摘 要

图像配准和拼接技术作为图像处理的一个重要分支有着不可取代的作用，人们可以使用图像拼接技术，将多幅有重叠区域的图像拼接成一幅高分辨率、广视角的全景图，这样的全景图可以应用到医学图像分析、全景视频监控、军事等领域。基于特征点的图像配准和拼接技术的核心步骤为特征提取，目前大多数特征检测算法的实时性和鲁棒性较差，得不到较好的拼接效果，而SURF算法是一种快速鲁棒性特征检测算法，有良好的检测性能，所以，基于SURF特征的图像配准和拼接技术有一定的优势。

本文主要研究的基于SURF特征的图像配准和拼接技术，其主要流程为图像采集、预处理、图像配准、图像融合等。其中图像配准是图像拼接的核心步骤，主要分为特征检测、特征匹配、变换模型估计3个方面的内容。在特征检测环节，本文主要研究Harris、FAST、SIFT、SURF特征检测算法的原理和特征提取过程，并从特征点数量、检测时间、可重复性三个方面对比四种算法的性能，结果表明，SURF算法的综合性能优于其他3种算法，最后对SURF算法的可重复性做了详细分析，包括旋转、尺度、模糊、亮度变化四个方面。在特征匹配环节，主要分为粗匹配和细匹配，在最短距离匹配的基础上，使用比率检验算法和交叉检查算法结合的改进粗匹配算法，细匹配则是利用RANSAC算法剔除误配对，经过粗匹配和细匹配后可以得到没有误配的匹配项。在变换模型估计环节，本文采用了单应性矩阵变换模型，并使用没有误配的匹配项和奇异值分解法估计单应矩阵参数。图像融合同样是图像拼接的重要步骤，本文主要研究像素级的图像融合算法，并实验比较均值、最大值、加权平均、改进加权平均融合算法的效果，实验表明，改进的加权平均融合算法能有效地去除拼接缝，而且能很好的保留图像细节信息。其次，本文主要研究柱面全景图拼接的实现步骤，并使用扭曲度的评价标准对拼接质量进行分析。另外，图像拼接只需要图像间的重合区域的细节信息，而传统的拼接算法在特征检测时对整幅图像进行检测，增加了特征匹配的计算量，影响了拼接的效率，因此本文使用基于重合区域的特征检测优化方法来提高拼接效率，实验表明，该方法大约可以减少一半的拼接时间，且拼接图像数量越多拼接效率越明显。最后，利用Qt Creator界面开发工具和Opencv计算机视觉库搭建了图像拼接软件系统，并对系统做调试分析。

综上所述，本文研究的图像拼接算法基本步骤为先对图像进行SURF特征提取，再进行特征匹配，然后求出单应矩阵并进行图像变换，最后经过图像融合后得到拼接结果。研究结果表明，此次的研究成果在实现两幅图像配准的基础上可以实现两幅或多幅图像的拼接，且均能得到令人满意的效果，误差较小，同时，与传统的拼接算法相比，拼接效率明显提升。

关键字：SURF算法、图像预处理、图像配准、单应矩阵、图像融合、柱面全景图拼接

Abstract

Image registration and stitching technology as an important branch of image processing technology has an irreplaceable role, one can use image stitching technology, multiple overlapping areas of the image spliced into a high-resolution, wide viewing angle of the panorama, it can be applied to medical image analysis, panoramic video surveillance, military and other fields. Finally, most of the feature detection algorithms are less real-time and robust, and no better detecting effect is obtained. The SURF algorithm is a kind of fast and robustness feature detection algorithm, a good detection performance, so, based on the SURF feature of the image registration and splicing technology has certain advantages.

In this paper, based on the SURF feature of image registration and stitching technology, the main process for image acquisition, preprocessing, image registration, image fusion and so on. The image registration is the core step of image stitching, which is divided into three aspects: feature detection, feature matching and transformation model estimation. In feature detection, the principle and feature extraction process of Harris, FAST, SIFT and SURF feature detection algorithms are studied in this paper. The performance of four algorithms is analyzed from three aspects: the number of feature points, the detection time and the repetition rate. Experiments show that the performance of SURF algorithm is better than that of the other three algorithms. Finally, the performance of SURF algorithm is analyzed in detail, including four aspects: rotation, scale, fuzzy and brightness. In the feature matching process, mainly divided into rough matching and fine matching, based on the shortest distance matching, using the ratio checking algorithm and the cross-checking algorithm combined with the improved feature rough matching algorithm, the fine matching is the use of RANSAC algorithm for mismatching , after rough matching and fine matching can be no match with the match. In the case of transformation model estimation, this paper uses a homography matrix transformation model and estimates the single matrix parameters using mismatched matches and singular value decomposition algorithm. Image fusion is also an important step in image splicing. In this paper, we mainly study the image fusion algorithm based on pixel level, and compare the mean value, the maximum value, the weighted average and improve the weighted average fusion algorithm. The experiment shows that the improved weighted average fusion algorithm effectively remove the stitching seam, and can be very good to retain the details of the image information. In addition, the image stitching only needs the detailed information of the coincidence area between the images, and the traditional stitching algorithm detects the whole image at the time of feature detection, increases the computational complexity and interference of the feature matching, and affects the efficiency of stitching. The experiment shows that the method can reduce the stitching time by about half, and the more the number of splicing images is more obvious, the more the stitching efficiency is. Secondly, this paper uses the evaluation criteria of distortion to analyze the quality of image stitching. Finally, the Qt Creator IDE tool and Opencv computer vision library are used to build the image mosaic software system, and the system is debugged and analyzed.

In summary, the basic steps of the image stitching algorithm in this paper are to extract the SURF feature of the image first, and then feature matching. Then, the single matrix is obtained and the image is transformed. Finally, the result of stitching is obtained after image fusion. The results show that the results of this study can achieve two or more images on the basis of two image registration, and can achieve satisfactory results and the error is small, at the same time, with the traditional stitching compared with the algorithm, the stitching efficiency is improved obviously.

Keywords: SURF algorithm, image preprocessing, image registration, homography matrix, image fusion, cylindrical panorama stitching