【作者】 刘旭；

【导师】 张素文；

【作者基本信息】 武汉理工大学 ， 控制科学与工程， 2012， 硕士

【摘要】 全景图像拼接技术是数字图像处理领域的一个重要的研究方向,在文物保护(古籍或古文字资料保护)、计算机视觉、遥感图像处理、医学图像分析、计算机图形学等领域有着重要的应用价值。一个完整的图像拼接过程主要包括图像获取,图像配准,投影,图像融合等几个步骤,而其中的图像配准和图像融合是关键,若是多幅图像的拼接,还得考虑图像的排序。因此本文的研究也就着重于这三个方面。在图像配准方面,本文通过对基于灰度信息与基于特征的几种配准算法进行较深入研究后,选择采用基于尺度不变特征变换(Scale Invariant Feature Transform, SIFT)的图像配准方式。针对研究中发现的问题,提出了一种改进的匹配算法,即距离、角度、随机抽样一致性(Random Sample Consensus, RANSAC)三者相结合的算法。同时将匹配阶段分为粗匹配和细匹配两个部分。在粗匹配阶段,利用SIFT的特性将距离和角度结合,利用相似性来剔除误配点。在细匹配阶段,使用RANSAC算法提纯特征点并得到图像间的单应性矩阵,完成配准。实验结果表明：该算法与其他所选方法相比,耗时相当,精度更高,是一种不错的匹配方案。在图像融合方面,首先叙述了图像融合中的几种典型算法,接着对加权平均法,渐进渐出法和金字塔式融合分别进行了实验。实验结果表明：加权平均法运算简单,但效果最差；渐进渐出法改善了加权平均法的缺点,效果较好；基于金字塔式的融合,效果最好,没有明显的缺陷,但耗时较高。最后,根据实验结果,选择使用渐进渐出的融合算法。在图像排序方面,利用图像配准阶段算法的特点,使用基于粗匹配阶段的相匹配特征点对数来进行排序,匹配对数量多的就可能有匹配关系,最后利用概率模型来确认这种匹配关系。在此基础上,提出了一种自动的全景图像拼接方案,并进行了实验,其结果表明,该方案能对多幅图像进行排序,能剔除图像集合中不相关的噪声图像,拼接后的图像效果较好,速度合理,达到了预期的目的。更多还原

【Abstract】 The panoramic image mosaic technology is an important research direction in the field of digital image processing. There has a wide range of important practical value on protection of cultural relics(ancient or ancient text data protection), computer vision, remote sensing image processing, medical image analysis, computer graphics etc.A complete image mosaic process includes image acquisition, image registration, projection and image fusion etc. The key steps are image registration and image fusion. It has to consider the sort of image if the mosaic object is an image sequence. So, this paper focuses on these three points.In image registration, this paper choose an image registration method based on Scale Invariant Feature Transform(SIFT) after study deeply on based on gray information and the fcature of image registration algorithm. Propose an improved matching algorithm which is a distance, angle and Random Sample Conscnsus(RANSAC) combination of the three algorithms. The matching stage is divided into coarse matching and fine matching of two parts. In the coarse matching stage, use the SIFT to combine with distance and angle and weed out mismatches points with similarity. In the fine matching stage, use the RANSAC algorithm purification feature points and obtain the homography between images to complete the registration. The experimental results show that proposed algorithm compare with other algorithm has higher accuracy and time-consuming almost equal. It’s a good matching plan.In image fusion, this paper describes several typical algorithms, and then do experiment on weighted average fusion, fading in and out fusion and pyramid fusion. These experimental results indicate that the weighted average fusion is the simplest but least effective, the fading in and out fusion improve the shortcomings of the weighted average fusion and the effect is good. The pyramid fusion is the best, no obvious defects, but time-consuming higher. Based on the experimental results, choose to use the fading in and out fusion algorithm. In image sorting, using the feature of the image registration stage which are the number of matching points to do sorting. The most number of the matching points is the most likely to have matching relations and then use a probabilistic method to make sure this relations.Based on these, propose an automated panoramic image mosaic program and do experiment. The results indicate that this program can sort the unsorted images automatically, and able to remove the independent images out of the sequence. After image mosaic, the result is good, the speed is reasonable, achieve the intended purpose.更多还原