experimental results show that the improved RANSAC algorithm is more efficient and can eliminate the mismatch.Finally, the wavelet transform image fusion algorithm based on searching image stitch line isproposed in the paper. The algorithm first searches the best splicing points with the smallestdifference of brightness value row by row within the overlap region between input images, so that theideal stitch line can be determined, and then according to each scale and band, determine size of thetransition region of each layer’s image after wavelet decomposition, finally, select appropriate fusionrules to complete image fusion in the transition region of each layer. In the experiment, compare thisalgorithm with weighted average image fusion algorithm which is representative in the traditionalimage fusion algorithm, by both subjective and objective evaluation methods, the results show that,the proposed algorithm has high-quality result of stitch tracer eliminate based on ensuring real-time,besides, it avoids image blur, and achieves better fusion effect.